drivers/block/mtip32xx/mtip32xx.c

   1 /*
   2  * Driver for the Micron P320 SSD
   3  *   Copyright (C) 2011 Micron Technology, Inc.
   4  *
   5  * Portions of this code were derived from works subjected to the
   6  * following copyright:
   7  *    Copyright (C) 2009 Integrated Device Technology, Inc.
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  */
  20
  21 #include <linux/pci.h>
  22 #include <linux/interrupt.h>
  23 #include <linux/ata.h>
  24 #include <linux/delay.h>
  25 #include <linux/hdreg.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/random.h>
  28 #include <linux/smp.h>
  29 #include <linux/compat.h>
  30 #include <linux/fs.h>
  31 #include <linux/module.h>
  32 #include <linux/genhd.h>
  33 #include <linux/blkdev.h>
  34 #include <linux/blk-mq.h>
  35 #include <linux/bio.h>
  36 #include <linux/dma-mapping.h>
  37 #include <linux/idr.h>
  38 #include <linux/kthread.h>
  39 #include <../drivers/ata/ahci.h>
  40 #include <linux/export.h>
  41 #include <linux/debugfs.h>
  42 #include <linux/prefetch.h>
  43 #include "mtip32xx.h"
  44
  45 #define HW_CMD_SLOT_SZ          (MTIP_MAX_COMMAND_SLOTS * 32)
  46
  47 /* DMA region containing RX Fis, Identify, RLE10, and SMART buffers */
  48 #define AHCI_RX_FIS_SZ          0x100
  49 #define AHCI_RX_FIS_OFFSET      0x0
  50 #define AHCI_IDFY_SZ            ATA_SECT_SIZE
  51 #define AHCI_IDFY_OFFSET        0x400
  52 #define AHCI_SECTBUF_SZ         ATA_SECT_SIZE
  53 #define AHCI_SECTBUF_OFFSET     0x800
  54 #define AHCI_SMARTBUF_SZ        ATA_SECT_SIZE
  55 #define AHCI_SMARTBUF_OFFSET    0xC00
  56 /* 0x100 + 0x200 + 0x200 + 0x200 is smaller than 4k but we pad it out */
  57 #define BLOCK_DMA_ALLOC_SZ      4096
  58
  59 /* DMA region containing command table (should be 8192 bytes) */
  60 #define AHCI_CMD_SLOT_SZ        sizeof(struct mtip_cmd_hdr)
  61 #define AHCI_CMD_TBL_SZ         (MTIP_MAX_COMMAND_SLOTS * AHCI_CMD_SLOT_SZ)
  62 #define AHCI_CMD_TBL_OFFSET     0x0
  63
  64 /* DMA region per command (contains header and SGL) */
  65 #define AHCI_CMD_TBL_HDR_SZ     0x80
  66 #define AHCI_CMD_TBL_HDR_OFFSET 0x0
  67 #define AHCI_CMD_TBL_SGL_SZ     (MTIP_MAX_SG * sizeof(struct mtip_cmd_sg))
  68 #define AHCI_CMD_TBL_SGL_OFFSET AHCI_CMD_TBL_HDR_SZ
  69 #define CMD_DMA_ALLOC_SZ        (AHCI_CMD_TBL_SGL_SZ + AHCI_CMD_TBL_HDR_SZ)
  70
  71
  72 #define HOST_CAP_NZDMA          (1 << 19)
  73 #define HOST_HSORG              0xFC
  74 #define HSORG_DISABLE_SLOTGRP_INTR (1<<24)
  75 #define HSORG_DISABLE_SLOTGRP_PXIS (1<<16)
  76 #define HSORG_HWREV             0xFF00
  77 #define HSORG_STYLE             0x8
  78 #define HSORG_SLOTGROUPS        0x7
  79
  80 #define PORT_COMMAND_ISSUE      0x38
  81 #define PORT_SDBV               0x7C
  82
  83 #define PORT_OFFSET             0x100
  84 #define PORT_MEM_SIZE           0x80
  85
  86 #define PORT_IRQ_ERR \
  87         (PORT_IRQ_HBUS_ERR | PORT_IRQ_IF_ERR | PORT_IRQ_CONNECT | \
  88          PORT_IRQ_PHYRDY | PORT_IRQ_UNK_FIS | PORT_IRQ_BAD_PMP | \
  89          PORT_IRQ_TF_ERR | PORT_IRQ_HBUS_DATA_ERR | PORT_IRQ_IF_NONFATAL | \
  90          PORT_IRQ_OVERFLOW)
  91 #define PORT_IRQ_LEGACY \
  92         (PORT_IRQ_PIOS_FIS | PORT_IRQ_D2H_REG_FIS)
  93 #define PORT_IRQ_HANDLED \
  94         (PORT_IRQ_SDB_FIS | PORT_IRQ_LEGACY | \
  95          PORT_IRQ_TF_ERR | PORT_IRQ_IF_ERR | \
  96          PORT_IRQ_CONNECT | PORT_IRQ_PHYRDY)
  97 #define DEF_PORT_IRQ \
  98         (PORT_IRQ_ERR | PORT_IRQ_LEGACY | PORT_IRQ_SDB_FIS)
  99
 100 /* product numbers */
 101 #define MTIP_PRODUCT_UNKNOWN    0x00
 102 #define MTIP_PRODUCT_ASICFPGA   0x11
 103
 104 /* Device instance number, incremented each time a device is probed. */
 105 static int instance;
 106
 107 static struct list_head online_list;
 108 static struct list_head removing_list;
 109 static spinlock_t dev_lock;
 110
 111 /*
 112  * Global variable used to hold the major block device number
 113  * allocated in mtip_init().
 114  */
 115 static int mtip_major;
 116 static struct dentry *dfs_parent;
 117 static struct dentry *dfs_device_status;
 118
 119 static u32 cpu_use[NR_CPUS];
 120
 121 static DEFINE_SPINLOCK(rssd_index_lock);
 122 static DEFINE_IDA(rssd_index_ida);
 123
 124 static int mtip_block_initialize(struct driver_data *dd);
 125
 126 #ifdef CONFIG_COMPAT
 127 struct mtip_compat_ide_task_request_s {
 128         __u8            io_ports[8];
 129         __u8            hob_ports[8];
 130         ide_reg_valid_t out_flags;
 131         ide_reg_valid_t in_flags;
 132         int             data_phase;
 133         int             req_cmd;
 134         compat_ulong_t  out_size;
 135         compat_ulong_t  in_size;
 136 };
 137 #endif
 138
 139 /*
 140  * This function check_for_surprise_removal is called
 141  * while card is removed from the system and it will
 142  * read the vendor id from the configration space
 143  *
 144  * @pdev Pointer to the pci_dev structure.
 145  *
 146  * return value
 147  *       true if device removed, else false
 148  */
 149 static bool mtip_check_surprise_removal(struct pci_dev *pdev)
 150 {
 151         u16 vendor_id = 0;
 152         struct driver_data *dd = pci_get_drvdata(pdev);
 153
 154         if (dd->sr)
 155                 return true;
 156
 157        /* Read the vendorID from the configuration space */
 158         pci_read_config_word(pdev, 0x00, &vendor_id);
 159         if (vendor_id == 0xFFFF) {
 160                 dd->sr = true;
 161                 if (dd->queue)
 162                         set_bit(QUEUE_FLAG_DEAD, &dd->queue->queue_flags);
 163                 else
 164                         dev_warn(&dd->pdev->dev,
 165                                 "%s: dd->queue is NULL\n", __func__);
 166                 return true; /* device removed */
 167         }
 168
 169         return false; /* device present */
 170 }
 171
 172 static struct mtip_cmd *mtip_get_int_command(struct driver_data *dd)
 173 {
 174         struct request *rq;
 175
 176         if (mtip_check_surprise_removal(dd->pdev))
 177                 return NULL;
 178
 179         rq = blk_mq_alloc_request(dd->queue, 0, BLK_MQ_REQ_RESERVED);
 180         if (IS_ERR(rq))
 181                 return NULL;
 182
 183         return blk_mq_rq_to_pdu(rq);
 184 }
 185
 186 static void mtip_put_int_command(struct driver_data *dd, struct mtip_cmd *cmd)
 187 {
 188         blk_put_request(blk_mq_rq_from_pdu(cmd));
 189 }
 190
 191 /*
 192  * Once we add support for one hctx per mtip group, this will change a bit
 193  */
 194 static struct request *mtip_rq_from_tag(struct driver_data *dd,
 195                                         unsigned int tag)
 196 {
 197         struct blk_mq_hw_ctx *hctx = dd->queue->queue_hw_ctx[0];
 198
 199         return blk_mq_tag_to_rq(hctx->tags, tag);
 200 }
 201
 202 static struct mtip_cmd *mtip_cmd_from_tag(struct driver_data *dd,
 203                                           unsigned int tag)
 204 {
 205         struct request *rq = mtip_rq_from_tag(dd, tag);
 206
 207         return blk_mq_rq_to_pdu(rq);
 208 }
 209
 210 /*
 211  * IO completion function.
 212  *
 213  * This completion function is called by the driver ISR when a
 214  * command that was issued by the kernel completes. It first calls the
 215  * asynchronous completion function which normally calls back into the block
 216  * layer passing the asynchronous callback data, then unmaps the
 217  * scatter list associated with the completed command, and finally
 218  * clears the allocated bit associated with the completed command.
 219  *
 220  * @port   Pointer to the port data structure.
 221  * @tag    Tag of the command.
 222  * @data   Pointer to driver_data.
 223  * @status Completion status.
 224  *
 225  * return value
 226  *      None
 227  */
 228 static void mtip_async_complete(struct mtip_port *port,
 229                                 int tag, struct mtip_cmd *cmd, int status)
 230 {
 231         struct driver_data *dd = port->dd;
 232         struct request *rq;
 233
 234         if (unlikely(!dd) || unlikely(!port))
 235                 return;
 236
 237         if (unlikely(status == PORT_IRQ_TF_ERR)) {
 238                 dev_warn(&port->dd->pdev->dev,
 239                         "Command tag %d failed due to TFE\n", tag);
 240         }
 241
 242         rq = mtip_rq_from_tag(dd, tag);
 243
 244         blk_mq_complete_request(rq, status);
 245 }
 246
 247 /*
 248  * Reset the HBA (without sleeping)
 249  *
 250  * @dd Pointer to the driver data structure.
 251  *
 252  * return value
 253  *      0       The reset was successful.
 254  *      -1      The HBA Reset bit did not clear.
 255  */
 256 static int mtip_hba_reset(struct driver_data *dd)
 257 {
 258         unsigned long timeout;
 259
 260         /* Set the reset bit */
 261         writel(HOST_RESET, dd->mmio + HOST_CTL);
 262
 263         /* Flush */
 264         readl(dd->mmio + HOST_CTL);
 265
 266         /*
 267          * Spin for up to 10 seconds waiting for reset acknowledgement. Spec
 268          * is 1 sec but in LUN failure conditions, up to 10 secs are required
 269          */
 270         timeout = jiffies + msecs_to_jiffies(10000);
 271         do {
 272                 mdelay(10);
 273                 if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag))
 274                         return -1;
 275
 276         } while ((readl(dd->mmio + HOST_CTL) & HOST_RESET)
 277                  && time_before(jiffies, timeout));
 278
 279         if (readl(dd->mmio + HOST_CTL) & HOST_RESET)
 280                 return -1;
 281
 282         return 0;
 283 }
 284
 285 /*
 286  * Issue a command to the hardware.
 287  *
 288  * Set the appropriate bit in the s_active and Command Issue hardware
 289  * registers, causing hardware command processing to begin.
 290  *
 291  * @port Pointer to the port structure.
 292  * @tag  The tag of the command to be issued.
 293  *
 294  * return value
 295  *      None
 296  */
 297 static inline void mtip_issue_ncq_command(struct mtip_port *port, int tag)
 298 {
 299         int group = tag >> 5;
 300
 301         /* guard SACT and CI registers */
 302         spin_lock(&port->cmd_issue_lock[group]);
 303         writel((1 << MTIP_TAG_BIT(tag)),
 304                         port->s_active[MTIP_TAG_INDEX(tag)]);
 305         writel((1 << MTIP_TAG_BIT(tag)),
 306                         port->cmd_issue[MTIP_TAG_INDEX(tag)]);
 307         spin_unlock(&port->cmd_issue_lock[group]);
 308 }
 309
 310 /*
 311  * Enable/disable the reception of FIS
 312  *
 313  * @port   Pointer to the port data structure
 314  * @enable 1 to enable, 0 to disable
 315  *
 316  * return value
 317  *      Previous state: 1 enabled, 0 disabled
 318  */
 319 static int mtip_enable_fis(struct mtip_port *port, int enable)
 320 {
 321         u32 tmp;
 322
 323         /* enable FIS reception */
 324         tmp = readl(port->mmio + PORT_CMD);
 325         if (enable)
 326                 writel(tmp | PORT_CMD_FIS_RX, port->mmio + PORT_CMD);
 327         else
 328                 writel(tmp & ~PORT_CMD_FIS_RX, port->mmio + PORT_CMD);
 329
 330         /* Flush */
 331         readl(port->mmio + PORT_CMD);
 332
 333         return (((tmp & PORT_CMD_FIS_RX) == PORT_CMD_FIS_RX));
 334 }
 335
 336 /*
 337  * Enable/disable the DMA engine
 338  *
 339  * @port   Pointer to the port data structure
 340  * @enable 1 to enable, 0 to disable
 341  *
 342  * return value
 343  *      Previous state: 1 enabled, 0 disabled.
 344  */
 345 static int mtip_enable_engine(struct mtip_port *port, int enable)
 346 {
 347         u32 tmp;
 348
 349         /* enable FIS reception */
 350         tmp = readl(port->mmio + PORT_CMD);
 351         if (enable)
 352                 writel(tmp | PORT_CMD_START, port->mmio + PORT_CMD);
 353         else
 354                 writel(tmp & ~PORT_CMD_START, port->mmio + PORT_CMD);
 355
 356         readl(port->mmio + PORT_CMD);
 357         return (((tmp & PORT_CMD_START) == PORT_CMD_START));
 358 }
 359
 360 /*
 361  * Enables the port DMA engine and FIS reception.
 362  *
 363  * return value
 364  *      None
 365  */
 366 static inline void mtip_start_port(struct mtip_port *port)
 367 {
 368         /* Enable FIS reception */
 369         mtip_enable_fis(port, 1);
 370
 371         /* Enable the DMA engine */
 372         mtip_enable_engine(port, 1);
 373 }
 374
 375 /*
 376  * Deinitialize a port by disabling port interrupts, the DMA engine,
 377  * and FIS reception.
 378  *
 379  * @port Pointer to the port structure
 380  *
 381  * return value
 382  *      None
 383  */
 384 static inline void mtip_deinit_port(struct mtip_port *port)
 385 {
 386         /* Disable interrupts on this port */
 387         writel(0, port->mmio + PORT_IRQ_MASK);
 388
 389         /* Disable the DMA engine */
 390         mtip_enable_engine(port, 0);
 391
 392         /* Disable FIS reception */
 393         mtip_enable_fis(port, 0);
 394 }
 395
 396 /*
 397  * Initialize a port.
 398  *
 399  * This function deinitializes the port by calling mtip_deinit_port() and
 400  * then initializes it by setting the command header and RX FIS addresses,
 401  * clearing the SError register and any pending port interrupts before
 402  * re-enabling the default set of port interrupts.
 403  *
 404  * @port Pointer to the port structure.
 405  *
 406  * return value
 407  *      None
 408  */
 409 static void mtip_init_port(struct mtip_port *port)
 410 {
 411         int i;
 412         mtip_deinit_port(port);
 413
 414         /* Program the command list base and FIS base addresses */
 415         if (readl(port->dd->mmio + HOST_CAP) & HOST_CAP_64) {
 416                 writel((port->command_list_dma >> 16) >> 16,
 417                          port->mmio + PORT_LST_ADDR_HI);
 418                 writel((port->rxfis_dma >> 16) >> 16,
 419                          port->mmio + PORT_FIS_ADDR_HI);
 420         }
 421
 422         writel(port->command_list_dma & 0xFFFFFFFF,
 423                         port->mmio + PORT_LST_ADDR);
 424         writel(port->rxfis_dma & 0xFFFFFFFF, port->mmio + PORT_FIS_ADDR);
 425
 426         /* Clear SError */
 427         writel(readl(port->mmio + PORT_SCR_ERR), port->mmio + PORT_SCR_ERR);
 428
 429         /* reset the completed registers.*/
 430         for (i = 0; i < port->dd->slot_groups; i++)
 431                 writel(0xFFFFFFFF, port->completed[i]);
 432
 433         /* Clear any pending interrupts for this port */
 434         writel(readl(port->mmio + PORT_IRQ_STAT), port->mmio + PORT_IRQ_STAT);
 435
 436         /* Clear any pending interrupts on the HBA. */
 437         writel(readl(port->dd->mmio + HOST_IRQ_STAT),
 438                                         port->dd->mmio + HOST_IRQ_STAT);
 439
 440         /* Enable port interrupts */
 441         writel(DEF_PORT_IRQ, port->mmio + PORT_IRQ_MASK);
 442 }
 443
 444 /*
 445  * Restart a port
 446  *
 447  * @port Pointer to the port data structure.
 448  *
 449  * return value
 450  *      None
 451  */
 452 static void mtip_restart_port(struct mtip_port *port)
 453 {
 454         unsigned long timeout;
 455
 456         /* Disable the DMA engine */
 457         mtip_enable_engine(port, 0);
 458
 459         /* Chip quirk: wait up to 500ms for PxCMD.CR == 0 */
 460         timeout = jiffies + msecs_to_jiffies(500);
 461         while ((readl(port->mmio + PORT_CMD) & PORT_CMD_LIST_ON)
 462                  && time_before(jiffies, timeout))
 463                 ;
 464
 465         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 466                 return;
 467
 468         /*
 469          * Chip quirk: escalate to hba reset if
 470          * PxCMD.CR not clear after 500 ms
 471          */
 472         if (readl(port->mmio + PORT_CMD) & PORT_CMD_LIST_ON) {
 473                 dev_warn(&port->dd->pdev->dev,
 474                         "PxCMD.CR not clear, escalating reset\n");
 475
 476                 if (mtip_hba_reset(port->dd))
 477                         dev_err(&port->dd->pdev->dev,
 478                                 "HBA reset escalation failed.\n");
 479
 480                 /* 30 ms delay before com reset to quiesce chip */
 481                 mdelay(30);
 482         }
 483
 484         dev_warn(&port->dd->pdev->dev, "Issuing COM reset\n");
 485
 486         /* Set PxSCTL.DET */
 487         writel(readl(port->mmio + PORT_SCR_CTL) |
 488                          1, port->mmio + PORT_SCR_CTL);
 489         readl(port->mmio + PORT_SCR_CTL);
 490
 491         /* Wait 1 ms to quiesce chip function */
 492         timeout = jiffies + msecs_to_jiffies(1);
 493         while (time_before(jiffies, timeout))
 494                 ;
 495
 496         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 497                 return;
 498
 499         /* Clear PxSCTL.DET */
 500         writel(readl(port->mmio + PORT_SCR_CTL) & ~1,
 501                          port->mmio + PORT_SCR_CTL);
 502         readl(port->mmio + PORT_SCR_CTL);
 503
 504         /* Wait 500 ms for bit 0 of PORT_SCR_STS to be set */
 505         timeout = jiffies + msecs_to_jiffies(500);
 506         while (((readl(port->mmio + PORT_SCR_STAT) & 0x01) == 0)
 507                          && time_before(jiffies, timeout))
 508                 ;
 509
 510         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 511                 return;
 512
 513         if ((readl(port->mmio + PORT_SCR_STAT) & 0x01) == 0)
 514                 dev_warn(&port->dd->pdev->dev,
 515                         "COM reset failed\n");
 516
 517         mtip_init_port(port);
 518         mtip_start_port(port);
 519
 520 }
 521
 522 static int mtip_device_reset(struct driver_data *dd)
 523 {
 524         int rv = 0;
 525
 526         if (mtip_check_surprise_removal(dd->pdev))
 527                 return 0;
 528
 529         if (mtip_hba_reset(dd) < 0)
 530                 rv = -EFAULT;
 531
 532         mdelay(1);
 533         mtip_init_port(dd->port);
 534         mtip_start_port(dd->port);
 535
 536         /* Enable interrupts on the HBA. */
 537         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
 538                                         dd->mmio + HOST_CTL);
 539         return rv;
 540 }
 541
 542 /*
 543  * Helper function for tag logging
 544  */
 545 static void print_tags(struct driver_data *dd,
 546                         char *msg,
 547                         unsigned long *tagbits,
 548                         int cnt)
 549 {
 550         unsigned char tagmap[128];
 551         int group, tagmap_len = 0;
 552
 553         memset(tagmap, 0, sizeof(tagmap));
 554         for (group = SLOTBITS_IN_LONGS; group > 0; group--)
 555                 tagmap_len += sprintf(tagmap + tagmap_len, "%016lX ",
 556                                                 tagbits[group-1]);
 557         dev_warn(&dd->pdev->dev,
 558                         "%d command(s) %s: tagmap [%s]", cnt, msg, tagmap);
 559 }
 560
 561 /*
 562  * Internal command completion callback function.
 563  *
 564  * This function is normally called by the driver ISR when an internal
 565  * command completed. This function signals the command completion by
 566  * calling complete().
 567  *
 568  * @port   Pointer to the port data structure.
 569  * @tag    Tag of the command that has completed.
 570  * @data   Pointer to a completion structure.
 571  * @status Completion status.
 572  *
 573  * return value
 574  *      None
 575  */
 576 static void mtip_completion(struct mtip_port *port,
 577                             int tag, struct mtip_cmd *command, int status)
 578 {
 579         struct completion *waiting = command->comp_data;
 580         if (unlikely(status == PORT_IRQ_TF_ERR))
 581                 dev_warn(&port->dd->pdev->dev,
 582                         "Internal command %d completed with TFE\n", tag);
 583
 584         command->comp_func = NULL;
 585         command->comp_data = NULL;
 586         complete(waiting);
 587 }
 588
 589 static void mtip_null_completion(struct mtip_port *port,
 590                             int tag, struct mtip_cmd *command, int status)
 591 {
 592 }
 593
 594 static int mtip_read_log_page(struct mtip_port *port, u8 page, u16 *buffer,
 595                                 dma_addr_t buffer_dma, unsigned int sectors);
 596 static int mtip_get_smart_attr(struct mtip_port *port, unsigned int id,
 597                                                 struct smart_attr *attrib);
 598 /*
 599  * Handle an error.
 600  *
 601  * @dd Pointer to the DRIVER_DATA structure.
 602  *
 603  * return value
 604  *      None
 605  */
 606 static void mtip_handle_tfe(struct driver_data *dd)
 607 {
 608         int group, tag, bit, reissue, rv;
 609         struct mtip_port *port;
 610         struct mtip_cmd  *cmd;
 611         u32 completed;
 612         struct host_to_dev_fis *fis;
 613         unsigned long tagaccum[SLOTBITS_IN_LONGS];
 614         unsigned int cmd_cnt = 0;
 615         unsigned char *buf;
 616         char *fail_reason = NULL;
 617         int fail_all_ncq_write = 0, fail_all_ncq_cmds = 0;
 618
 619         dev_warn(&dd->pdev->dev, "Taskfile error\n");
 620
 621         port = dd->port;
 622
 623         if (test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags)) {
 624                 cmd = mtip_cmd_from_tag(dd, MTIP_TAG_INTERNAL);
 625                 dbg_printk(MTIP_DRV_NAME " TFE for the internal command\n");
 626
 627                 if (cmd->comp_data && cmd->comp_func) {
 628                         cmd->comp_func(port, MTIP_TAG_INTERNAL,
 629                                         cmd, PORT_IRQ_TF_ERR);
 630                 }
 631                 return;
 632         }
 633
 634         /* clear the tag accumulator */
 635         memset(tagaccum, 0, SLOTBITS_IN_LONGS * sizeof(long));
 636
 637         /* Loop through all the groups */
 638         for (group = 0; group < dd->slot_groups; group++) {
 639                 completed = readl(port->completed[group]);
 640
 641                 dev_warn(&dd->pdev->dev, "g=%u, comp=%x\n", group, completed);
 642
 643                 /* clear completed status register in the hardware.*/
 644                 writel(completed, port->completed[group]);
 645
 646                 /* Process successfully completed commands */
 647                 for (bit = 0; bit < 32 && completed; bit++) {
 648                         if (!(completed & (1<<bit)))
 649                                 continue;
 650                         tag = (group << 5) + bit;
 651
 652                         /* Skip the internal command slot */
 653                         if (tag == MTIP_TAG_INTERNAL)
 654                                 continue;
 655
 656                         cmd = mtip_cmd_from_tag(dd, tag);
 657                         if (likely(cmd->comp_func)) {
 658                                 set_bit(tag, tagaccum);
 659                                 cmd_cnt++;
 660                                 cmd->comp_func(port, tag, cmd, 0);
 661                         } else {
 662                                 dev_err(&port->dd->pdev->dev,
 663                                         "Missing completion func for tag %d",
 664                                         tag);
 665                                 if (mtip_check_surprise_removal(dd->pdev)) {
 666                                         /* don't proceed further */
 667                                         return;
 668                                 }
 669                         }
 670                 }
 671         }
 672
 673         print_tags(dd, "completed (TFE)", tagaccum, cmd_cnt);
 674
 675         /* Restart the port */
 676         mdelay(20);
 677         mtip_restart_port(port);
 678
 679         /* Trying to determine the cause of the error */
 680         rv = mtip_read_log_page(dd->port, ATA_LOG_SATA_NCQ,
 681                                 dd->port->log_buf,
 682                                 dd->port->log_buf_dma, 1);
 683         if (rv) {
 684                 dev_warn(&dd->pdev->dev,
 685                         "Error in READ LOG EXT (10h) command\n");
 686                 /* non-critical error, don't fail the load */
 687         } else {
 688                 buf = (unsigned char *)dd->port->log_buf;
 689                 if (buf[259] & 0x1) {
 690                         dev_info(&dd->pdev->dev,
 691                                 "Write protect bit is set.\n");
 692                         set_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag);
 693                         fail_all_ncq_write = 1;
 694                         fail_reason = "write protect";
 695                 }
 696                 if (buf[288] == 0xF7) {
 697                         dev_info(&dd->pdev->dev,
 698                                 "Exceeded Tmax, drive in thermal shutdown.\n");
 699                         set_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag);
 700                         fail_all_ncq_cmds = 1;
 701                         fail_reason = "thermal shutdown";
 702                 }
 703                 if (buf[288] == 0xBF) {
 704                         set_bit(MTIP_DDF_REBUILD_FAILED_BIT, &dd->dd_flag);
 705                         dev_info(&dd->pdev->dev,
 706                                 "Drive indicates rebuild has failed. Secure erase required.\n");
 707                         fail_all_ncq_cmds = 1;
 708                         fail_reason = "rebuild failed";
 709                 }
 710         }
 711
 712         /* clear the tag accumulator */
 713         memset(tagaccum, 0, SLOTBITS_IN_LONGS * sizeof(long));
 714
 715         /* Loop through all the groups */
 716         for (group = 0; group < dd->slot_groups; group++) {
 717                 for (bit = 0; bit < 32; bit++) {
 718                         reissue = 1;
 719                         tag = (group << 5) + bit;
 720                         cmd = mtip_cmd_from_tag(dd, tag);
 721
 722                         fis = (struct host_to_dev_fis *)cmd->command;
 723
 724                         /* Should re-issue? */
 725                         if (tag == MTIP_TAG_INTERNAL ||
 726                             fis->command == ATA_CMD_SET_FEATURES)
 727                                 reissue = 0;
 728                         else {
 729                                 if (fail_all_ncq_cmds ||
 730                                         (fail_all_ncq_write &&
 731                                         fis->command == ATA_CMD_FPDMA_WRITE)) {
 732                                         dev_warn(&dd->pdev->dev,
 733                                         "  Fail: %s w/tag %d [%s].\n",
 734                                         fis->command == ATA_CMD_FPDMA_WRITE ?
 735                                                 "write" : "read",
 736                                         tag,
 737                                         fail_reason != NULL ?
 738                                                 fail_reason : "unknown");
 739                                         if (cmd->comp_func) {
 740                                                 cmd->comp_func(port, tag,
 741                                                         cmd, -ENODATA);
 742                                         }
 743                                         continue;
 744                                 }
 745                         }
 746
 747                         /*
 748                          * First check if this command has
 749                          *  exceeded its retries.
 750                          */
 751                         if (reissue && (cmd->retries-- > 0)) {
 752
 753                                 set_bit(tag, tagaccum);
 754
 755                                 /* Re-issue the command. */
 756                                 mtip_issue_ncq_command(port, tag);
 757
 758                                 continue;
 759                         }
 760
 761                         /* Retire a command that will not be reissued */
 762                         dev_warn(&port->dd->pdev->dev,
 763                                 "retiring tag %d\n", tag);
 764
 765                         if (cmd->comp_func)
 766                                 cmd->comp_func(port, tag, cmd, PORT_IRQ_TF_ERR);
 767                         else
 768                                 dev_warn(&port->dd->pdev->dev,
 769                                         "Bad completion for tag %d\n",
 770                                         tag);
 771                 }
 772         }
 773         print_tags(dd, "reissued (TFE)", tagaccum, cmd_cnt);
 774 }
 775
 776 /*
 777  * Handle a set device bits interrupt
 778  */
 779 static inline void mtip_workq_sdbfx(struct mtip_port *port, int group,
 780                                                         u32 completed)
 781 {
 782         struct driver_data *dd = port->dd;
 783         int tag, bit;
 784         struct mtip_cmd *command;
 785
 786         if (!completed) {
 787                 WARN_ON_ONCE(!completed);
 788                 return;
 789         }
 790         /* clear completed status register in the hardware.*/
 791         writel(completed, port->completed[group]);
 792
 793         /* Process completed commands. */
 794         for (bit = 0; (bit < 32) && completed; bit++) {
 795                 if (completed & 0x01) {
 796                         tag = (group << 5) | bit;
 797
 798                         /* skip internal command slot. */
 799                         if (unlikely(tag == MTIP_TAG_INTERNAL))
 800                                 continue;
 801
 802                         command = mtip_cmd_from_tag(dd, tag);
 803                         if (likely(command->comp_func))
 804                                 command->comp_func(port, tag, command, 0);
 805                         else {
 806                                 dev_dbg(&dd->pdev->dev,
 807                                         "Null completion for tag %d",
 808                                         tag);
 809
 810                                 if (mtip_check_surprise_removal(
 811                                         dd->pdev)) {
 812                                         return;
 813                                 }
 814                         }
 815                 }
 816                 completed >>= 1;
 817         }
 818
 819         /* If last, re-enable interrupts */
 820         if (atomic_dec_return(&dd->irq_workers_active) == 0)
 821                 writel(0xffffffff, dd->mmio + HOST_IRQ_STAT);
 822 }
 823
 824 /*
 825  * Process legacy pio and d2h interrupts
 826  */
 827 static inline void mtip_process_legacy(struct driver_data *dd, u32 port_stat)
 828 {
 829         struct mtip_port *port = dd->port;
 830         struct mtip_cmd *cmd = mtip_cmd_from_tag(dd, MTIP_TAG_INTERNAL);
 831
 832         if (test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags) &&
 833             (cmd != NULL) && !(readl(port->cmd_issue[MTIP_TAG_INTERNAL])
 834                 & (1 << MTIP_TAG_INTERNAL))) {
 835                 if (cmd->comp_func) {
 836                         cmd->comp_func(port, MTIP_TAG_INTERNAL, cmd, 0);
 837                         return;
 838                 }
 839         }
 840
 841         return;
 842 }
 843
 844 /*
 845  * Demux and handle errors
 846  */
 847 static inline void mtip_process_errors(struct driver_data *dd, u32 port_stat)
 848 {
 849
 850         if (unlikely(port_stat & PORT_IRQ_CONNECT)) {
 851                 dev_warn(&dd->pdev->dev,
 852                         "Clearing PxSERR.DIAG.x\n");
 853                 writel((1 << 26), dd->port->mmio + PORT_SCR_ERR);
 854         }
 855
 856         if (unlikely(port_stat & PORT_IRQ_PHYRDY)) {
 857                 dev_warn(&dd->pdev->dev,
 858                         "Clearing PxSERR.DIAG.n\n");
 859                 writel((1 << 16), dd->port->mmio + PORT_SCR_ERR);
 860         }
 861
 862         if (unlikely(port_stat & ~PORT_IRQ_HANDLED)) {
 863                 dev_warn(&dd->pdev->dev,
 864                         "Port stat errors %x unhandled\n",
 865                         (port_stat & ~PORT_IRQ_HANDLED));
 866                 if (mtip_check_surprise_removal(dd->pdev))
 867                         return;
 868         }
 869         if (likely(port_stat & (PORT_IRQ_TF_ERR | PORT_IRQ_IF_ERR))) {
 870                 set_bit(MTIP_PF_EH_ACTIVE_BIT, &dd->port->flags);
 871                 wake_up_interruptible(&dd->port->svc_wait);
 872         }
 873 }
 874
 875 static inline irqreturn_t mtip_handle_irq(struct driver_data *data)
 876 {
 877         struct driver_data *dd = (struct driver_data *) data;
 878         struct mtip_port *port = dd->port;
 879         u32 hba_stat, port_stat;
 880         int rv = IRQ_NONE;
 881         int do_irq_enable = 1, i, workers;
 882         struct mtip_work *twork;
 883
 884         hba_stat = readl(dd->mmio + HOST_IRQ_STAT);
 885         if (hba_stat) {
 886                 rv = IRQ_HANDLED;
 887
 888                 /* Acknowledge the interrupt status on the port.*/
 889                 port_stat = readl(port->mmio + PORT_IRQ_STAT);
 890                 if (unlikely(port_stat == 0xFFFFFFFF)) {
 891                         mtip_check_surprise_removal(dd->pdev);
 892                         return IRQ_HANDLED;
 893                 }
 894                 writel(port_stat, port->mmio + PORT_IRQ_STAT);
 895
 896                 /* Demux port status */
 897                 if (likely(port_stat & PORT_IRQ_SDB_FIS)) {
 898                         do_irq_enable = 0;
 899                         WARN_ON_ONCE(atomic_read(&dd->irq_workers_active) != 0);
 900
 901                         /* Start at 1: group zero is always local? */
 902                         for (i = 0, workers = 0; i < MTIP_MAX_SLOT_GROUPS;
 903                                                                         i++) {
 904                                 twork = &dd->work[i];
 905                                 twork->completed = readl(port->completed[i]);
 906                                 if (twork->completed)
 907                                         workers++;
 908                         }
 909
 910                         atomic_set(&dd->irq_workers_active, workers);
 911                         if (workers) {
 912                                 for (i = 1; i < MTIP_MAX_SLOT_GROUPS; i++) {
 913                                         twork = &dd->work[i];
 914                                         if (twork->completed)
 915                                                 queue_work_on(
 916                                                         twork->cpu_binding,
 917                                                         dd->isr_workq,
 918                                                         &twork->work);
 919                                 }
 920
 921                                 if (likely(dd->work[0].completed))
 922                                         mtip_workq_sdbfx(port, 0,
 923                                                         dd->work[0].completed);
 924
 925                         } else {
 926                                 /*
 927                                  * Chip quirk: SDB interrupt but nothing
 928                                  * to complete
 929                                  */
 930                                 do_irq_enable = 1;
 931                         }
 932                 }
 933
 934                 if (unlikely(port_stat & PORT_IRQ_ERR)) {
 935                         if (unlikely(mtip_check_surprise_removal(dd->pdev))) {
 936                                 /* don't proceed further */
 937                                 return IRQ_HANDLED;
 938                         }
 939                         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
 940                                                         &dd->dd_flag))
 941                                 return rv;
 942
 943                         mtip_process_errors(dd, port_stat & PORT_IRQ_ERR);
 944                 }
 945
 946                 if (unlikely(port_stat & PORT_IRQ_LEGACY))
 947                         mtip_process_legacy(dd, port_stat & PORT_IRQ_LEGACY);
 948         }
 949
 950         /* acknowledge interrupt */
 951         if (unlikely(do_irq_enable))
 952                 writel(hba_stat, dd->mmio + HOST_IRQ_STAT);
 953
 954         return rv;
 955 }
 956
 957 /*
 958  * HBA interrupt subroutine.
 959  *
 960  * @irq         IRQ number.
 961  * @instance    Pointer to the driver data structure.
 962  *
 963  * return value
 964  *      IRQ_HANDLED     A HBA interrupt was pending and handled.
 965  *      IRQ_NONE        This interrupt was not for the HBA.
 966  */
 967 static irqreturn_t mtip_irq_handler(int irq, void *instance)
 968 {
 969         struct driver_data *dd = instance;
 970
 971         return mtip_handle_irq(dd);
 972 }
 973
 974 static void mtip_issue_non_ncq_command(struct mtip_port *port, int tag)
 975 {
 976         writel(1 << MTIP_TAG_BIT(tag),
 977                 port->cmd_issue[MTIP_TAG_INDEX(tag)]);
 978 }
 979
 980 static bool mtip_pause_ncq(struct mtip_port *port,
 981                                 struct host_to_dev_fis *fis)
 982 {
 983         struct host_to_dev_fis *reply;
 984         unsigned long task_file_data;
 985
 986         reply = port->rxfis + RX_FIS_D2H_REG;
 987         task_file_data = readl(port->mmio+PORT_TFDATA);
 988
 989         if ((task_file_data & 1))
 990                 return false;
 991
 992         if (fis->command == ATA_CMD_SEC_ERASE_PREP) {
 993                 port->ic_pause_timer = jiffies;
 994                 return true;
 995         } else if ((fis->command == ATA_CMD_DOWNLOAD_MICRO) &&
 996                                         (fis->features == 0x03)) {
 997                 set_bit(MTIP_PF_DM_ACTIVE_BIT, &port->flags);
 998                 port->ic_pause_timer = jiffies;
 999                 return true;
1000         } else if ((fis->command == ATA_CMD_SEC_ERASE_UNIT) ||
1001                 ((fis->command == 0xFC) &&
1002                         (fis->features == 0x27 || fis->features == 0x72 ||
1003                          fis->features == 0x62 || fis->features == 0x26))) {
1004                 clear_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1005                 clear_bit(MTIP_DDF_REBUILD_FAILED_BIT, &port->dd->dd_flag);
1006                 /* Com reset after secure erase or lowlevel format */
1007                 mtip_restart_port(port);
1008                 clear_bit(MTIP_PF_SE_ACTIVE_BIT, &port->flags);
1009                 return false;
1010         }
1011
1012         return false;
1013 }
1014
1015 /*
1016  * Wait for port to quiesce
1017  *
1018  * @port    Pointer to port data structure
1019  * @timeout Max duration to wait (ms)
1020  * @atomic  gfp_t flag to indicate blockable context or not
1021  *
1022  * return value
1023  *      0       Success
1024  *      -EBUSY  Commands still active
1025  */
1026 static int mtip_quiesce_io(struct mtip_port *port, unsigned long timeout,
1027                                                                 gfp_t atomic)
1028 {
1029         unsigned long to;
1030         unsigned int n;
1031         unsigned int active = 1;
1032
1033         blk_mq_stop_hw_queues(port->dd->queue);
1034
1035         to = jiffies + msecs_to_jiffies(timeout);
1036         do {
1037                 if (test_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags) &&
1038                         test_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags) &&
1039                         atomic == GFP_KERNEL) {
1040                         msleep(20);
1041                         continue; /* svc thd is actively issuing commands */
1042                 }
1043
1044                 if (atomic == GFP_KERNEL)
1045                         msleep(100);
1046                 else {
1047                         cpu_relax();
1048                         udelay(100);
1049                 }
1050
1051                 if (mtip_check_surprise_removal(port->dd->pdev))
1052                         goto err_fault;
1053
1054                 /*
1055                  * Ignore s_active bit 0 of array element 0.
1056                  * This bit will always be set
1057                  */
1058                 active = readl(port->s_active[0]) & 0xFFFFFFFE;
1059                 for (n = 1; n < port->dd->slot_groups; n++)
1060                         active |= readl(port->s_active[n]);
1061
1062                 if (!active)
1063                         break;
1064         } while (time_before(jiffies, to));
1065
1066         blk_mq_start_stopped_hw_queues(port->dd->queue, true);
1067         return active ? -EBUSY : 0;
1068 err_fault:
1069         blk_mq_start_stopped_hw_queues(port->dd->queue, true);
1070         return -EFAULT;
1071 }
1072
1073 /*
1074  * Execute an internal command and wait for the completion.
1075  *
1076  * @port    Pointer to the port data structure.
1077  * @fis     Pointer to the FIS that describes the command.
1078  * @fis_len  Length in WORDS of the FIS.
1079  * @buffer  DMA accessible for command data.
1080  * @buf_len  Length, in bytes, of the data buffer.
1081  * @opts    Command header options, excluding the FIS length
1082  *             and the number of PRD entries.
1083  * @timeout Time in ms to wait for the command to complete.
1084  *
1085  * return value
1086  *      0        Command completed successfully.
1087  *      -EFAULT  The buffer address is not correctly aligned.
1088  *      -EBUSY   Internal command or other IO in progress.
1089  *      -EAGAIN  Time out waiting for command to complete.
1090  */
1091 static int mtip_exec_internal_command(struct mtip_port *port,
1092                                         struct host_to_dev_fis *fis,
1093                                         int fis_len,
1094                                         dma_addr_t buffer,
1095                                         int buf_len,
1096                                         u32 opts,
1097                                         gfp_t atomic,
1098                                         unsigned long timeout)
1099 {
1100         struct mtip_cmd_sg *command_sg;
1101         DECLARE_COMPLETION_ONSTACK(wait);
1102         struct mtip_cmd *int_cmd;
1103         struct driver_data *dd = port->dd;
1104         int rv = 0;
1105         unsigned long start;
1106
1107         /* Make sure the buffer is 8 byte aligned. This is asic specific. */
1108         if (buffer & 0x00000007) {
1109                 dev_err(&dd->pdev->dev, "SG buffer is not 8 byte aligned\n");
1110                 return -EFAULT;
1111         }
1112
1113         int_cmd = mtip_get_int_command(dd);
1114         if (!int_cmd) {
1115                 dbg_printk(MTIP_DRV_NAME "Unable to allocate tag for PIO cmd\n");
1116                 return -EFAULT;
1117         }
1118
1119         set_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1120
1121         if (fis->command == ATA_CMD_SEC_ERASE_PREP)
1122                 set_bit(MTIP_PF_SE_ACTIVE_BIT, &port->flags);
1123
1124         clear_bit(MTIP_PF_DM_ACTIVE_BIT, &port->flags);
1125
1126         if (atomic == GFP_KERNEL) {
1127                 if (fis->command != ATA_CMD_STANDBYNOW1) {
1128                         /* wait for io to complete if non atomic */
1129                         if (mtip_quiesce_io(port,
1130                                 MTIP_QUIESCE_IO_TIMEOUT_MS, atomic) < 0) {
1131                                 dev_warn(&dd->pdev->dev,
1132                                         "Failed to quiesce IO\n");
1133                                 mtip_put_int_command(dd, int_cmd);
1134                                 clear_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1135                                 wake_up_interruptible(&port->svc_wait);
1136                                 return -EBUSY;
1137                         }
1138                 }
1139
1140                 /* Set the completion function and data for the command. */
1141                 int_cmd->comp_data = &wait;
1142                 int_cmd->comp_func = mtip_completion;
1143
1144         } else {
1145                 /* Clear completion - we're going to poll */
1146                 int_cmd->comp_data = NULL;
1147                 int_cmd->comp_func = mtip_null_completion;
1148         }
1149
1150         /* Copy the command to the command table */
1151         memcpy(int_cmd->command, fis, fis_len*4);
1152
1153         /* Populate the SG list */
1154         int_cmd->command_header->opts =
1155                  __force_bit2int cpu_to_le32(opts | fis_len);
1156         if (buf_len) {
1157                 command_sg = int_cmd->command + AHCI_CMD_TBL_HDR_SZ;
1158
1159                 command_sg->info =
1160                         __force_bit2int cpu_to_le32((buf_len-1) & 0x3FFFFF);
1161                 command_sg->dba =
1162                         __force_bit2int cpu_to_le32(buffer & 0xFFFFFFFF);
1163                 command_sg->dba_upper =
1164                         __force_bit2int cpu_to_le32((buffer >> 16) >> 16);
1165
1166                 int_cmd->command_header->opts |=
1167                         __force_bit2int cpu_to_le32((1 << 16));
1168         }
1169
1170         /* Populate the command header */
1171         int_cmd->command_header->byte_count = 0;
1172
1173         start = jiffies;
1174
1175         /* Issue the command to the hardware */
1176         mtip_issue_non_ncq_command(port, MTIP_TAG_INTERNAL);
1177
1178         if (atomic == GFP_KERNEL) {
1179                 /* Wait for the command to complete or timeout. */
1180                 if ((rv = wait_for_completion_interruptible_timeout(
1181                                 &wait,
1182                                 msecs_to_jiffies(timeout))) <= 0) {
1183
1184                         if (rv == -ERESTARTSYS) { /* interrupted */
1185                                 dev_err(&dd->pdev->dev,
1186                                         "Internal command [%02X] was interrupted after %u ms\n",
1187                                         fis->command,
1188                                         jiffies_to_msecs(jiffies - start));
1189                                 rv = -EINTR;
1190                                 goto exec_ic_exit;
1191                         } else if (rv == 0) /* timeout */
1192                                 dev_err(&dd->pdev->dev,
1193                                         "Internal command did not complete [%02X] within timeout of  %lu ms\n",
1194                                         fis->command, timeout);
1195                         else
1196                                 dev_err(&dd->pdev->dev,
1197                                         "Internal command [%02X] wait returned code [%d] after %lu ms - unhandled\n",
1198                                         fis->command, rv, timeout);
1199
1200                         if (mtip_check_surprise_removal(dd->pdev) ||
1201                                 test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
1202                                                 &dd->dd_flag)) {
1203                                 dev_err(&dd->pdev->dev,
1204                                         "Internal command [%02X] wait returned due to SR\n",
1205                                         fis->command);
1206                                 rv = -ENXIO;
1207                                 goto exec_ic_exit;
1208                         }
1209                         mtip_device_reset(dd); /* recover from timeout issue */
1210                         rv = -EAGAIN;
1211                         goto exec_ic_exit;
1212                 }
1213         } else {
1214                 u32 hba_stat, port_stat;
1215
1216                 /* Spin for <timeout> checking if command still outstanding */
1217                 timeout = jiffies + msecs_to_jiffies(timeout);
1218                 while ((readl(port->cmd_issue[MTIP_TAG_INTERNAL])
1219                                 & (1 << MTIP_TAG_INTERNAL))
1220                                 && time_before(jiffies, timeout)) {
1221                         if (mtip_check_surprise_removal(dd->pdev)) {
1222                                 rv = -ENXIO;
1223                                 goto exec_ic_exit;
1224                         }
1225                         if ((fis->command != ATA_CMD_STANDBYNOW1) &&
1226                                 test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
1227                                                 &dd->dd_flag)) {
1228                                 rv = -ENXIO;
1229                                 goto exec_ic_exit;
1230                         }
1231                         port_stat = readl(port->mmio + PORT_IRQ_STAT);
1232                         if (!port_stat)
1233                                 continue;
1234
1235                         if (port_stat & PORT_IRQ_ERR) {
1236                                 dev_err(&dd->pdev->dev,
1237                                         "Internal command [%02X] failed\n",
1238                                         fis->command);
1239                                 mtip_device_reset(dd);
1240                                 rv = -EIO;
1241                                 goto exec_ic_exit;
1242                         } else {
1243                                 writel(port_stat, port->mmio + PORT_IRQ_STAT);
1244                                 hba_stat = readl(dd->mmio + HOST_IRQ_STAT);
1245                                 if (hba_stat)
1246                                         writel(hba_stat,
1247                                                 dd->mmio + HOST_IRQ_STAT);
1248                         }
1249                         break;
1250                 }
1251         }
1252
1253         if (readl(port->cmd_issue[MTIP_TAG_INTERNAL])
1254                         & (1 << MTIP_TAG_INTERNAL)) {
1255                 rv = -ENXIO;
1256                 if (!test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)) {
1257                         mtip_device_reset(dd);
1258                         rv = -EAGAIN;
1259                 }
1260         }
1261 exec_ic_exit:
1262         /* Clear the allocated and active bits for the internal command. */
1263         mtip_put_int_command(dd, int_cmd);
1264         clear_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1265         if (rv >= 0 && mtip_pause_ncq(port, fis)) {
1266                 /* NCQ paused */
1267                 return rv;
1268         }
1269         wake_up_interruptible(&port->svc_wait);
1270
1271         return rv;
1272 }
1273
1274 /*
1275  * Byte-swap ATA ID strings.
1276  *
1277  * ATA identify data contains strings in byte-swapped 16-bit words.
1278  * They must be swapped (on all architectures) to be usable as C strings.
1279  * This function swaps bytes in-place.
1280  *
1281  * @buf The buffer location of the string
1282  * @len The number of bytes to swap
1283  *
1284  * return value
1285  *      None
1286  */
1287 static inline void ata_swap_string(u16 *buf, unsigned int len)
1288 {
1289         int i;
1290         for (i = 0; i < (len/2); i++)
1291                 be16_to_cpus(&buf[i]);
1292 }
1293
1294 static void mtip_set_timeout(struct driver_data *dd,
1295                                         struct host_to_dev_fis *fis,
1296                                         unsigned int *timeout, u8 erasemode)
1297 {
1298         switch (fis->command) {
1299         case ATA_CMD_DOWNLOAD_MICRO:
1300                 *timeout = 120000; /* 2 minutes */
1301                 break;
1302         case ATA_CMD_SEC_ERASE_UNIT:
1303         case 0xFC:
1304                 if (erasemode)
1305                         *timeout = ((*(dd->port->identify + 90) * 2) * 60000);
1306                 else
1307                         *timeout = ((*(dd->port->identify + 89) * 2) * 60000);
1308                 break;
1309         case ATA_CMD_STANDBYNOW1:
1310                 *timeout = 120000;  /* 2 minutes */
1311                 break;
1312         case 0xF7:
1313         case 0xFA:
1314                 *timeout = 60000;  /* 60 seconds */
1315                 break;
1316         case ATA_CMD_SMART:
1317                 *timeout = 15000;  /* 15 seconds */
1318                 break;
1319         default:
1320                 *timeout = MTIP_IOCTL_CMD_TIMEOUT_MS;
1321                 break;
1322         }
1323 }
1324
1325 /*
1326  * Request the device identity information.
1327  *
1328  * If a user space buffer is not specified, i.e. is NULL, the
1329  * identify information is still read from the drive and placed
1330  * into the identify data buffer (@e port->identify) in the
1331  * port data structure.
1332  * When the identify buffer contains valid identify information @e
1333  * port->identify_valid is non-zero.
1334  *
1335  * @port         Pointer to the port structure.
1336  * @user_buffer  A user space buffer where the identify data should be
1337  *                    copied.
1338  *
1339  * return value
1340  *      0       Command completed successfully.
1341  *      -EFAULT An error occurred while coping data to the user buffer.
1342  *      -1      Command failed.
1343  */
1344 static int mtip_get_identify(struct mtip_port *port, void __user *user_buffer)
1345 {
1346         int rv = 0;
1347         struct host_to_dev_fis fis;
1348
1349         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
1350                 return -EFAULT;
1351
1352         /* Build the FIS. */
1353         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1354         fis.type        = 0x27;
1355         fis.opts        = 1 << 7;
1356         fis.command     = ATA_CMD_ID_ATA;
1357
1358         /* Set the identify information as invalid. */
1359         port->identify_valid = 0;
1360
1361         /* Clear the identify information. */
1362         memset(port->identify, 0, sizeof(u16) * ATA_ID_WORDS);
1363
1364         /* Execute the command. */
1365         if (mtip_exec_internal_command(port,
1366                                 &fis,
1367                                 5,
1368                                 port->identify_dma,
1369                                 sizeof(u16) * ATA_ID_WORDS,
1370                                 0,
1371                                 GFP_KERNEL,
1372                                 MTIP_INT_CMD_TIMEOUT_MS)
1373                                 < 0) {
1374                 rv = -1;
1375                 goto out;
1376         }
1377
1378         /*
1379          * Perform any necessary byte-swapping.  Yes, the kernel does in fact
1380          * perform field-sensitive swapping on the string fields.
1381          * See the kernel use of ata_id_string() for proof of this.
1382          */
1383 #ifdef __LITTLE_ENDIAN
1384         ata_swap_string(port->identify + 27, 40);  /* model string*/
1385         ata_swap_string(port->identify + 23, 8);   /* firmware string*/
1386         ata_swap_string(port->identify + 10, 20);  /* serial# string*/
1387 #else
1388         {
1389                 int i;
1390                 for (i = 0; i < ATA_ID_WORDS; i++)
1391                         port->identify[i] = le16_to_cpu(port->identify[i]);
1392         }
1393 #endif
1394
1395         /* Check security locked state */
1396         if (port->identify[128] & 0x4)
1397                 set_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1398         else
1399                 clear_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1400
1401 #ifdef MTIP_TRIM /* Disabling TRIM support temporarily */
1402         /* Demux ID.DRAT & ID.RZAT to determine trim support */
1403         if (port->identify[69] & (1 << 14) && port->identify[69] & (1 << 5))
1404                 port->dd->trim_supp = true;
1405         else
1406 #endif
1407                 port->dd->trim_supp = false;
1408
1409         /* Set the identify buffer as valid. */
1410         port->identify_valid = 1;
1411
1412         if (user_buffer) {
1413                 if (copy_to_user(
1414                         user_buffer,
1415                         port->identify,
1416                         ATA_ID_WORDS * sizeof(u16))) {
1417                         rv = -EFAULT;
1418                         goto out;
1419                 }
1420         }
1421
1422 out:
1423         return rv;
1424 }
1425
1426 /*
1427  * Issue a standby immediate command to the device.
1428  *
1429  * @port Pointer to the port structure.
1430  *
1431  * return value
1432  *      0       Command was executed successfully.
1433  *      -1      An error occurred while executing the command.
1434  */
1435 static int mtip_standby_immediate(struct mtip_port *port)
1436 {
1437         int rv;
1438         struct host_to_dev_fis  fis;
1439         unsigned long start;
1440         unsigned int timeout;
1441
1442         /* Build the FIS. */
1443         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1444         fis.type        = 0x27;
1445         fis.opts        = 1 << 7;
1446         fis.command     = ATA_CMD_STANDBYNOW1;
1447
1448         mtip_set_timeout(port->dd, &fis, &timeout, 0);
1449
1450         start = jiffies;
1451         rv = mtip_exec_internal_command(port,
1452                                         &fis,
1453                                         5,
1454                                         0,
1455                                         0,
1456                                         0,
1457                                         GFP_ATOMIC,
1458                                         timeout);
1459         dbg_printk(MTIP_DRV_NAME "Time taken to complete standby cmd: %d ms\n",
1460                         jiffies_to_msecs(jiffies - start));
1461         if (rv)
1462                 dev_warn(&port->dd->pdev->dev,
1463                         "STANDBY IMMEDIATE command failed.\n");
1464
1465         return rv;
1466 }
1467
1468 /*
1469  * Issue a READ LOG EXT command to the device.
1470  *
1471  * @port        pointer to the port structure.
1472  * @page        page number to fetch
1473  * @buffer      pointer to buffer
1474  * @buffer_dma  dma address corresponding to @buffer
1475  * @sectors     page length to fetch, in sectors
1476  *
1477  * return value
1478  *      @rv     return value from mtip_exec_internal_command()
1479  */
1480 static int mtip_read_log_page(struct mtip_port *port, u8 page, u16 *buffer,
1481                                 dma_addr_t buffer_dma, unsigned int sectors)
1482 {
1483         struct host_to_dev_fis fis;
1484
1485         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1486         fis.type        = 0x27;
1487         fis.opts        = 1 << 7;
1488         fis.command     = ATA_CMD_READ_LOG_EXT;
1489         fis.sect_count  = sectors & 0xFF;
1490         fis.sect_cnt_ex = (sectors >> 8) & 0xFF;
1491         fis.lba_low     = page;
1492         fis.lba_mid     = 0;
1493         fis.device      = ATA_DEVICE_OBS;
1494
1495         memset(buffer, 0, sectors * ATA_SECT_SIZE);
1496
1497         return mtip_exec_internal_command(port,
1498                                         &fis,
1499                                         5,
1500                                         buffer_dma,
1501                                         sectors * ATA_SECT_SIZE,
1502                                         0,
1503                                         GFP_ATOMIC,
1504                                         MTIP_INT_CMD_TIMEOUT_MS);
1505 }
1506
1507 /*
1508  * Issue a SMART READ DATA command to the device.
1509  *
1510  * @port        pointer to the port structure.
1511  * @buffer      pointer to buffer
1512  * @buffer_dma  dma address corresponding to @buffer
1513  *
1514  * return value
1515  *      @rv     return value from mtip_exec_internal_command()
1516  */
1517 static int mtip_get_smart_data(struct mtip_port *port, u8 *buffer,
1518                                         dma_addr_t buffer_dma)
1519 {
1520         struct host_to_dev_fis fis;
1521
1522         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1523         fis.type        = 0x27;
1524         fis.opts        = 1 << 7;
1525         fis.command     = ATA_CMD_SMART;
1526         fis.features    = 0xD0;
1527         fis.sect_count  = 1;
1528         fis.lba_mid     = 0x4F;
1529         fis.lba_hi      = 0xC2;
1530         fis.device      = ATA_DEVICE_OBS;
1531
1532         return mtip_exec_internal_command(port,
1533                                         &fis,
1534                                         5,
1535                                         buffer_dma,
1536                                         ATA_SECT_SIZE,
1537                                         0,
1538                                         GFP_ATOMIC,
1539                                         15000);
1540 }
1541
1542 /*
1543  * Get the value of a smart attribute
1544  *
1545  * @port        pointer to the port structure
1546  * @id          attribute number
1547  * @attrib      pointer to return attrib information corresponding to @id
1548  *
1549  * return value
1550  *      -EINVAL NULL buffer passed or unsupported attribute @id.
1551  *      -EPERM  Identify data not valid, SMART not supported or not enabled
1552  */
1553 static int mtip_get_smart_attr(struct mtip_port *port, unsigned int id,
1554                                                 struct smart_attr *attrib)
1555 {
1556         int rv, i;
1557         struct smart_attr *pattr;
1558
1559         if (!attrib)
1560                 return -EINVAL;
1561
1562         if (!port->identify_valid) {
1563                 dev_warn(&port->dd->pdev->dev, "IDENTIFY DATA not valid\n");
1564                 return -EPERM;
1565         }
1566         if (!(port->identify[82] & 0x1)) {
1567                 dev_warn(&port->dd->pdev->dev, "SMART not supported\n");
1568                 return -EPERM;
1569         }
1570         if (!(port->identify[85] & 0x1)) {
1571                 dev_warn(&port->dd->pdev->dev, "SMART not enabled\n");
1572                 return -EPERM;
1573         }
1574
1575         memset(port->smart_buf, 0, ATA_SECT_SIZE);
1576         rv = mtip_get_smart_data(port, port->smart_buf, port->smart_buf_dma);
1577         if (rv) {
1578                 dev_warn(&port->dd->pdev->dev, "Failed to ge SMART data\n");
1579                 return rv;
1580         }
1581
1582         pattr = (struct smart_attr *)(port->smart_buf + 2);
1583         for (i = 0; i < 29; i++, pattr++)
1584                 if (pattr->attr_id == id) {
1585                         memcpy(attrib, pattr, sizeof(struct smart_attr));
1586                         break;
1587                 }
1588
1589         if (i == 29) {
1590                 dev_warn(&port->dd->pdev->dev,
1591                         "Query for invalid SMART attribute ID\n");
1592                 rv = -EINVAL;
1593         }
1594
1595         return rv;
1596 }
1597
1598 /*
1599  * Trim unused sectors
1600  *
1601  * @dd          pointer to driver_data structure
1602  * @lba         starting lba
1603  * @len         # of 512b sectors to trim
1604  *
1605  * return value
1606  *      -ENOMEM         Out of dma memory
1607  *      -EINVAL         Invalid parameters passed in, trim not supported
1608  *      -EIO            Error submitting trim request to hw
1609  */
1610 static int mtip_send_trim(struct driver_data *dd, unsigned int lba,
1611                                 unsigned int len)
1612 {
1613         int i, rv = 0;
1614         u64 tlba, tlen, sect_left;
1615         struct mtip_trim_entry *buf;
1616         dma_addr_t dma_addr;
1617         struct host_to_dev_fis fis;
1618
1619         if (!len || dd->trim_supp == false)
1620                 return -EINVAL;
1621
1622         /* Trim request too big */
1623         WARN_ON(len > (MTIP_MAX_TRIM_ENTRY_LEN * MTIP_MAX_TRIM_ENTRIES));
1624
1625         /* Trim request not aligned on 4k boundary */
1626         WARN_ON(len % 8 != 0);
1627
1628         /* Warn if vu_trim structure is too big */
1629         WARN_ON(sizeof(struct mtip_trim) > ATA_SECT_SIZE);
1630
1631         /* Allocate a DMA buffer for the trim structure */
1632         buf = dmam_alloc_coherent(&dd->pdev->dev, ATA_SECT_SIZE, &dma_addr,
1633                                                                 GFP_KERNEL);
1634         if (!buf)
1635                 return -ENOMEM;
1636         memset(buf, 0, ATA_SECT_SIZE);
1637
1638         for (i = 0, sect_left = len, tlba = lba;
1639                         i < MTIP_MAX_TRIM_ENTRIES && sect_left;
1640                         i++) {
1641                 tlen = (sect_left >= MTIP_MAX_TRIM_ENTRY_LEN ?
1642                                         MTIP_MAX_TRIM_ENTRY_LEN :
1643                                         sect_left);
1644                 buf[i].lba = __force_bit2int cpu_to_le32(tlba);
1645                 buf[i].range = __force_bit2int cpu_to_le16(tlen);
1646                 tlba += tlen;
1647                 sect_left -= tlen;
1648         }
1649         WARN_ON(sect_left != 0);
1650
1651         /* Build the fis */
1652         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1653         fis.type       = 0x27;
1654         fis.opts       = 1 << 7;
1655         fis.command    = 0xfb;
1656         fis.features   = 0x60;
1657         fis.sect_count = 1;
1658         fis.device     = ATA_DEVICE_OBS;
1659
1660         if (mtip_exec_internal_command(dd->port,
1661                                         &fis,
1662                                         5,
1663                                         dma_addr,
1664                                         ATA_SECT_SIZE,
1665                                         0,
1666                                         GFP_KERNEL,
1667                                         MTIP_TRIM_TIMEOUT_MS) < 0)
1668                 rv = -EIO;
1669
1670         dmam_free_coherent(&dd->pdev->dev, ATA_SECT_SIZE, buf, dma_addr);
1671         return rv;
1672 }
1673
1674 /*
1675  * Get the drive capacity.
1676  *
1677  * @dd      Pointer to the device data structure.
1678  * @sectors Pointer to the variable that will receive the sector count.
1679  *
1680  * return value
1681  *      1 Capacity was returned successfully.
1682  *      0 The identify information is invalid.
1683  */
1684 static bool mtip_hw_get_capacity(struct driver_data *dd, sector_t *sectors)
1685 {
1686         struct mtip_port *port = dd->port;
1687         u64 total, raw0, raw1, raw2, raw3;
1688         raw0 = port->identify[100];
1689         raw1 = port->identify[101];
1690         raw2 = port->identify[102];
1691         raw3 = port->identify[103];
1692         total = raw0 | raw1<<16 | raw2<<32 | raw3<<48;
1693         *sectors = total;
1694         return (bool) !!port->identify_valid;
1695 }
1696
1697 /*
1698  * Display the identify command data.
1699  *
1700  * @port Pointer to the port data structure.
1701  *
1702  * return value
1703  *      None
1704  */
1705 static void mtip_dump_identify(struct mtip_port *port)
1706 {
1707         sector_t sectors;
1708         unsigned short revid;
1709         char cbuf[42];
1710
1711         if (!port->identify_valid)
1712                 return;
1713
1714         strlcpy(cbuf, (char *)(port->identify+10), 21);
1715         dev_info(&port->dd->pdev->dev,
1716                 "Serial No.: %s\n", cbuf);
1717
1718         strlcpy(cbuf, (char *)(port->identify+23), 9);
1719         dev_info(&port->dd->pdev->dev,
1720                 "Firmware Ver.: %s\n", cbuf);
1721
1722         strlcpy(cbuf, (char *)(port->identify+27), 41);
1723         dev_info(&port->dd->pdev->dev, "Model: %s\n", cbuf);
1724
1725         dev_info(&port->dd->pdev->dev, "Security: %04x %s\n",
1726                 port->identify[128],
1727                 port->identify[128] & 0x4 ? "(LOCKED)" : "");
1728
1729         if (mtip_hw_get_capacity(port->dd, &sectors))
1730                 dev_info(&port->dd->pdev->dev,
1731                         "Capacity: %llu sectors (%llu MB)\n",
1732                          (u64)sectors,
1733                          ((u64)sectors) * ATA_SECT_SIZE >> 20);
1734
1735         pci_read_config_word(port->dd->pdev, PCI_REVISION_ID, &revid);
1736         switch (revid & 0xFF) {
1737         case 0x1:
1738                 strlcpy(cbuf, "A0", 3);
1739                 break;
1740         case 0x3:
1741                 strlcpy(cbuf, "A2", 3);
1742                 break;
1743         default:
1744                 strlcpy(cbuf, "?", 2);
1745                 break;
1746         }
1747         dev_info(&port->dd->pdev->dev,
1748                 "Card Type: %s\n", cbuf);
1749 }
1750
1751 /*
1752  * Map the commands scatter list into the command table.
1753  *
1754  * @command Pointer to the command.
1755  * @nents Number of scatter list entries.
1756  *
1757  * return value
1758  *      None
1759  */
1760 static inline void fill_command_sg(struct driver_data *dd,
1761                                 struct mtip_cmd *command,
1762                                 int nents)
1763 {
1764         int n;
1765         unsigned int dma_len;
1766         struct mtip_cmd_sg *command_sg;
1767         struct scatterlist *sg = command->sg;
1768
1769         command_sg = command->command + AHCI_CMD_TBL_HDR_SZ;
1770
1771         for (n = 0; n < nents; n++) {
1772                 dma_len = sg_dma_len(sg);
1773                 if (dma_len > 0x400000)
1774                         dev_err(&dd->pdev->dev,
1775                                 "DMA segment length truncated\n");
1776                 command_sg->info = __force_bit2int
1777                         cpu_to_le32((dma_len-1) & 0x3FFFFF);
1778                 command_sg->dba = __force_bit2int
1779                         cpu_to_le32(sg_dma_address(sg));
1780                 command_sg->dba_upper = __force_bit2int
1781                         cpu_to_le32((sg_dma_address(sg) >> 16) >> 16);
1782                 command_sg++;
1783                 sg++;
1784         }
1785 }
1786
1787 /*
1788  * @brief Execute a drive command.
1789  *
1790  * return value 0 The command completed successfully.
1791  * return value -1 An error occurred while executing the command.
1792  */
1793 static int exec_drive_task(struct mtip_port *port, u8 *command)
1794 {
1795         struct host_to_dev_fis  fis;
1796         struct host_to_dev_fis *reply = (port->rxfis + RX_FIS_D2H_REG);
1797         unsigned int to;
1798
1799         /* Build the FIS. */
1800         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1801         fis.type        = 0x27;
1802         fis.opts        = 1 << 7;
1803         fis.command     = command[0];
1804         fis.features    = command[1];
1805         fis.sect_count  = command[2];
1806         fis.sector      = command[3];
1807         fis.cyl_low     = command[4];
1808         fis.cyl_hi      = command[5];
1809         fis.device      = command[6] & ~0x10; /* Clear the dev bit*/
1810
1811         mtip_set_timeout(port->dd, &fis, &to, 0);
1812
1813         dbg_printk(MTIP_DRV_NAME " %s: User Command: cmd %x, feat %x, nsect %x, sect %x, lcyl %x, hcyl %x, sel %x\n",
1814                 __func__,
1815                 command[0],
1816                 command[1],
1817                 command[2],
1818                 command[3],
1819                 command[4],
1820                 command[5],
1821                 command[6]);
1822
1823         /* Execute the command. */
1824         if (mtip_exec_internal_command(port,
1825                                  &fis,
1826                                  5,
1827                                  0,
1828                                  0,
1829                                  0,
1830                                  GFP_KERNEL,
1831                                  to) < 0) {
1832                 return -1;
1833         }
1834
1835         command[0] = reply->command; /* Status*/
1836         command[1] = reply->features; /* Error*/
1837         command[4] = reply->cyl_low;
1838         command[5] = reply->cyl_hi;
1839
1840         dbg_printk(MTIP_DRV_NAME " %s: Completion Status: stat %x, err %x , cyl_lo %x cyl_hi %x\n",
1841                 __func__,
1842                 command[0],
1843                 command[1],
1844                 command[4],
1845                 command[5]);
1846
1847         return 0;
1848 }
1849
1850 /*
1851  * @brief Execute a drive command.
1852  *
1853  * @param port Pointer to the port data structure.
1854  * @param command Pointer to the user specified command parameters.
1855  * @param user_buffer Pointer to the user space buffer where read sector
1856  *                   data should be copied.
1857  *
1858  * return value 0 The command completed successfully.
1859  * return value -EFAULT An error occurred while copying the completion
1860  *                 data to the user space buffer.
1861  * return value -1 An error occurred while executing the command.
1862  */
1863 static int exec_drive_command(struct mtip_port *port, u8 *command,
1864                                 void __user *user_buffer)
1865 {
1866         struct host_to_dev_fis  fis;
1867         struct host_to_dev_fis *reply;
1868         u8 *buf = NULL;
1869         dma_addr_t dma_addr = 0;
1870         int rv = 0, xfer_sz = command[3];
1871         unsigned int to;
1872
1873         if (xfer_sz) {
1874                 if (!user_buffer)
1875                         return -EFAULT;
1876
1877                 buf = dmam_alloc_coherent(&port->dd->pdev->dev,
1878                                 ATA_SECT_SIZE * xfer_sz,
1879                                 &dma_addr,
1880                                 GFP_KERNEL);
1881                 if (!buf) {
1882                         dev_err(&port->dd->pdev->dev,
1883                                 "Memory allocation failed (%d bytes)\n",
1884                                 ATA_SECT_SIZE * xfer_sz);
1885                         return -ENOMEM;
1886                 }
1887                 memset(buf, 0, ATA_SECT_SIZE * xfer_sz);
1888         }
1889
1890         /* Build the FIS. */
1891         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1892         fis.type        = 0x27;
1893         fis.opts        = 1 << 7;
1894         fis.command     = command[0];
1895         fis.features    = command[2];
1896         fis.sect_count  = command[3];
1897         if (fis.command == ATA_CMD_SMART) {
1898                 fis.sector      = command[1];
1899                 fis.cyl_low     = 0x4F;
1900                 fis.cyl_hi      = 0xC2;
1901         }
1902
1903         mtip_set_timeout(port->dd, &fis, &to, 0);
1904
1905         if (xfer_sz)
1906                 reply = (port->rxfis + RX_FIS_PIO_SETUP);
1907         else
1908                 reply = (port->rxfis + RX_FIS_D2H_REG);
1909
1910         dbg_printk(MTIP_DRV_NAME
1911                 " %s: User Command: cmd %x, sect %x, "
1912                 "feat %x, sectcnt %x\n",
1913                 __func__,
1914                 command[0],
1915                 command[1],
1916                 command[2],
1917                 command[3]);
1918
1919         /* Execute the command. */
1920         if (mtip_exec_internal_command(port,
1921                                 &fis,
1922                                  5,
1923                                  (xfer_sz ? dma_addr : 0),
1924                                  (xfer_sz ? ATA_SECT_SIZE * xfer_sz : 0),
1925                                  0,
1926                                  GFP_KERNEL,
1927                                  to)
1928                                  < 0) {
1929                 rv = -EFAULT;
1930                 goto exit_drive_command;
1931         }
1932
1933         /* Collect the completion status. */
1934         command[0] = reply->command; /* Status*/
1935         command[1] = reply->features; /* Error*/
1936         command[2] = reply->sect_count;
1937
1938         dbg_printk(MTIP_DRV_NAME
1939                 " %s: Completion Status: stat %x, "
1940                 "err %x, nsect %x\n",
1941                 __func__,
1942                 command[0],
1943                 command[1],
1944                 command[2]);
1945
1946         if (xfer_sz) {
1947                 if (copy_to_user(user_buffer,
1948                                  buf,
1949                                  ATA_SECT_SIZE * command[3])) {
1950                         rv = -EFAULT;
1951                         goto exit_drive_command;
1952                 }
1953         }
1954 exit_drive_command:
1955         if (buf)
1956                 dmam_free_coherent(&port->dd->pdev->dev,
1957                                 ATA_SECT_SIZE * xfer_sz, buf, dma_addr);
1958         return rv;
1959 }
1960
1961 /*
1962  *  Indicates whether a command has a single sector payload.
1963  *
1964  *  @command passed to the device to perform the certain event.
1965  *  @features passed to the device to perform the certain event.
1966  *
1967  *  return value
1968  *      1       command is one that always has a single sector payload,
1969  *              regardless of the value in the Sector Count field.
1970  *      0       otherwise
1971  *
1972  */
1973 static unsigned int implicit_sector(unsigned char command,
1974                                     unsigned char features)
1975 {
1976         unsigned int rv = 0;
1977
1978         /* list of commands that have an implicit sector count of 1 */
1979         switch (command) {
1980         case ATA_CMD_SEC_SET_PASS:
1981         case ATA_CMD_SEC_UNLOCK:
1982         case ATA_CMD_SEC_ERASE_PREP:
1983         case ATA_CMD_SEC_ERASE_UNIT:
1984         case ATA_CMD_SEC_FREEZE_LOCK:
1985         case ATA_CMD_SEC_DISABLE_PASS:
1986         case ATA_CMD_PMP_READ:
1987         case ATA_CMD_PMP_WRITE:
1988                 rv = 1;
1989                 break;
1990         case ATA_CMD_SET_MAX:
1991                 if (features == ATA_SET_MAX_UNLOCK)
1992                         rv = 1;
1993                 break;
1994         case ATA_CMD_SMART:
1995                 if ((features == ATA_SMART_READ_VALUES) ||
1996                                 (features == ATA_SMART_READ_THRESHOLDS))
1997                         rv = 1;
1998                 break;
1999         case ATA_CMD_CONF_OVERLAY:
2000                 if ((features == ATA_DCO_IDENTIFY) ||
2001                                 (features == ATA_DCO_SET))
2002                         rv = 1;
2003                 break;
2004         }
2005         return rv;
2006 }
2007
2008 /*
2009  * Executes a taskfile
2010  * See ide_taskfile_ioctl() for derivation
2011  */
2012 static int exec_drive_taskfile(struct driver_data *dd,
2013                                void __user *buf,
2014                                ide_task_request_t *req_task,
2015                                int outtotal)
2016 {
2017         struct host_to_dev_fis  fis;
2018         struct host_to_dev_fis *reply;
2019         u8 *outbuf = NULL;
2020         u8 *inbuf = NULL;
2021         dma_addr_t outbuf_dma = 0;
2022         dma_addr_t inbuf_dma = 0;
2023         dma_addr_t dma_buffer = 0;
2024         int err = 0;
2025         unsigned int taskin = 0;
2026         unsigned int taskout = 0;
2027         u8 nsect = 0;
2028         unsigned int timeout;
2029         unsigned int force_single_sector;
2030         unsigned int transfer_size;
2031         unsigned long task_file_data;
2032         int intotal = outtotal + req_task->out_size;
2033         int erasemode = 0;
2034
2035         taskout = req_task->out_size;
2036         taskin = req_task->in_size;
2037         /* 130560 = 512 * 0xFF*/
2038         if (taskin > 130560 || taskout > 130560)
2039                 return -EINVAL;
2040
2041         if (taskout) {
2042                 outbuf = memdup_user(buf + outtotal, taskout);
2043                 if (IS_ERR(outbuf))
2044                         return PTR_ERR(outbuf);
2045
2046                 outbuf_dma = pci_map_single(dd->pdev,
2047                                          outbuf,
2048                                          taskout,
2049                                          DMA_TO_DEVICE);
2050                 if (pci_dma_mapping_error(dd->pdev, outbuf_dma)) {
2051                         err = -ENOMEM;
2052                         goto abort;
2053                 }
2054                 dma_buffer = outbuf_dma;
2055         }
2056
2057         if (taskin) {
2058                 inbuf = memdup_user(buf + intotal, taskin);
2059                 if (IS_ERR(inbuf)) {
2060                         err = PTR_ERR(inbuf);
2061                         inbuf = NULL;
2062                         goto abort;
2063                 }
2064                 inbuf_dma = pci_map_single(dd->pdev,
2065                                          inbuf,
2066                                          taskin, DMA_FROM_DEVICE);
2067                 if (pci_dma_mapping_error(dd->pdev, inbuf_dma)) {
2068                         err = -ENOMEM;
2069                         goto abort;
2070                 }
2071                 dma_buffer = inbuf_dma;
2072         }
2073
2074         /* only supports PIO and non-data commands from this ioctl. */
2075         switch (req_task->data_phase) {
2076         case TASKFILE_OUT:
2077                 nsect = taskout / ATA_SECT_SIZE;
2078                 reply = (dd->port->rxfis + RX_FIS_PIO_SETUP);
2079                 break;
2080         case TASKFILE_IN:
2081                 reply = (dd->port->rxfis + RX_FIS_PIO_SETUP);
2082                 break;
2083         case TASKFILE_NO_DATA:
2084                 reply = (dd->port->rxfis + RX_FIS_D2H_REG);
2085                 break;
2086         default:
2087                 err = -EINVAL;
2088                 goto abort;
2089         }
2090
2091         /* Build the FIS. */
2092         memset(&fis, 0, sizeof(struct host_to_dev_fis));
2093
2094         fis.type        = 0x27;
2095         fis.opts        = 1 << 7;
2096         fis.command     = req_task->io_ports[7];
2097         fis.features    = req_task->io_ports[1];
2098         fis.sect_count  = req_task->io_ports[2];
2099         fis.lba_low     = req_task->io_ports[3];
2100         fis.lba_mid     = req_task->io_ports[4];
2101         fis.lba_hi      = req_task->io_ports[5];
2102          /* Clear the dev bit*/
2103         fis.device      = req_task->io_ports[6] & ~0x10;
2104
2105         if ((req_task->in_flags.all == 0) && (req_task->out_flags.all & 1)) {
2106                 req_task->in_flags.all  =
2107                         IDE_TASKFILE_STD_IN_FLAGS |
2108                         (IDE_HOB_STD_IN_FLAGS << 8);
2109                 fis.lba_low_ex          = req_task->hob_ports[3];
2110                 fis.lba_mid_ex          = req_task->hob_ports[4];
2111                 fis.lba_hi_ex           = req_task->hob_ports[5];
2112                 fis.features_ex         = req_task->hob_ports[1];
2113                 fis.sect_cnt_ex         = req_task->hob_ports[2];
2114
2115         } else {
2116                 req_task->in_flags.all = IDE_TASKFILE_STD_IN_FLAGS;
2117         }
2118
2119         force_single_sector = implicit_sector(fis.command, fis.features);
2120
2121         if ((taskin || taskout) && (!fis.sect_count)) {
2122                 if (nsect)
2123                         fis.sect_count = nsect;
2124                 else {
2125                         if (!force_single_sector) {
2126                                 dev_warn(&dd->pdev->dev,
2127                                         "data movement but "
2128                                         "sect_count is 0\n");
2129                                         err = -EINVAL;
2130                                         goto abort;
2131                         }
2132                 }
2133         }
2134
2135         dbg_printk(MTIP_DRV_NAME
2136                 " %s: cmd %x, feat %x, nsect %x,"
2137                 " sect/lbal %x, lcyl/lbam %x, hcyl/lbah %x,"
2138                 " head/dev %x\n",
2139                 __func__,
2140                 fis.command,
2141                 fis.features,
2142                 fis.sect_count,
2143                 fis.lba_low,
2144                 fis.lba_mid,
2145                 fis.lba_hi,
2146                 fis.device);
2147
2148         /* check for erase mode support during secure erase.*/
2149         if ((fis.command == ATA_CMD_SEC_ERASE_UNIT) && outbuf &&
2150                                         (outbuf[0] & MTIP_SEC_ERASE_MODE)) {
2151                 erasemode = 1;
2152         }
2153
2154         mtip_set_timeout(dd, &fis, &timeout, erasemode);
2155
2156         /* Determine the correct transfer size.*/
2157         if (force_single_sector)
2158                 transfer_size = ATA_SECT_SIZE;
2159         else
2160                 transfer_size = ATA_SECT_SIZE * fis.sect_count;
2161
2162         /* Execute the command.*/
2163         if (mtip_exec_internal_command(dd->port,
2164                                  &fis,
2165                                  5,
2166                                  dma_buffer,
2167                                  transfer_size,
2168                                  0,
2169                                  GFP_KERNEL,
2170                                  timeout) < 0) {
2171                 err = -EIO;
2172                 goto abort;
2173         }
2174
2175         task_file_data = readl(dd->port->mmio+PORT_TFDATA);
2176
2177         if ((req_task->data_phase == TASKFILE_IN) && !(task_file_data & 1)) {
2178                 reply = dd->port->rxfis + RX_FIS_PIO_SETUP;
2179                 req_task->io_ports[7] = reply->control;
2180         } else {
2181                 reply = dd->port->rxfis + RX_FIS_D2H_REG;
2182                 req_task->io_ports[7] = reply->command;
2183         }
2184
2185         /* reclaim the DMA buffers.*/
2186         if (inbuf_dma)
2187                 pci_unmap_single(dd->pdev, inbuf_dma,
2188                         taskin, DMA_FROM_DEVICE);
2189         if (outbuf_dma)
2190                 pci_unmap_single(dd->pdev, outbuf_dma,
2191                         taskout, DMA_TO_DEVICE);
2192         inbuf_dma  = 0;
2193         outbuf_dma = 0;
2194
2195         /* return the ATA registers to the caller.*/
2196         req_task->io_ports[1] = reply->features;
2197         req_task->io_ports[2] = reply->sect_count;
2198         req_task->io_ports[3] = reply->lba_low;
2199         req_task->io_ports[4] = reply->lba_mid;
2200         req_task->io_ports[5] = reply->lba_hi;
2201         req_task->io_ports[6] = reply->device;
2202
2203         if (req_task->out_flags.all & 1)  {
2204
2205                 req_task->hob_ports[3] = reply->lba_low_ex;
2206                 req_task->hob_ports[4] = reply->lba_mid_ex;
2207                 req_task->hob_ports[5] = reply->lba_hi_ex;
2208                 req_task->hob_ports[1] = reply->features_ex;
2209                 req_task->hob_ports[2] = reply->sect_cnt_ex;
2210         }
2211         dbg_printk(MTIP_DRV_NAME
2212                 " %s: Completion: stat %x,"
2213                 "err %x, sect_cnt %x, lbalo %x,"
2214                 "lbamid %x, lbahi %x, dev %x\n",
2215                 __func__,
2216                 req_task->io_ports[7],
2217                 req_task->io_ports[1],
2218                 req_task->io_ports[2],
2219                 req_task->io_ports[3],
2220                 req_task->io_ports[4],
2221                 req_task->io_ports[5],
2222                 req_task->io_ports[6]);
2223
2224         if (taskout) {
2225                 if (copy_to_user(buf + outtotal, outbuf, taskout)) {
2226                         err = -EFAULT;
2227                         goto abort;
2228                 }
2229         }
2230         if (taskin) {
2231                 if (copy_to_user(buf + intotal, inbuf, taskin)) {
2232                         err = -EFAULT;
2233                         goto abort;
2234                 }
2235         }
2236 abort:
2237         if (inbuf_dma)
2238                 pci_unmap_single(dd->pdev, inbuf_dma,
2239                                         taskin, DMA_FROM_DEVICE);
2240         if (outbuf_dma)
2241                 pci_unmap_single(dd->pdev, outbuf_dma,
2242                                         taskout, DMA_TO_DEVICE);
2243         kfree(outbuf);
2244         kfree(inbuf);
2245
2246         return err;
2247 }
2248
2249 /*
2250  * Handle IOCTL calls from the Block Layer.
2251  *
2252  * This function is called by the Block Layer when it receives an IOCTL
2253  * command that it does not understand. If the IOCTL command is not supported
2254  * this function returns -ENOTTY.
2255  *
2256  * @dd  Pointer to the driver data structure.
2257  * @cmd IOCTL command passed from the Block Layer.
2258  * @arg IOCTL argument passed from the Block Layer.
2259  *
2260  * return value
2261  *      0       The IOCTL completed successfully.
2262  *      -ENOTTY The specified command is not supported.
2263  *      -EFAULT An error occurred copying data to a user space buffer.
2264  *      -EIO    An error occurred while executing the command.
2265  */
2266 static int mtip_hw_ioctl(struct driver_data *dd, unsigned int cmd,
2267                          unsigned long arg)
2268 {
2269         switch (cmd) {
2270         case HDIO_GET_IDENTITY:
2271         {
2272                 if (copy_to_user((void __user *)arg, dd->port->identify,
2273                                                 sizeof(u16) * ATA_ID_WORDS))
2274                         return -EFAULT;
2275                 break;
2276         }
2277         case HDIO_DRIVE_CMD:
2278         {
2279                 u8 drive_command[4];
2280
2281                 /* Copy the user command info to our buffer. */
2282                 if (copy_from_user(drive_command,
2283                                          (void __user *) arg,
2284                                          sizeof(drive_command)))
2285                         return -EFAULT;
2286
2287                 /* Execute the drive command. */
2288                 if (exec_drive_command(dd->port,
2289                                          drive_command,
2290                                          (void __user *) (arg+4)))
2291                         return -EIO;
2292
2293                 /* Copy the status back to the users buffer. */
2294                 if (copy_to_user((void __user *) arg,
2295                                          drive_command,
2296                                          sizeof(drive_command)))
2297                         return -EFAULT;
2298
2299                 break;
2300         }
2301         case HDIO_DRIVE_TASK:
2302         {
2303                 u8 drive_command[7];
2304
2305                 /* Copy the user command info to our buffer. */
2306                 if (copy_from_user(drive_command,
2307                                          (void __user *) arg,
2308                                          sizeof(drive_command)))
2309                         return -EFAULT;
2310
2311                 /* Execute the drive command. */
2312                 if (exec_drive_task(dd->port, drive_command))
2313                         return -EIO;
2314
2315                 /* Copy the status back to the users buffer. */
2316                 if (copy_to_user((void __user *) arg,
2317                                          drive_command,
2318                                          sizeof(drive_command)))
2319                         return -EFAULT;
2320
2321                 break;
2322         }
2323         case HDIO_DRIVE_TASKFILE: {
2324                 ide_task_request_t req_task;
2325                 int ret, outtotal;
2326
2327                 if (copy_from_user(&req_task, (void __user *) arg,
2328                                         sizeof(req_task)))
2329                         return -EFAULT;
2330
2331                 outtotal = sizeof(req_task);
2332
2333                 ret = exec_drive_taskfile(dd, (void __user *) arg,
2334                                                 &req_task, outtotal);
2335
2336                 if (copy_to_user((void __user *) arg, &req_task,
2337                                                         sizeof(req_task)))
2338                         return -EFAULT;
2339
2340                 return ret;
2341         }
2342
2343         default:
2344                 return -EINVAL;
2345         }
2346         return 0;
2347 }
2348
2349 /*
2350  * Submit an IO to the hw
2351  *
2352  * This function is called by the block layer to issue an io
2353  * to the device. Upon completion, the callback function will
2354  * be called with the data parameter passed as the callback data.
2355  *
2356  * @dd       Pointer to the driver data structure.
2357  * @start    First sector to read.
2358  * @nsect    Number of sectors to read.
2359  * @nents    Number of entries in scatter list for the read command.
2360  * @tag      The tag of this read command.
2361  * @callback Pointer to the function that should be called
2362  *           when the read completes.
2363  * @data     Callback data passed to the callback function
2364  *           when the read completes.
2365  * @dir      Direction (read or write)
2366  *
2367  * return value
2368  *      None
2369  */
2370 static void mtip_hw_submit_io(struct driver_data *dd, struct request *rq,
2371                               struct mtip_cmd *command, int nents,
2372                               struct blk_mq_hw_ctx *hctx)
2373 {
2374         struct host_to_dev_fis  *fis;
2375         struct mtip_port *port = dd->port;
2376         int dma_dir = rq_data_dir(rq) == READ ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
2377         u64 start = blk_rq_pos(rq);
2378         unsigned int nsect = blk_rq_sectors(rq);
2379
2380         /* Map the scatter list for DMA access */
2381         nents = dma_map_sg(&dd->pdev->dev, command->sg, nents, dma_dir);
2382
2383         prefetch(&port->flags);
2384
2385         command->scatter_ents = nents;
2386
2387         /*
2388          * The number of retries for this command before it is
2389          * reported as a failure to the upper layers.
2390          */
2391         command->retries = MTIP_MAX_RETRIES;
2392
2393         /* Fill out fis */
2394         fis = command->command;
2395         fis->type        = 0x27;
2396         fis->opts        = 1 << 7;
2397         if (dma_dir == DMA_FROM_DEVICE)
2398                 fis->command = ATA_CMD_FPDMA_READ;
2399         else
2400                 fis->command = ATA_CMD_FPDMA_WRITE;
2401         fis->lba_low     = start & 0xFF;
2402         fis->lba_mid     = (start >> 8) & 0xFF;
2403         fis->lba_hi      = (start >> 16) & 0xFF;
2404         fis->lba_low_ex  = (start >> 24) & 0xFF;
2405         fis->lba_mid_ex  = (start >> 32) & 0xFF;
2406         fis->lba_hi_ex   = (start >> 40) & 0xFF;
2407         fis->device      = 1 << 6;
2408         fis->features    = nsect & 0xFF;
2409         fis->features_ex = (nsect >> 8) & 0xFF;
2410         fis->sect_count  = ((rq->tag << 3) | (rq->tag >> 5));
2411         fis->sect_cnt_ex = 0;
2412         fis->control     = 0;
2413         fis->res2        = 0;
2414         fis->res3        = 0;
2415         fill_command_sg(dd, command, nents);
2416
2417         if (unlikely(command->unaligned))
2418                 fis->device |= 1 << 7;
2419
2420         /* Populate the command header */
2421         command->command_header->opts =
2422                         __force_bit2int cpu_to_le32(
2423                                 (nents << 16) | 5 | AHCI_CMD_PREFETCH);
2424         command->command_header->byte_count = 0;
2425
2426         /*
2427          * Set the completion function and data for the command
2428          * within this layer.
2429          */
2430         command->comp_data = dd;
2431         command->comp_func = mtip_async_complete;
2432         command->direction = dma_dir;
2433
2434         /*
2435          * To prevent this command from being issued
2436          * if an internal command is in progress or error handling is active.
2437          */
2438         if (unlikely(port->flags & MTIP_PF_PAUSE_IO)) {
2439                 set_bit(rq->tag, port->cmds_to_issue);
2440                 set_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags);
2441                 return;
2442         }
2443
2444         /* Issue the command to the hardware */
2445         mtip_issue_ncq_command(port, rq->tag);
2446 }
2447
2448 /*
2449  * Sysfs status dump.
2450  *
2451  * @dev  Pointer to the device structure, passed by the kernrel.
2452  * @attr Pointer to the device_attribute structure passed by the kernel.
2453  * @buf  Pointer to the char buffer that will receive the stats info.
2454  *
2455  * return value
2456  *      The size, in bytes, of the data copied into buf.
2457  */
2458 static ssize_t mtip_hw_show_status(struct device *dev,
2459                                 struct device_attribute *attr,
2460                                 char *buf)
2461 {
2462         struct driver_data *dd = dev_to_disk(dev)->private_data;
2463         int size = 0;
2464
2465         if (test_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag))
2466                 size += sprintf(buf, "%s", "thermal_shutdown\n");
2467         else if (test_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag))
2468                 size += sprintf(buf, "%s", "write_protect\n");
2469         else
2470                 size += sprintf(buf, "%s", "online\n");
2471
2472         return size;
2473 }
2474
2475 static DEVICE_ATTR(status, S_IRUGO, mtip_hw_show_status, NULL);
2476
2477 /* debugsfs entries */
2478
2479 static ssize_t show_device_status(struct device_driver *drv, char *buf)
2480 {
2481         int size = 0;
2482         struct driver_data *dd, *tmp;
2483         unsigned long flags;
2484         char id_buf[42];
2485         u16 status = 0;
2486
2487         spin_lock_irqsave(&dev_lock, flags);
2488         size += sprintf(&buf[size], "Devices Present:\n");
2489         list_for_each_entry_safe(dd, tmp, &online_list, online_list) {
2490                 if (dd->pdev) {
2491                         if (dd->port &&
2492                             dd->port->identify &&
2493                             dd->port->identify_valid) {
2494                                 strlcpy(id_buf,
2495                                         (char *) (dd->port->identify + 10), 21);
2496                                 status = *(dd->port->identify + 141);
2497                         } else {
2498                                 memset(id_buf, 0, 42);
2499                                 status = 0;
2500                         }
2501
2502                         if (dd->port &&
2503                             test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags)) {
2504                                 size += sprintf(&buf[size],
2505                                         " device %s %s (ftl rebuild %d %%)\n",
2506                                         dev_name(&dd->pdev->dev),
2507                                         id_buf,
2508                                         status);
2509                         } else {
2510                                 size += sprintf(&buf[size],
2511                                         " device %s %s\n",
2512                                         dev_name(&dd->pdev->dev),
2513                                         id_buf);
2514                         }
2515                 }
2516         }
2517
2518         size += sprintf(&buf[size], "Devices Being Removed:\n");
2519         list_for_each_entry_safe(dd, tmp, &removing_list, remove_list) {
2520                 if (dd->pdev) {
2521                         if (dd->port &&
2522                             dd->port->identify &&
2523                             dd->port->identify_valid) {
2524                                 strlcpy(id_buf,
2525                                         (char *) (dd->port->identify+10), 21);
2526                                 status = *(dd->port->identify + 141);
2527                         } else {
2528                                 memset(id_buf, 0, 42);
2529                                 status = 0;
2530                         }
2531
2532                         if (dd->port &&
2533                             test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags)) {
2534                                 size += sprintf(&buf[size],
2535                                         " device %s %s (ftl rebuild %d %%)\n",
2536                                         dev_name(&dd->pdev->dev),
2537                                         id_buf,
2538                                         status);
2539                         } else {
2540                                 size += sprintf(&buf[size],
2541                                         " device %s %s\n",
2542                                         dev_name(&dd->pdev->dev),
2543                                         id_buf);
2544                         }
2545                 }
2546         }
2547         spin_unlock_irqrestore(&dev_lock, flags);
2548
2549         return size;
2550 }
2551
2552 static ssize_t mtip_hw_read_device_status(struct file *f, char __user *ubuf,
2553                                                 size_t len, loff_t *offset)
2554 {
2555         struct driver_data *dd =  (struct driver_data *)f->private_data;
2556         int size = *offset;
2557         char *buf;
2558         int rv = 0;
2559
2560         if (!len || *offset)
2561                 return 0;
2562
2563         buf = kzalloc(MTIP_DFS_MAX_BUF_SIZE, GFP_KERNEL);
2564         if (!buf) {
2565                 dev_err(&dd->pdev->dev,
2566                         "Memory allocation: status buffer\n");
2567                 return -ENOMEM;
2568         }
2569
2570         size += show_device_status(NULL, buf);
2571
2572         *offset = size <= len ? size : len;
2573         size = copy_to_user(ubuf, buf, *offset);
2574         if (size)
2575                 rv = -EFAULT;
2576
2577         kfree(buf);
2578         return rv ? rv : *offset;
2579 }
2580
2581 static ssize_t mtip_hw_read_registers(struct file *f, char __user *ubuf,
2582                                   size_t len, loff_t *offset)
2583 {
2584         struct driver_data *dd =  (struct driver_data *)f->private_data;
2585         char *buf;
2586         u32 group_allocated;
2587         int size = *offset;
2588         int n, rv = 0;
2589
2590         if (!len || size)
2591                 return 0;
2592
2593         buf = kzalloc(MTIP_DFS_MAX_BUF_SIZE, GFP_KERNEL);
2594         if (!buf) {
2595                 dev_err(&dd->pdev->dev,
2596                         "Memory allocation: register buffer\n");
2597                 return -ENOMEM;
2598         }
2599
2600         size += sprintf(&buf[size], "H/ S ACTive      : [ 0x");
2601
2602         for (n = dd->slot_groups-1; n >= 0; n--)
2603                 size += sprintf(&buf[size], "%08X ",
2604                                          readl(dd->port->s_active[n]));
2605
2606         size += sprintf(&buf[size], "]\n");
2607         size += sprintf(&buf[size], "H/ Command Issue : [ 0x");
2608
2609         for (n = dd->slot_groups-1; n >= 0; n--)
2610                 size += sprintf(&buf[size], "%08X ",
2611                                         readl(dd->port->cmd_issue[n]));
2612
2613         size += sprintf(&buf[size], "]\n");
2614         size += sprintf(&buf[size], "H/ Completed     : [ 0x");
2615
2616         for (n = dd->slot_groups-1; n >= 0; n--)
2617                 size += sprintf(&buf[size], "%08X ",
2618                                 readl(dd->port->completed[n]));
2619
2620         size += sprintf(&buf[size], "]\n");
2621         size += sprintf(&buf[size], "H/ PORT IRQ STAT : [ 0x%08X ]\n",
2622                                 readl(dd->port->mmio + PORT_IRQ_STAT));
2623         size += sprintf(&buf[size], "H/ HOST IRQ STAT : [ 0x%08X ]\n",
2624                                 readl(dd->mmio + HOST_IRQ_STAT));
2625         size += sprintf(&buf[size], "\n");
2626
2627         size += sprintf(&buf[size], "L/ Commands in Q : [ 0x");
2628
2629         for (n = dd->slot_groups-1; n >= 0; n--) {
2630                 if (sizeof(long) > sizeof(u32))
2631                         group_allocated =
2632                                 dd->port->cmds_to_issue[n/2] >> (32*(n&1));
2633                 else
2634                         group_allocated = dd->port->cmds_to_issue[n];
2635                 size += sprintf(&buf[size], "%08X ", group_allocated);
2636         }
2637         size += sprintf(&buf[size], "]\n");
2638
2639         *offset = size <= len ? size : len;
2640         size = copy_to_user(ubuf, buf, *offset);
2641         if (size)
2642                 rv = -EFAULT;
2643
2644         kfree(buf);
2645         return rv ? rv : *offset;
2646 }
2647
2648 static ssize_t mtip_hw_read_flags(struct file *f, char __user *ubuf,
2649                                   size_t len, loff_t *offset)
2650 {
2651         struct driver_data *dd =  (struct driver_data *)f->private_data;
2652         char *buf;
2653         int size = *offset;
2654         int rv = 0;
2655
2656         if (!len || size)
2657                 return 0;
2658
2659         buf = kzalloc(MTIP_DFS_MAX_BUF_SIZE, GFP_KERNEL);
2660         if (!buf) {
2661                 dev_err(&dd->pdev->dev,
2662                         "Memory allocation: flag buffer\n");
2663                 return -ENOMEM;
2664         }
2665
2666         size += sprintf(&buf[size], "Flag-port : [ %08lX ]\n",
2667                                                         dd->port->flags);
2668         size += sprintf(&buf[size], "Flag-dd   : [ %08lX ]\n",
2669                                                         dd->dd_flag);
2670
2671         *offset = size <= len ? size : len;
2672         size = copy_to_user(ubuf, buf, *offset);
2673         if (size)
2674                 rv = -EFAULT;
2675
2676         kfree(buf);
2677         return rv ? rv : *offset;
2678 }
2679
2680 static const struct file_operations mtip_device_status_fops = {
2681         .owner  = THIS_MODULE,
2682         .open   = simple_open,
2683         .read   = mtip_hw_read_device_status,
2684         .llseek = no_llseek,
2685 };
2686
2687 static const struct file_operations mtip_regs_fops = {
2688         .owner  = THIS_MODULE,
2689         .open   = simple_open,
2690         .read   = mtip_hw_read_registers,
2691         .llseek = no_llseek,
2692 };
2693
2694 static const struct file_operations mtip_flags_fops = {
2695         .owner  = THIS_MODULE,
2696         .open   = simple_open,
2697         .read   = mtip_hw_read_flags,
2698         .llseek = no_llseek,
2699 };
2700
2701 /*
2702  * Create the sysfs related attributes.
2703  *
2704  * @dd   Pointer to the driver data structure.
2705  * @kobj Pointer to the kobj for the block device.
2706  *
2707  * return value
2708  *      0       Operation completed successfully.
2709  *      -EINVAL Invalid parameter.
2710  */
2711 static int mtip_hw_sysfs_init(struct driver_data *dd, struct kobject *kobj)
2712 {
2713         if (!kobj || !dd)
2714                 return -EINVAL;
2715
2716         if (sysfs_create_file(kobj, &dev_attr_status.attr))
2717                 dev_warn(&dd->pdev->dev,
2718                         "Error creating 'status' sysfs entry\n");
2719         return 0;
2720 }
2721
2722 /*
2723  * Remove the sysfs related attributes.
2724  *
2725  * @dd   Pointer to the driver data structure.
2726  * @kobj Pointer to the kobj for the block device.
2727  *
2728  * return value
2729  *      0       Operation completed successfully.
2730  *      -EINVAL Invalid parameter.
2731  */
2732 static int mtip_hw_sysfs_exit(struct driver_data *dd, struct kobject *kobj)
2733 {
2734         if (!kobj || !dd)
2735                 return -EINVAL;
2736
2737         sysfs_remove_file(kobj, &dev_attr_status.attr);
2738
2739         return 0;
2740 }
2741
2742 static int mtip_hw_debugfs_init(struct driver_data *dd)
2743 {
2744         if (!dfs_parent)
2745                 return -1;
2746
2747         dd->dfs_node = debugfs_create_dir(dd->disk->disk_name, dfs_parent);
2748         if (IS_ERR_OR_NULL(dd->dfs_node)) {
2749                 dev_warn(&dd->pdev->dev,
2750                         "Error creating node %s under debugfs\n",
2751                                                 dd->disk->disk_name);
2752                 dd->dfs_node = NULL;
2753                 return -1;
2754         }
2755
2756         debugfs_create_file("flags", S_IRUGO, dd->dfs_node, dd,
2757                                                         &mtip_flags_fops);
2758         debugfs_create_file("registers", S_IRUGO, dd->dfs_node, dd,
2759                                                         &mtip_regs_fops);
2760
2761         return 0;
2762 }
2763
2764 static void mtip_hw_debugfs_exit(struct driver_data *dd)
2765 {
2766         if (dd->dfs_node)
2767                 debugfs_remove_recursive(dd->dfs_node);
2768 }
2769
2770 /*
2771  * Perform any init/resume time hardware setup
2772  *
2773  * @dd Pointer to the driver data structure.
2774  *
2775  * return value
2776  *      None
2777  */
2778 static inline void hba_setup(struct driver_data *dd)
2779 {
2780         u32 hwdata;
2781         hwdata = readl(dd->mmio + HOST_HSORG);
2782
2783         /* interrupt bug workaround: use only 1 IS bit.*/
2784         writel(hwdata |
2785                 HSORG_DISABLE_SLOTGRP_INTR |
2786                 HSORG_DISABLE_SLOTGRP_PXIS,
2787                 dd->mmio + HOST_HSORG);
2788 }
2789
2790 static int mtip_device_unaligned_constrained(struct driver_data *dd)
2791 {
2792         return (dd->pdev->device == P420M_DEVICE_ID ? 1 : 0);
2793 }
2794
2795 /*
2796  * Detect the details of the product, and store anything needed
2797  * into the driver data structure.  This includes product type and
2798  * version and number of slot groups.
2799  *
2800  * @dd Pointer to the driver data structure.
2801  *
2802  * return value
2803  *      None
2804  */
2805 static void mtip_detect_product(struct driver_data *dd)
2806 {
2807         u32 hwdata;
2808         unsigned int rev, slotgroups;
2809
2810         /*
2811          * HBA base + 0xFC [15:0] - vendor-specific hardware interface
2812          * info register:
2813          * [15:8] hardware/software interface rev#
2814          * [   3] asic-style interface
2815          * [ 2:0] number of slot groups, minus 1 (only valid for asic-style).
2816          */
2817         hwdata = readl(dd->mmio + HOST_HSORG);
2818
2819         dd->product_type = MTIP_PRODUCT_UNKNOWN;
2820         dd->slot_groups = 1;
2821
2822         if (hwdata & 0x8) {
2823                 dd->product_type = MTIP_PRODUCT_ASICFPGA;
2824                 rev = (hwdata & HSORG_HWREV) >> 8;
2825                 slotgroups = (hwdata & HSORG_SLOTGROUPS) + 1;
2826                 dev_info(&dd->pdev->dev,
2827                         "ASIC-FPGA design, HS rev 0x%x, "
2828                         "%i slot groups [%i slots]\n",
2829                          rev,
2830                          slotgroups,
2831                          slotgroups * 32);
2832
2833                 if (slotgroups > MTIP_MAX_SLOT_GROUPS) {
2834                         dev_warn(&dd->pdev->dev,
2835                                 "Warning: driver only supports "
2836                                 "%i slot groups.\n", MTIP_MAX_SLOT_GROUPS);
2837                         slotgroups = MTIP_MAX_SLOT_GROUPS;
2838                 }
2839                 dd->slot_groups = slotgroups;
2840                 return;
2841         }
2842
2843         dev_warn(&dd->pdev->dev, "Unrecognized product id\n");
2844 }
2845
2846 /*
2847  * Blocking wait for FTL rebuild to complete
2848  *
2849  * @dd Pointer to the DRIVER_DATA structure.
2850  *
2851  * return value
2852  *      0       FTL rebuild completed successfully
2853  *      -EFAULT FTL rebuild error/timeout/interruption
2854  */
2855 static int mtip_ftl_rebuild_poll(struct driver_data *dd)
2856 {
2857         unsigned long timeout, cnt = 0, start;
2858
2859         dev_warn(&dd->pdev->dev,
2860                 "FTL rebuild in progress. Polling for completion.\n");
2861
2862         start = jiffies;
2863         timeout = jiffies + msecs_to_jiffies(MTIP_FTL_REBUILD_TIMEOUT_MS);
2864
2865         do {
2866                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
2867                                 &dd->dd_flag)))
2868                         return -EFAULT;
2869                 if (mtip_check_surprise_removal(dd->pdev))
2870                         return -EFAULT;
2871
2872                 if (mtip_get_identify(dd->port, NULL) < 0)
2873                         return -EFAULT;
2874
2875                 if (*(dd->port->identify + MTIP_FTL_REBUILD_OFFSET) ==
2876                         MTIP_FTL_REBUILD_MAGIC) {
2877                         ssleep(1);
2878                         /* Print message every 3 minutes */
2879                         if (cnt++ >= 180) {
2880                                 dev_warn(&dd->pdev->dev,
2881                                 "FTL rebuild in progress (%d secs).\n",
2882                                 jiffies_to_msecs(jiffies - start) / 1000);
2883                                 cnt = 0;
2884                         }
2885                 } else {
2886                         dev_warn(&dd->pdev->dev,
2887                                 "FTL rebuild complete (%d secs).\n",
2888                         jiffies_to_msecs(jiffies - start) / 1000);
2889                         mtip_block_initialize(dd);
2890                         return 0;
2891                 }
2892         } while (time_before(jiffies, timeout));
2893
2894         /* Check for timeout */
2895         dev_err(&dd->pdev->dev,
2896                 "Timed out waiting for FTL rebuild to complete (%d secs).\n",
2897                 jiffies_to_msecs(jiffies - start) / 1000);
2898         return -EFAULT;
2899 }
2900
2901 static void mtip_softirq_done_fn(struct request *rq)
2902 {
2903         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
2904         struct driver_data *dd = rq->q->queuedata;
2905
2906         /* Unmap the DMA scatter list entries */
2907         dma_unmap_sg(&dd->pdev->dev, cmd->sg, cmd->scatter_ents,
2908                                                         cmd->direction);
2909
2910         if (unlikely(cmd->unaligned))
2911                 up(&dd->port->cmd_slot_unal);
2912
2913         blk_mq_end_request(rq, rq->errors);
2914 }
2915
2916 static void mtip_abort_cmd(struct request *req, void *data,
2917                                                         bool reserved)
2918 {
2919         struct driver_data *dd = data;
2920
2921         dbg_printk(MTIP_DRV_NAME " Aborting request, tag = %d\n", req->tag);
2922
2923         clear_bit(req->tag, dd->port->cmds_to_issue);
2924         req->errors = -EIO;
2925         mtip_softirq_done_fn(req);
2926 }
2927
2928 static void mtip_queue_cmd(struct request *req, void *data,
2929                                                         bool reserved)
2930 {
2931         struct driver_data *dd = data;
2932
2933         set_bit(req->tag, dd->port->cmds_to_issue);
2934         blk_abort_request(req);
2935 }
2936
2937 /*
2938  * service thread to issue queued commands
2939  *
2940  * @data Pointer to the driver data structure.
2941  *
2942  * return value
2943  *      0
2944  */
2945
2946 static int mtip_service_thread(void *data)
2947 {
2948         struct driver_data *dd = (struct driver_data *)data;
2949         unsigned long slot, slot_start, slot_wrap, to;
2950         unsigned int num_cmd_slots = dd->slot_groups * 32;
2951         struct mtip_port *port = dd->port;
2952
2953         while (1) {
2954                 if (kthread_should_stop() ||
2955                         test_bit(MTIP_PF_SVC_THD_STOP_BIT, &port->flags))
2956                         goto st_out;
2957                 clear_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags);
2958
2959                 /*
2960                  * the condition is to check neither an internal command is
2961                  * is in progress nor error handling is active
2962                  */
2963                 wait_event_interruptible(port->svc_wait, (port->flags) &&
2964                         (port->flags & MTIP_PF_SVC_THD_WORK));
2965
2966                 if (kthread_should_stop() ||
2967                         test_bit(MTIP_PF_SVC_THD_STOP_BIT, &port->flags))
2968                         goto st_out;
2969
2970                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
2971                                 &dd->dd_flag)))
2972                         goto st_out;
2973
2974                 set_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags);
2975
2976 restart_eh:
2977                 /* Demux bits: start with error handling */
2978                 if (test_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags)) {
2979                         mtip_handle_tfe(dd);
2980                         clear_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags);
2981                 }
2982
2983                 if (test_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags))
2984                         goto restart_eh;
2985
2986                 if (test_bit(MTIP_PF_TO_ACTIVE_BIT, &port->flags)) {
2987                         to = jiffies + msecs_to_jiffies(5000);
2988
2989                         do {
2990                                 mdelay(100);
2991                         } while (atomic_read(&dd->irq_workers_active) != 0 &&
2992                                 time_before(jiffies, to));
2993
2994                         if (atomic_read(&dd->irq_workers_active) != 0)
2995                                 dev_warn(&dd->pdev->dev,
2996                                         "Completion workers still active!");
2997
2998                         spin_lock(dd->queue->queue_lock);
2999                         blk_mq_tagset_busy_iter(&dd->tags,
3000                                                         mtip_queue_cmd, dd);
3001                         spin_unlock(dd->queue->queue_lock);
3002
3003                         set_bit(MTIP_PF_ISSUE_CMDS_BIT, &dd->port->flags);
3004
3005                         if (mtip_device_reset(dd))
3006                                 blk_mq_tagset_busy_iter(&dd->tags,
3007                                                         mtip_abort_cmd, dd);
3008
3009                         clear_bit(MTIP_PF_TO_ACTIVE_BIT, &dd->port->flags);
3010                 }
3011
3012                 if (test_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags)) {
3013                         slot = 1;
3014                         /* used to restrict the loop to one iteration */
3015                         slot_start = num_cmd_slots;
3016                         slot_wrap = 0;
3017                         while (1) {
3018                                 slot = find_next_bit(port->cmds_to_issue,
3019                                                 num_cmd_slots, slot);
3020                                 if (slot_wrap == 1) {
3021                                         if ((slot_start >= slot) ||
3022                                                 (slot >= num_cmd_slots))
3023                                                 break;
3024                                 }
3025                                 if (unlikely(slot_start == num_cmd_slots))
3026                                         slot_start = slot;
3027
3028                                 if (unlikely(slot == num_cmd_slots)) {
3029                                         slot = 1;
3030                                         slot_wrap = 1;
3031                                         continue;
3032                                 }
3033
3034                                 /* Issue the command to the hardware */
3035                                 mtip_issue_ncq_command(port, slot);
3036
3037                                 clear_bit(slot, port->cmds_to_issue);
3038                         }
3039
3040                         clear_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags);
3041                 }
3042
3043                 if (test_bit(MTIP_PF_REBUILD_BIT, &port->flags)) {
3044                         if (mtip_ftl_rebuild_poll(dd) == 0)
3045                                 clear_bit(MTIP_PF_REBUILD_BIT, &port->flags);
3046                 }
3047         }
3048
3049 st_out:
3050         return 0;
3051 }
3052
3053 /*
3054  * DMA region teardown
3055  *
3056  * @dd Pointer to driver_data structure
3057  *
3058  * return value
3059  *      None
3060  */
3061 static void mtip_dma_free(struct driver_data *dd)
3062 {
3063         struct mtip_port *port = dd->port;
3064
3065         if (port->block1)
3066                 dmam_free_coherent(&dd->pdev->dev, BLOCK_DMA_ALLOC_SZ,
3067                                         port->block1, port->block1_dma);
3068
3069         if (port->command_list) {
3070                 dmam_free_coherent(&dd->pdev->dev, AHCI_CMD_TBL_SZ,
3071                                 port->command_list, port->command_list_dma);
3072         }
3073 }
3074
3075 /*
3076  * DMA region setup
3077  *
3078  * @dd Pointer to driver_data structure
3079  *
3080  * return value
3081  *      -ENOMEM Not enough free DMA region space to initialize driver
3082  */
3083 static int mtip_dma_alloc(struct driver_data *dd)
3084 {
3085         struct mtip_port *port = dd->port;
3086
3087         /* Allocate dma memory for RX Fis, Identify, and Sector Bufffer */
3088         port->block1 =
3089                 dmam_alloc_coherent(&dd->pdev->dev, BLOCK_DMA_ALLOC_SZ,
3090                                         &port->block1_dma, GFP_KERNEL);
3091         if (!port->block1)
3092                 return -ENOMEM;
3093         memset(port->block1, 0, BLOCK_DMA_ALLOC_SZ);
3094
3095         /* Allocate dma memory for command list */
3096         port->command_list =
3097                 dmam_alloc_coherent(&dd->pdev->dev, AHCI_CMD_TBL_SZ,
3098                                         &port->command_list_dma, GFP_KERNEL);
3099         if (!port->command_list) {
3100                 dmam_free_coherent(&dd->pdev->dev, BLOCK_DMA_ALLOC_SZ,
3101                                         port->block1, port->block1_dma);
3102                 port->block1 = NULL;
3103                 port->block1_dma = 0;
3104                 return -ENOMEM;
3105         }
3106         memset(port->command_list, 0, AHCI_CMD_TBL_SZ);
3107
3108         /* Setup all pointers into first DMA region */
3109         port->rxfis         = port->block1 + AHCI_RX_FIS_OFFSET;
3110         port->rxfis_dma     = port->block1_dma + AHCI_RX_FIS_OFFSET;
3111         port->identify      = port->block1 + AHCI_IDFY_OFFSET;
3112         port->identify_dma  = port->block1_dma + AHCI_IDFY_OFFSET;
3113         port->log_buf       = port->block1 + AHCI_SECTBUF_OFFSET;
3114         port->log_buf_dma   = port->block1_dma + AHCI_SECTBUF_OFFSET;
3115         port->smart_buf     = port->block1 + AHCI_SMARTBUF_OFFSET;
3116         port->smart_buf_dma = port->block1_dma + AHCI_SMARTBUF_OFFSET;
3117
3118         return 0;
3119 }
3120
3121 static int mtip_hw_get_identify(struct driver_data *dd)
3122 {
3123         struct smart_attr attr242;
3124         unsigned char *buf;
3125         int rv;
3126
3127         if (mtip_get_identify(dd->port, NULL) < 0)
3128                 return -EFAULT;
3129
3130         if (*(dd->port->identify + MTIP_FTL_REBUILD_OFFSET) ==
3131                 MTIP_FTL_REBUILD_MAGIC) {
3132                 set_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags);
3133                 return MTIP_FTL_REBUILD_MAGIC;
3134         }
3135         mtip_dump_identify(dd->port);
3136
3137         /* check write protect, over temp and rebuild statuses */
3138         rv = mtip_read_log_page(dd->port, ATA_LOG_SATA_NCQ,
3139                                 dd->port->log_buf,
3140                                 dd->port->log_buf_dma, 1);
3141         if (rv) {
3142                 dev_warn(&dd->pdev->dev,
3143                         "Error in READ LOG EXT (10h) command\n");
3144                 /* non-critical error, don't fail the load */
3145         } else {
3146                 buf = (unsigned char *)dd->port->log_buf;
3147                 if (buf[259] & 0x1) {
3148                         dev_info(&dd->pdev->dev,
3149                                 "Write protect bit is set.\n");
3150                         set_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag);
3151                 }
3152                 if (buf[288] == 0xF7) {
3153                         dev_info(&dd->pdev->dev,
3154                                 "Exceeded Tmax, drive in thermal shutdown.\n");
3155                         set_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag);
3156                 }
3157                 if (buf[288] == 0xBF) {
3158                         dev_info(&dd->pdev->dev,
3159                                 "Drive indicates rebuild has failed.\n");
3160                         set_bit(MTIP_DDF_REBUILD_FAILED_BIT, &dd->dd_flag);
3161                 }
3162         }
3163
3164         /* get write protect progess */
3165         memset(&attr242, 0, sizeof(struct smart_attr));
3166         if (mtip_get_smart_attr(dd->port, 242, &attr242))
3167                 dev_warn(&dd->pdev->dev,
3168                                 "Unable to check write protect progress\n");
3169         else
3170                 dev_info(&dd->pdev->dev,
3171                                 "Write protect progress: %u%% (%u blocks)\n",
3172                                 attr242.cur, le32_to_cpu(attr242.data));
3173
3174         return rv;
3175 }
3176
3177 /*
3178  * Called once for each card.
3179  *
3180  * @dd Pointer to the driver data structure.
3181  *
3182  * return value
3183  *      0 on success, else an error code.
3184  */
3185 static int mtip_hw_init(struct driver_data *dd)
3186 {
3187         int i;
3188         int rv;
3189         unsigned int num_command_slots;
3190         unsigned long timeout, timetaken;
3191
3192         dd->mmio = pcim_iomap_table(dd->pdev)[MTIP_ABAR];
3193
3194         mtip_detect_product(dd);
3195         if (dd->product_type == MTIP_PRODUCT_UNKNOWN) {
3196                 rv = -EIO;
3197                 goto out1;
3198         }
3199         num_command_slots = dd->slot_groups * 32;
3200
3201         hba_setup(dd);
3202
3203         dd->port = kzalloc_node(sizeof(struct mtip_port), GFP_KERNEL,
3204                                 dd->numa_node);
3205         if (!dd->port) {
3206                 dev_err(&dd->pdev->dev,
3207                         "Memory allocation: port structure\n");
3208                 return -ENOMEM;
3209         }
3210
3211         /* Continue workqueue setup */
3212         for (i = 0; i < MTIP_MAX_SLOT_GROUPS; i++)
3213                 dd->work[i].port = dd->port;
3214
3215         /* Enable unaligned IO constraints for some devices */
3216         if (mtip_device_unaligned_constrained(dd))
3217                 dd->unal_qdepth = MTIP_MAX_UNALIGNED_SLOTS;
3218         else
3219                 dd->unal_qdepth = 0;
3220
3221         sema_init(&dd->port->cmd_slot_unal, dd->unal_qdepth);
3222
3223         /* Spinlock to prevent concurrent issue */
3224         for (i = 0; i < MTIP_MAX_SLOT_GROUPS; i++)
3225                 spin_lock_init(&dd->port->cmd_issue_lock[i]);
3226
3227         /* Set the port mmio base address. */
3228         dd->port->mmio  = dd->mmio + PORT_OFFSET;
3229         dd->port->dd    = dd;
3230
3231         /* DMA allocations */
3232         rv = mtip_dma_alloc(dd);
3233         if (rv < 0)
3234                 goto out1;
3235
3236         /* Setup the pointers to the extended s_active and CI registers. */
3237         for (i = 0; i < dd->slot_groups; i++) {
3238                 dd->port->s_active[i] =
3239                         dd->port->mmio + i*0x80 + PORT_SCR_ACT;
3240                 dd->port->cmd_issue[i] =
3241                         dd->port->mmio + i*0x80 + PORT_COMMAND_ISSUE;
3242                 dd->port->completed[i] =
3243                         dd->port->mmio + i*0x80 + PORT_SDBV;
3244         }
3245
3246         timetaken = jiffies;
3247         timeout = jiffies + msecs_to_jiffies(30000);
3248         while (((readl(dd->port->mmio + PORT_SCR_STAT) & 0x0F) != 0x03) &&
3249                  time_before(jiffies, timeout)) {
3250                 mdelay(100);
3251         }
3252         if (unlikely(mtip_check_surprise_removal(dd->pdev))) {
3253                 timetaken = jiffies - timetaken;
3254                 dev_warn(&dd->pdev->dev,
3255                         "Surprise removal detected at %u ms\n",
3256                         jiffies_to_msecs(timetaken));
3257                 rv = -ENODEV;
3258                 goto out2 ;
3259         }
3260         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag))) {
3261                 timetaken = jiffies - timetaken;
3262                 dev_warn(&dd->pdev->dev,
3263                         "Removal detected at %u ms\n",
3264                         jiffies_to_msecs(timetaken));
3265                 rv = -EFAULT;
3266                 goto out2;
3267         }
3268
3269         /* Conditionally reset the HBA. */
3270         if (!(readl(dd->mmio + HOST_CAP) & HOST_CAP_NZDMA)) {
3271                 if (mtip_hba_reset(dd) < 0) {
3272                         dev_err(&dd->pdev->dev,
3273                                 "Card did not reset within timeout\n");
3274                         rv = -EIO;
3275                         goto out2;
3276                 }
3277         } else {
3278                 /* Clear any pending interrupts on the HBA */
3279                 writel(readl(dd->mmio + HOST_IRQ_STAT),
3280                         dd->mmio + HOST_IRQ_STAT);
3281         }
3282
3283         mtip_init_port(dd->port);
3284         mtip_start_port(dd->port);
3285
3286         /* Setup the ISR and enable interrupts. */
3287         rv = devm_request_irq(&dd->pdev->dev,
3288                                 dd->pdev->irq,
3289                                 mtip_irq_handler,
3290                                 IRQF_SHARED,
3291                                 dev_driver_string(&dd->pdev->dev),
3292                                 dd);
3293
3294         if (rv) {
3295                 dev_err(&dd->pdev->dev,
3296                         "Unable to allocate IRQ %d\n", dd->pdev->irq);
3297                 goto out2;
3298         }
3299         irq_set_affinity_hint(dd->pdev->irq, get_cpu_mask(dd->isr_binding));
3300
3301         /* Enable interrupts on the HBA. */
3302         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
3303                                         dd->mmio + HOST_CTL);
3304
3305         init_waitqueue_head(&dd->port->svc_wait);
3306
3307         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)) {
3308                 rv = -EFAULT;
3309                 goto out3;
3310         }
3311
3312         return rv;
3313
3314 out3:
3315         /* Disable interrupts on the HBA. */
3316         writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3317                         dd->mmio + HOST_CTL);
3318
3319         /* Release the IRQ. */
3320         irq_set_affinity_hint(dd->pdev->irq, NULL);
3321         devm_free_irq(&dd->pdev->dev, dd->pdev->irq, dd);
3322
3323 out2:
3324         mtip_deinit_port(dd->port);
3325         mtip_dma_free(dd);
3326
3327 out1:
3328         /* Free the memory allocated for the for structure. */
3329         kfree(dd->port);
3330
3331         return rv;
3332 }
3333
3334 static int mtip_standby_drive(struct driver_data *dd)
3335 {
3336         int rv = 0;
3337
3338         if (dd->sr || !dd->port)
3339                 return -ENODEV;
3340         /*
3341          * Send standby immediate (E0h) to the drive so that it
3342          * saves its state.
3343          */
3344         if (!test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags) &&
3345             !test_bit(MTIP_DDF_REBUILD_FAILED_BIT, &dd->dd_flag) &&
3346             !test_bit(MTIP_DDF_SEC_LOCK_BIT, &dd->dd_flag)) {
3347                 rv = mtip_standby_immediate(dd->port);
3348                 if (rv)
3349                         dev_warn(&dd->pdev->dev,
3350                                 "STANDBY IMMEDIATE failed\n");
3351         }
3352         return rv;
3353 }
3354
3355 /*
3356  * Called to deinitialize an interface.
3357  *
3358  * @dd Pointer to the driver data structure.
3359  *
3360  * return value
3361  *      0
3362  */
3363 static int mtip_hw_exit(struct driver_data *dd)
3364 {
3365         if (!dd->sr) {
3366                 /* de-initialize the port. */
3367                 mtip_deinit_port(dd->port);
3368
3369                 /* Disable interrupts on the HBA. */
3370                 writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3371                                 dd->mmio + HOST_CTL);
3372         }
3373
3374         /* Release the IRQ. */
3375         irq_set_affinity_hint(dd->pdev->irq, NULL);
3376         devm_free_irq(&dd->pdev->dev, dd->pdev->irq, dd);
3377         msleep(1000);
3378
3379         /* Free dma regions */
3380         mtip_dma_free(dd);
3381
3382         /* Free the memory allocated for the for structure. */
3383         kfree(dd->port);
3384         dd->port = NULL;
3385
3386         return 0;
3387 }
3388
3389 /*
3390  * Issue a Standby Immediate command to the device.
3391  *
3392  * This function is called by the Block Layer just before the
3393  * system powers off during a shutdown.
3394  *
3395  * @dd Pointer to the driver data structure.
3396  *
3397  * return value
3398  *      0
3399  */
3400 static int mtip_hw_shutdown(struct driver_data *dd)
3401 {
3402         /*
3403          * Send standby immediate (E0h) to the drive so that it
3404          * saves its state.
3405          */
3406         mtip_standby_drive(dd);
3407
3408         return 0;
3409 }
3410
3411 /*
3412  * Suspend function
3413  *
3414  * This function is called by the Block Layer just before the
3415  * system hibernates.
3416  *
3417  * @dd Pointer to the driver data structure.
3418  *
3419  * return value
3420  *      0       Suspend was successful
3421  *      -EFAULT Suspend was not successful
3422  */
3423 static int mtip_hw_suspend(struct driver_data *dd)
3424 {
3425         /*
3426          * Send standby immediate (E0h) to the drive
3427          * so that it saves its state.
3428          */
3429         if (mtip_standby_drive(dd) != 0) {
3430                 dev_err(&dd->pdev->dev,
3431                         "Failed standby-immediate command\n");
3432                 return -EFAULT;
3433         }
3434
3435         /* Disable interrupts on the HBA.*/
3436         writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3437                         dd->mmio + HOST_CTL);
3438         mtip_deinit_port(dd->port);
3439
3440         return 0;
3441 }
3442
3443 /*
3444  * Resume function
3445  *
3446  * This function is called by the Block Layer as the
3447  * system resumes.
3448  *
3449  * @dd Pointer to the driver data structure.
3450  *
3451  * return value
3452  *      0       Resume was successful
3453  *      -EFAULT Resume was not successful
3454  */
3455 static int mtip_hw_resume(struct driver_data *dd)
3456 {
3457         /* Perform any needed hardware setup steps */
3458         hba_setup(dd);
3459
3460         /* Reset the HBA */
3461         if (mtip_hba_reset(dd) != 0) {
3462                 dev_err(&dd->pdev->dev,
3463                         "Unable to reset the HBA\n");
3464                 return -EFAULT;
3465         }
3466
3467         /*
3468          * Enable the port, DMA engine, and FIS reception specific
3469          * h/w in controller.
3470          */
3471         mtip_init_port(dd->port);
3472         mtip_start_port(dd->port);
3473
3474         /* Enable interrupts on the HBA.*/
3475         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
3476                         dd->mmio + HOST_CTL);
3477
3478         return 0;
3479 }
3480
3481 /*
3482  * Helper function for reusing disk name
3483  * upon hot insertion.
3484  */
3485 static int rssd_disk_name_format(char *prefix,
3486                                  int index,
3487                                  char *buf,
3488                                  int buflen)
3489 {
3490         const int base = 'z' - 'a' + 1;
3491         char *begin = buf + strlen(prefix);
3492         char *end = buf + buflen;
3493         char *p;
3494         int unit;
3495
3496         p = end - 1;
3497         *p = '\0';
3498         unit = base;
3499         do {
3500                 if (p == begin)
3501                         return -EINVAL;
3502                 *--p = 'a' + (index % unit);
3503                 index = (index / unit) - 1;
3504         } while (index >= 0);
3505
3506         memmove(begin, p, end - p);
3507         memcpy(buf, prefix, strlen(prefix));
3508
3509         return 0;
3510 }
3511
3512 /*
3513  * Block layer IOCTL handler.
3514  *
3515  * @dev Pointer to the block_device structure.
3516  * @mode ignored
3517  * @cmd IOCTL command passed from the user application.
3518  * @arg Argument passed from the user application.
3519  *
3520  * return value
3521  *      0        IOCTL completed successfully.
3522  *      -ENOTTY  IOCTL not supported or invalid driver data
3523  *                 structure pointer.
3524  */
3525 static int mtip_block_ioctl(struct block_device *dev,
3526                             fmode_t mode,
3527                             unsigned cmd,
3528                             unsigned long arg)
3529 {
3530         struct driver_data *dd = dev->bd_disk->private_data;
3531
3532         if (!capable(CAP_SYS_ADMIN))
3533                 return -EACCES;
3534
3535         if (!dd)
3536                 return -ENOTTY;
3537
3538         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)))
3539                 return -ENOTTY;
3540
3541         switch (cmd) {
3542         case BLKFLSBUF:
3543                 return -ENOTTY;
3544         default:
3545                 return mtip_hw_ioctl(dd, cmd, arg);
3546         }
3547 }
3548
3549 #ifdef CONFIG_COMPAT
3550 /*
3551  * Block layer compat IOCTL handler.
3552  *
3553  * @dev Pointer to the block_device structure.
3554  * @mode ignored
3555  * @cmd IOCTL command passed from the user application.
3556  * @arg Argument passed from the user application.
3557  *
3558  * return value
3559  *      0        IOCTL completed successfully.
3560  *      -ENOTTY  IOCTL not supported or invalid driver data
3561  *                 structure pointer.
3562  */
3563 static int mtip_block_compat_ioctl(struct block_device *dev,
3564                             fmode_t mode,
3565                             unsigned cmd,
3566                             unsigned long arg)
3567 {
3568         struct driver_data *dd = dev->bd_disk->private_data;
3569
3570         if (!capable(CAP_SYS_ADMIN))
3571                 return -EACCES;
3572
3573         if (!dd)
3574                 return -ENOTTY;
3575
3576         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)))
3577                 return -ENOTTY;
3578
3579         switch (cmd) {
3580         case BLKFLSBUF:
3581                 return -ENOTTY;
3582         case HDIO_DRIVE_TASKFILE: {
3583                 struct mtip_compat_ide_task_request_s __user *compat_req_task;
3584                 ide_task_request_t req_task;
3585                 int compat_tasksize, outtotal, ret;
3586
3587                 compat_tasksize =
3588                         sizeof(struct mtip_compat_ide_task_request_s);
3589
3590                 compat_req_task =
3591                         (struct mtip_compat_ide_task_request_s __user *) arg;
3592
3593                 if (copy_from_user(&req_task, (void __user *) arg,
3594                         compat_tasksize - (2 * sizeof(compat_long_t))))
3595                         return -EFAULT;
3596
3597                 if (get_user(req_task.out_size, &compat_req_task->out_size))
3598                         return -EFAULT;
3599
3600                 if (get_user(req_task.in_size, &compat_req_task->in_size))
3601                         return -EFAULT;
3602
3603                 outtotal = sizeof(struct mtip_compat_ide_task_request_s);
3604
3605                 ret = exec_drive_taskfile(dd, (void __user *) arg,
3606                                                 &req_task, outtotal);
3607
3608                 if (copy_to_user((void __user *) arg, &req_task,
3609                                 compat_tasksize -
3610                                 (2 * sizeof(compat_long_t))))
3611                         return -EFAULT;
3612
3613                 if (put_user(req_task.out_size, &compat_req_task->out_size))
3614                         return -EFAULT;
3615
3616                 if (put_user(req_task.in_size, &compat_req_task->in_size))
3617                         return -EFAULT;
3618
3619                 return ret;
3620         }
3621         default:
3622                 return mtip_hw_ioctl(dd, cmd, arg);
3623         }
3624 }
3625 #endif
3626
3627 /*
3628  * Obtain the geometry of the device.
3629  *
3630  * You may think that this function is obsolete, but some applications,
3631  * fdisk for example still used CHS values. This function describes the
3632  * device as having 224 heads and 56 sectors per cylinder. These values are
3633  * chosen so that each cylinder is aligned on a 4KB boundary. Since a
3634  * partition is described in terms of a start and end cylinder this means
3635  * that each partition is also 4KB aligned. Non-aligned partitions adversely
3636  * affects performance.
3637  *
3638  * @dev Pointer to the block_device strucutre.
3639  * @geo Pointer to a hd_geometry structure.
3640  *
3641  * return value
3642  *      0       Operation completed successfully.
3643  *      -ENOTTY An error occurred while reading the drive capacity.
3644  */
3645 static int mtip_block_getgeo(struct block_device *dev,
3646                                 struct hd_geometry *geo)
3647 {
3648         struct driver_data *dd = dev->bd_disk->private_data;
3649         sector_t capacity;
3650
3651         if (!dd)
3652                 return -ENOTTY;
3653
3654         if (!(mtip_hw_get_capacity(dd, &capacity))) {
3655                 dev_warn(&dd->pdev->dev,
3656                         "Could not get drive capacity.\n");
3657                 return -ENOTTY;
3658         }
3659
3660         geo->heads = 224;
3661         geo->sectors = 56;
3662         sector_div(capacity, (geo->heads * geo->sectors));
3663         geo->cylinders = capacity;
3664         return 0;
3665 }
3666
3667 static int mtip_block_open(struct block_device *dev, fmode_t mode)
3668 {
3669         struct driver_data *dd;
3670
3671         if (dev && dev->bd_disk) {
3672                 dd = (struct driver_data *) dev->bd_disk->private_data;
3673
3674                 if (dd) {
3675                         if (test_bit(MTIP_DDF_REMOVAL_BIT,
3676                                                         &dd->dd_flag)) {
3677                                 return -ENODEV;
3678                         }
3679                         return 0;
3680                 }
3681         }
3682         return -ENODEV;
3683 }
3684
3685 static void mtip_block_release(struct gendisk *disk, fmode_t mode)
3686 {
3687 }
3688
3689 /*
3690  * Block device operation function.
3691  *
3692  * This structure contains pointers to the functions required by the block
3693  * layer.
3694  */
3695 static const struct block_device_operations mtip_block_ops = {
3696         .open           = mtip_block_open,
3697         .release        = mtip_block_release,
3698         .ioctl          = mtip_block_ioctl,
3699 #ifdef CONFIG_COMPAT
3700         .compat_ioctl   = mtip_block_compat_ioctl,
3701 #endif
3702         .getgeo         = mtip_block_getgeo,
3703         .owner          = THIS_MODULE
3704 };
3705
3706 static inline bool is_se_active(struct driver_data *dd)
3707 {
3708         if (unlikely(test_bit(MTIP_PF_SE_ACTIVE_BIT, &dd->port->flags))) {
3709                 if (dd->port->ic_pause_timer) {
3710                         unsigned long to = dd->port->ic_pause_timer +
3711                                                         msecs_to_jiffies(1000);
3712                         if (time_after(jiffies, to)) {
3713                                 clear_bit(MTIP_PF_SE_ACTIVE_BIT,
3714                                                         &dd->port->flags);
3715                                 clear_bit(MTIP_DDF_SEC_LOCK_BIT, &dd->dd_flag);
3716                                 dd->port->ic_pause_timer = 0;
3717                                 wake_up_interruptible(&dd->port->svc_wait);
3718                                 return false;
3719                         }
3720                 }
3721                 return true;
3722         }
3723         return false;
3724 }
3725
3726 /*
3727  * Block layer make request function.
3728  *
3729  * This function is called by the kernel to process a BIO for
3730  * the P320 device.
3731  *
3732  * @queue Pointer to the request queue. Unused other than to obtain
3733  *              the driver data structure.
3734  * @rq    Pointer to the request.
3735  *
3736  */
3737 static int mtip_submit_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
3738 {
3739         struct driver_data *dd = hctx->queue->queuedata;
3740         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
3741         unsigned int nents;
3742
3743         if (is_se_active(dd))
3744                 return -ENODATA;
3745
3746         if (unlikely(dd->dd_flag & MTIP_DDF_STOP_IO)) {
3747                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
3748                                                         &dd->dd_flag))) {
3749                         return -ENXIO;
3750                 }
3751                 if (unlikely(test_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag))) {
3752                         return -ENODATA;
3753                 }
3754                 if (unlikely(test_bit(MTIP_DDF_WRITE_PROTECT_BIT,
3755                                                         &dd->dd_flag) &&
3756                                 rq_data_dir(rq))) {
3757                         return -ENODATA;
3758                 }
3759                 if (unlikely(test_bit(MTIP_DDF_SEC_LOCK_BIT, &dd->dd_flag) ||
3760                         test_bit(MTIP_DDF_REBUILD_FAILED_BIT, &dd->dd_flag)))
3761                         return -ENODATA;
3762         }
3763
3764         if (req_op(rq) == REQ_OP_DISCARD) {
3765                 int err;
3766
3767                 err = mtip_send_trim(dd, blk_rq_pos(rq), blk_rq_sectors(rq));
3768                 blk_mq_end_request(rq, err);
3769                 return 0;
3770         }
3771
3772         /* Create the scatter list for this request. */
3773         nents = blk_rq_map_sg(hctx->queue, rq, cmd->sg);
3774
3775         /* Issue the read/write. */
3776         mtip_hw_submit_io(dd, rq, cmd, nents, hctx);
3777         return 0;
3778 }
3779
3780 static bool mtip_check_unal_depth(struct blk_mq_hw_ctx *hctx,
3781                                   struct request *rq)
3782 {
3783         struct driver_data *dd = hctx->queue->queuedata;
3784         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
3785
3786         if (rq_data_dir(rq) == READ || !dd->unal_qdepth)
3787                 return false;
3788
3789         /*
3790          * If unaligned depth must be limited on this controller, mark it
3791          * as unaligned if the IO isn't on a 4k boundary (start of length).
3792          */
3793         if (blk_rq_sectors(rq) <= 64) {
3794                 if ((blk_rq_pos(rq) & 7) || (blk_rq_sectors(rq) & 7))
3795                         cmd->unaligned = 1;
3796         }
3797
3798         if (cmd->unaligned && down_trylock(&dd->port->cmd_slot_unal))
3799                 return true;
3800
3801         return false;
3802 }
3803
3804 static int mtip_queue_rq(struct blk_mq_hw_ctx *hctx,
3805                          const struct blk_mq_queue_data *bd)
3806 {
3807         struct request *rq = bd->rq;
3808         int ret;
3809
3810         if (unlikely(mtip_check_unal_depth(hctx, rq)))
3811                 return BLK_MQ_RQ_QUEUE_BUSY;
3812
3813         blk_mq_start_request(rq);
3814
3815         ret = mtip_submit_request(hctx, rq);
3816         if (likely(!ret))
3817                 return BLK_MQ_RQ_QUEUE_OK;
3818
3819         rq->errors = ret;
3820         return BLK_MQ_RQ_QUEUE_ERROR;
3821 }
3822
3823 static void mtip_free_cmd(void *data, struct request *rq,
3824                           unsigned int hctx_idx, unsigned int request_idx)
3825 {
3826         struct driver_data *dd = data;
3827         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
3828
3829         if (!cmd->command)
3830                 return;
3831
3832         dmam_free_coherent(&dd->pdev->dev, CMD_DMA_ALLOC_SZ,
3833                                 cmd->command, cmd->command_dma);
3834 }
3835
3836 static int mtip_init_cmd(void *data, struct request *rq, unsigned int hctx_idx,
3837                          unsigned int request_idx, unsigned int numa_node)
3838 {
3839         struct driver_data *dd = data;
3840         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
3841         u32 host_cap_64 = readl(dd->mmio + HOST_CAP) & HOST_CAP_64;
3842
3843         /*
3844          * For flush requests, request_idx starts at the end of the
3845          * tag space.  Since we don't support FLUSH/FUA, simply return
3846          * 0 as there's nothing to be done.
3847          */
3848         if (request_idx >= MTIP_MAX_COMMAND_SLOTS)
3849                 return 0;
3850
3851         cmd->command = dmam_alloc_coherent(&dd->pdev->dev, CMD_DMA_ALLOC_SZ,
3852                         &cmd->command_dma, GFP_KERNEL);
3853         if (!cmd->command)
3854                 return -ENOMEM;
3855
3856         memset(cmd->command, 0, CMD_DMA_ALLOC_SZ);
3857
3858         /* Point the command headers at the command tables. */
3859         cmd->command_header = dd->port->command_list +
3860                                 (sizeof(struct mtip_cmd_hdr) * request_idx);
3861         cmd->command_header_dma = dd->port->command_list_dma +
3862                                 (sizeof(struct mtip_cmd_hdr) * request_idx);
3863
3864         if (host_cap_64)
3865                 cmd->command_header->ctbau = __force_bit2int cpu_to_le32((cmd->command_dma >> 16) >> 16);
3866
3867         cmd->command_header->ctba = __force_bit2int cpu_to_le32(cmd->command_dma & 0xFFFFFFFF);
3868
3869         sg_init_table(cmd->sg, MTIP_MAX_SG);
3870         return 0;
3871 }
3872
3873 static enum blk_eh_timer_return mtip_cmd_timeout(struct request *req,
3874                                                                 bool reserved)
3875 {
3876         struct driver_data *dd = req->q->queuedata;
3877
3878         if (reserved)
3879                 goto exit_handler;
3880
3881         if (test_bit(req->tag, dd->port->cmds_to_issue))
3882                 goto exit_handler;
3883
3884         if (test_and_set_bit(MTIP_PF_TO_ACTIVE_BIT, &dd->port->flags))
3885                 goto exit_handler;
3886
3887         wake_up_interruptible(&dd->port->svc_wait);
3888 exit_handler:
3889         return BLK_EH_RESET_TIMER;
3890 }
3891
3892 static struct blk_mq_ops mtip_mq_ops = {
3893         .queue_rq       = mtip_queue_rq,
3894         .init_request   = mtip_init_cmd,
3895         .exit_request   = mtip_free_cmd,
3896         .complete       = mtip_softirq_done_fn,
3897         .timeout        = mtip_cmd_timeout,
3898 };
3899
3900 /*
3901  * Block layer initialization function.
3902  *
3903  * This function is called once by the PCI layer for each P320
3904  * device that is connected to the system.
3905  *
3906  * @dd Pointer to the driver data structure.
3907  *
3908  * return value
3909  *      0 on success else an error code.
3910  */
3911 static int mtip_block_initialize(struct driver_data *dd)
3912 {
3913         int rv = 0, wait_for_rebuild = 0;
3914         sector_t capacity;
3915         unsigned int index = 0;
3916         struct kobject *kobj;
3917
3918         if (dd->disk)
3919                 goto skip_create_disk; /* hw init done, before rebuild */
3920
3921         if (mtip_hw_init(dd)) {
3922                 rv = -EINVAL;
3923                 goto protocol_init_error;
3924         }
3925
3926         dd->disk = alloc_disk_node(MTIP_MAX_MINORS, dd->numa_node);
3927         if (dd->disk  == NULL) {
3928                 dev_err(&dd->pdev->dev,
3929                         "Unable to allocate gendisk structure\n");
3930                 rv = -EINVAL;
3931                 goto alloc_disk_error;
3932         }
3933
3934         /* Generate the disk name, implemented same as in sd.c */
3935         do {
3936                 if (!ida_pre_get(&rssd_index_ida, GFP_KERNEL)) {
3937                         rv = -ENOMEM;
3938                         goto ida_get_error;
3939                 }
3940
3941                 spin_lock(&rssd_index_lock);
3942                 rv = ida_get_new(&rssd_index_ida, &index);
3943                 spin_unlock(&rssd_index_lock);
3944         } while (rv == -EAGAIN);
3945
3946         if (rv)
3947                 goto ida_get_error;
3948
3949         rv = rssd_disk_name_format("rssd",
3950                                 index,
3951                                 dd->disk->disk_name,
3952                                 DISK_NAME_LEN);
3953         if (rv)
3954                 goto disk_index_error;
3955
3956         dd->disk->major         = dd->major;
3957         dd->disk->first_minor   = index * MTIP_MAX_MINORS;
3958         dd->disk->minors        = MTIP_MAX_MINORS;
3959         dd->disk->fops          = &mtip_block_ops;
3960         dd->disk->private_data  = dd;
3961         dd->index               = index;
3962
3963         mtip_hw_debugfs_init(dd);
3964
3965         memset(&dd->tags, 0, sizeof(dd->tags));
3966         dd->tags.ops = &mtip_mq_ops;
3967         dd->tags.nr_hw_queues = 1;
3968         dd->tags.queue_depth = MTIP_MAX_COMMAND_SLOTS;
3969         dd->tags.reserved_tags = 1;
3970         dd->tags.cmd_size = sizeof(struct mtip_cmd);
3971         dd->tags.numa_node = dd->numa_node;
3972         dd->tags.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_NO_SCHED;
3973         dd->tags.driver_data = dd;
3974         dd->tags.timeout = MTIP_NCQ_CMD_TIMEOUT_MS;
3975
3976         rv = blk_mq_alloc_tag_set(&dd->tags);
3977         if (rv) {
3978                 dev_err(&dd->pdev->dev,
3979                         "Unable to allocate request queue\n");
3980                 goto block_queue_alloc_tag_error;
3981         }
3982
3983         /* Allocate the request queue. */
3984         dd->queue = blk_mq_init_queue(&dd->tags);
3985         if (IS_ERR(dd->queue)) {
3986                 dev_err(&dd->pdev->dev,
3987                         "Unable to allocate request queue\n");
3988                 rv = -ENOMEM;
3989                 goto block_queue_alloc_init_error;
3990         }
3991
3992         dd->disk->queue         = dd->queue;
3993         dd->queue->queuedata    = dd;
3994
3995 skip_create_disk:
3996         /* Initialize the protocol layer. */
3997         wait_for_rebuild = mtip_hw_get_identify(dd);
3998         if (wait_for_rebuild < 0) {
3999                 dev_err(&dd->pdev->dev,
4000                         "Protocol layer initialization failed\n");
4001                 rv = -EINVAL;
4002                 goto init_hw_cmds_error;
4003         }
4004
4005         /*
4006          * if rebuild pending, start the service thread, and delay the block
4007          * queue creation and device_add_disk()
4008          */
4009         if (wait_for_rebuild == MTIP_FTL_REBUILD_MAGIC)
4010                 goto start_service_thread;
4011
4012         /* Set device limits. */
4013         set_bit(QUEUE_FLAG_NONROT, &dd->queue->queue_flags);
4014         clear_bit(QUEUE_FLAG_ADD_RANDOM, &dd->queue->queue_flags);
4015         blk_queue_max_segments(dd->queue, MTIP_MAX_SG);
4016         blk_queue_physical_block_size(dd->queue, 4096);
4017         blk_queue_max_hw_sectors(dd->queue, 0xffff);
4018         blk_queue_max_segment_size(dd->queue, 0x400000);
4019         blk_queue_io_min(dd->queue, 4096);
4020         blk_queue_bounce_limit(dd->queue, dd->pdev->dma_mask);
4021
4022         /* Signal trim support */
4023         if (dd->trim_supp == true) {
4024                 set_bit(QUEUE_FLAG_DISCARD, &dd->queue->queue_flags);
4025                 dd->queue->limits.discard_granularity = 4096;
4026                 blk_queue_max_discard_sectors(dd->queue,
4027                         MTIP_MAX_TRIM_ENTRY_LEN * MTIP_MAX_TRIM_ENTRIES);
4028                 dd->queue->limits.discard_zeroes_data = 0;
4029         }
4030
4031         /* Set the capacity of the device in 512 byte sectors. */
4032         if (!(mtip_hw_get_capacity(dd, &capacity))) {
4033                 dev_warn(&dd->pdev->dev,
4034                         "Could not read drive capacity\n");
4035                 rv = -EIO;
4036                 goto read_capacity_error;
4037         }
4038         set_capacity(dd->disk, capacity);
4039
4040         /* Enable the block device and add it to /dev */
4041         device_add_disk(&dd->pdev->dev, dd->disk);
4042
4043         dd->bdev = bdget_disk(dd->disk, 0);
4044         /*
4045          * Now that the disk is active, initialize any sysfs attributes
4046          * managed by the protocol layer.
4047          */
4048         kobj = kobject_get(&disk_to_dev(dd->disk)->kobj);
4049         if (kobj) {
4050                 mtip_hw_sysfs_init(dd, kobj);
4051                 kobject_put(kobj);
4052         }
4053
4054         if (dd->mtip_svc_handler) {
4055                 set_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag);
4056                 return rv; /* service thread created for handling rebuild */
4057         }
4058
4059 start_service_thread:
4060         dd->mtip_svc_handler = kthread_create_on_node(mtip_service_thread,
4061                                                 dd, dd->numa_node,
4062                                                 "mtip_svc_thd_%02d", index);
4063
4064         if (IS_ERR(dd->mtip_svc_handler)) {
4065                 dev_err(&dd->pdev->dev, "service thread failed to start\n");
4066                 dd->mtip_svc_handler = NULL;
4067                 rv = -EFAULT;
4068                 goto kthread_run_error;
4069         }
4070         wake_up_process(dd->mtip_svc_handler);
4071         if (wait_for_rebuild == MTIP_FTL_REBUILD_MAGIC)
4072                 rv = wait_for_rebuild;
4073
4074         return rv;
4075
4076 kthread_run_error:
4077         bdput(dd->bdev);
4078         dd->bdev = NULL;
4079
4080         /* Delete our gendisk. This also removes the device from /dev */
4081         del_gendisk(dd->disk);
4082
4083 read_capacity_error:
4084 init_hw_cmds_error:
4085         blk_cleanup_queue(dd->queue);
4086 block_queue_alloc_init_error:
4087         blk_mq_free_tag_set(&dd->tags);
4088 block_queue_alloc_tag_error:
4089         mtip_hw_debugfs_exit(dd);
4090 disk_index_error:
4091         spin_lock(&rssd_index_lock);
4092         ida_remove(&rssd_index_ida, index);
4093         spin_unlock(&rssd_index_lock);
4094
4095 ida_get_error:
4096         put_disk(dd->disk);
4097
4098 alloc_disk_error:
4099         mtip_hw_exit(dd); /* De-initialize the protocol layer. */
4100
4101 protocol_init_error:
4102         return rv;
4103 }
4104
4105 static void mtip_no_dev_cleanup(struct request *rq, void *data, bool reserv)
4106 {
4107         struct driver_data *dd = (struct driver_data *)data;
4108         struct mtip_cmd *cmd;
4109
4110         if (likely(!reserv))
4111                 blk_mq_complete_request(rq, -ENODEV);
4112         else if (test_bit(MTIP_PF_IC_ACTIVE_BIT, &dd->port->flags)) {
4113
4114                 cmd = mtip_cmd_from_tag(dd, MTIP_TAG_INTERNAL);
4115                 if (cmd->comp_func)
4116                         cmd->comp_func(dd->port, MTIP_TAG_INTERNAL,
4117                                         cmd, -ENODEV);
4118         }
4119 }
4120
4121 /*
4122  * Block layer deinitialization function.
4123  *
4124  * Called by the PCI layer as each P320 device is removed.
4125  *
4126  * @dd Pointer to the driver data structure.
4127  *
4128  * return value
4129  *      0
4130  */
4131 static int mtip_block_remove(struct driver_data *dd)
4132 {
4133         struct kobject *kobj;
4134
4135         mtip_hw_debugfs_exit(dd);
4136
4137         if (dd->mtip_svc_handler) {
4138                 set_bit(MTIP_PF_SVC_THD_STOP_BIT, &dd->port->flags);
4139                 wake_up_interruptible(&dd->port->svc_wait);
4140                 kthread_stop(dd->mtip_svc_handler);
4141         }
4142
4143         /* Clean up the sysfs attributes, if created */
4144         if (test_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag)) {
4145                 kobj = kobject_get(&disk_to_dev(dd->disk)->kobj);
4146                 if (kobj) {
4147                         mtip_hw_sysfs_exit(dd, kobj);
4148                         kobject_put(kobj);
4149                 }
4150         }
4151
4152         if (!dd->sr) {
4153                 /*
4154                  * Explicitly wait here for IOs to quiesce,
4155                  * as mtip_standby_drive usually won't wait for IOs.
4156                  */
4157                 if (!mtip_quiesce_io(dd->port, MTIP_QUIESCE_IO_TIMEOUT_MS,
4158                                                                 GFP_KERNEL))
4159                         mtip_standby_drive(dd);
4160         }
4161         else
4162                 dev_info(&dd->pdev->dev, "device %s surprise removal\n",
4163                                                 dd->disk->disk_name);
4164
4165         blk_mq_freeze_queue_start(dd->queue);
4166         blk_mq_stop_hw_queues(dd->queue);
4167         blk_mq_tagset_busy_iter(&dd->tags, mtip_no_dev_cleanup, dd);
4168
4169         /*
4170          * Delete our gendisk structure. This also removes the device
4171          * from /dev
4172          */
4173         if (dd->bdev) {
4174                 bdput(dd->bdev);
4175                 dd->bdev = NULL;
4176         }
4177         if (dd->disk) {
4178                 if (test_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag))
4179                         del_gendisk(dd->disk);
4180                 if (dd->disk->queue) {
4181                         blk_cleanup_queue(dd->queue);
4182                         blk_mq_free_tag_set(&dd->tags);
4183                         dd->queue = NULL;
4184                 }
4185                 put_disk(dd->disk);
4186         }
4187         dd->disk  = NULL;
4188
4189         spin_lock(&rssd_index_lock);
4190         ida_remove(&rssd_index_ida, dd->index);
4191         spin_unlock(&rssd_index_lock);
4192
4193         /* De-initialize the protocol layer. */
4194         mtip_hw_exit(dd);
4195
4196         return 0;
4197 }
4198
4199 /*
4200  * Function called by the PCI layer when just before the
4201  * machine shuts down.
4202  *
4203  * If a protocol layer shutdown function is present it will be called
4204  * by this function.
4205  *
4206  * @dd Pointer to the driver data structure.
4207  *
4208  * return value
4209  *      0
4210  */
4211 static int mtip_block_shutdown(struct driver_data *dd)
4212 {
4213         mtip_hw_shutdown(dd);
4214
4215         /* Delete our gendisk structure, and cleanup the blk queue. */
4216         if (dd->disk) {
4217                 dev_info(&dd->pdev->dev,
4218                         "Shutting down %s ...\n", dd->disk->disk_name);
4219
4220                 if (test_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag))
4221                         del_gendisk(dd->disk);
4222                 if (dd->disk->queue) {
4223                         blk_cleanup_queue(dd->queue);
4224                         blk_mq_free_tag_set(&dd->tags);
4225                 }
4226                 put_disk(dd->disk);
4227                 dd->disk  = NULL;
4228                 dd->queue = NULL;
4229         }
4230
4231         spin_lock(&rssd_index_lock);
4232         ida_remove(&rssd_index_ida, dd->index);
4233         spin_unlock(&rssd_index_lock);
4234         return 0;
4235 }
4236
4237 static int mtip_block_suspend(struct driver_data *dd)
4238 {
4239         dev_info(&dd->pdev->dev,
4240                 "Suspending %s ...\n", dd->disk->disk_name);
4241         mtip_hw_suspend(dd);
4242         return 0;
4243 }
4244
4245 static int mtip_block_resume(struct driver_data *dd)
4246 {
4247         dev_info(&dd->pdev->dev, "Resuming %s ...\n",
4248                 dd->disk->disk_name);
4249         mtip_hw_resume(dd);
4250         return 0;
4251 }
4252
4253 static void drop_cpu(int cpu)
4254 {
4255         cpu_use[cpu]--;
4256 }
4257
4258 static int get_least_used_cpu_on_node(int node)
4259 {
4260         int cpu, least_used_cpu, least_cnt;
4261         const struct cpumask *node_mask;
4262
4263         node_mask = cpumask_of_node(node);
4264         least_used_cpu = cpumask_first(node_mask);
4265         least_cnt = cpu_use[least_used_cpu];
4266         cpu = least_used_cpu;
4267
4268         for_each_cpu(cpu, node_mask) {
4269                 if (cpu_use[cpu] < least_cnt) {
4270                         least_used_cpu = cpu;
4271                         least_cnt = cpu_use[cpu];
4272                 }
4273         }
4274         cpu_use[least_used_cpu]++;
4275         return least_used_cpu;
4276 }
4277
4278 /* Helper for selecting a node in round robin mode */
4279 static inline int mtip_get_next_rr_node(void)
4280 {
4281         static int next_node = -1;
4282
4283         if (next_node == -1) {
4284                 next_node = first_online_node;
4285                 return next_node;
4286         }
4287
4288         next_node = next_online_node(next_node);
4289         if (next_node == MAX_NUMNODES)
4290                 next_node = first_online_node;
4291         return next_node;
4292 }
4293
4294 static DEFINE_HANDLER(0);
4295 static DEFINE_HANDLER(1);
4296 static DEFINE_HANDLER(2);
4297 static DEFINE_HANDLER(3);
4298 static DEFINE_HANDLER(4);
4299 static DEFINE_HANDLER(5);
4300 static DEFINE_HANDLER(6);
4301 static DEFINE_HANDLER(7);
4302
4303 static void mtip_disable_link_opts(struct driver_data *dd, struct pci_dev *pdev)
4304 {
4305         int pos;
4306         unsigned short pcie_dev_ctrl;
4307
4308         pos = pci_find_capability(pdev, PCI_CAP_ID_EXP);
4309         if (pos) {
4310                 pci_read_config_word(pdev,
4311                         pos + PCI_EXP_DEVCTL,
4312                         &pcie_dev_ctrl);
4313                 if (pcie_dev_ctrl & (1 << 11) ||
4314                     pcie_dev_ctrl & (1 << 4)) {
4315                         dev_info(&dd->pdev->dev,
4316                                 "Disabling ERO/No-Snoop on bridge device %04x:%04x\n",
4317                                         pdev->vendor, pdev->device);
4318                         pcie_dev_ctrl &= ~(PCI_EXP_DEVCTL_NOSNOOP_EN |
4319                                                 PCI_EXP_DEVCTL_RELAX_EN);
4320                         pci_write_config_word(pdev,
4321                                 pos + PCI_EXP_DEVCTL,
4322                                 pcie_dev_ctrl);
4323                 }
4324         }
4325 }
4326
4327 static void mtip_fix_ero_nosnoop(struct driver_data *dd, struct pci_dev *pdev)
4328 {
4329         /*
4330          * This workaround is specific to AMD/ATI chipset with a PCI upstream
4331          * device with device id 0x5aXX
4332          */
4333         if (pdev->bus && pdev->bus->self) {
4334                 if (pdev->bus->self->vendor == PCI_VENDOR_ID_ATI &&
4335                     ((pdev->bus->self->device & 0xff00) == 0x5a00)) {
4336                         mtip_disable_link_opts(dd, pdev->bus->self);
4337                 } else {
4338                         /* Check further up the topology */
4339                         struct pci_dev *parent_dev = pdev->bus->self;
4340                         if (parent_dev->bus &&
4341                                 parent_dev->bus->parent &&
4342                                 parent_dev->bus->parent->self &&
4343                                 parent_dev->bus->parent->self->vendor ==
4344                                          PCI_VENDOR_ID_ATI &&
4345                                 (parent_dev->bus->parent->self->device &
4346                                         0xff00) == 0x5a00) {
4347                                 mtip_disable_link_opts(dd,
4348                                         parent_dev->bus->parent->self);
4349                         }
4350                 }
4351         }
4352 }
4353
4354 /*
4355  * Called for each supported PCI device detected.
4356  *
4357  * This function allocates the private data structure, enables the
4358  * PCI device and then calls the block layer initialization function.
4359  *
4360  * return value
4361  *      0 on success else an error code.
4362  */
4363 static int mtip_pci_probe(struct pci_dev *pdev,
4364                         const struct pci_device_id *ent)
4365 {
4366         int rv = 0;
4367         struct driver_data *dd = NULL;
4368         char cpu_list[256];
4369         const struct cpumask *node_mask;
4370         int cpu, i = 0, j = 0;
4371         int my_node = NUMA_NO_NODE;
4372         unsigned long flags;
4373
4374         /* Allocate memory for this devices private data. */
4375         my_node = pcibus_to_node(pdev->bus);
4376         if (my_node != NUMA_NO_NODE) {
4377                 if (!node_online(my_node))
4378                         my_node = mtip_get_next_rr_node();
4379         } else {
4380                 dev_info(&pdev->dev, "Kernel not reporting proximity, choosing a node\n");
4381                 my_node = mtip_get_next_rr_node();
4382         }
4383         dev_info(&pdev->dev, "NUMA node %d (closest: %d,%d, probe on %d:%d)\n",
4384                 my_node, pcibus_to_node(pdev->bus), dev_to_node(&pdev->dev),
4385                 cpu_to_node(raw_smp_processor_id()), raw_smp_processor_id());
4386
4387         dd = kzalloc_node(sizeof(struct driver_data), GFP_KERNEL, my_node);
4388         if (dd == NULL) {
4389                 dev_err(&pdev->dev,
4390                         "Unable to allocate memory for driver data\n");
4391                 return -ENOMEM;
4392         }
4393
4394         /* Attach the private data to this PCI device.  */
4395         pci_set_drvdata(pdev, dd);
4396
4397         rv = pcim_enable_device(pdev);
4398         if (rv < 0) {
4399                 dev_err(&pdev->dev, "Unable to enable device\n");
4400                 goto iomap_err;
4401         }
4402
4403         /* Map BAR5 to memory. */
4404         rv = pcim_iomap_regions(pdev, 1 << MTIP_ABAR, MTIP_DRV_NAME);
4405         if (rv < 0) {
4406                 dev_err(&pdev->dev, "Unable to map regions\n");
4407                 goto iomap_err;
4408         }
4409
4410         if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(64))) {
4411                 rv = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
4412
4413                 if (rv) {
4414                         rv = pci_set_consistent_dma_mask(pdev,
4415                                                 DMA_BIT_MASK(32));
4416                         if (rv) {
4417                                 dev_warn(&pdev->dev,
4418                                         "64-bit DMA enable failed\n");
4419                                 goto setmask_err;
4420                         }
4421                 }
4422         }
4423
4424         /* Copy the info we may need later into the private data structure. */
4425         dd->major       = mtip_major;
4426         dd->instance    = instance;
4427         dd->pdev        = pdev;
4428         dd->numa_node   = my_node;
4429
4430         INIT_LIST_HEAD(&dd->online_list);
4431         INIT_LIST_HEAD(&dd->remove_list);
4432
4433         memset(dd->workq_name, 0, 32);
4434         snprintf(dd->workq_name, 31, "mtipq%d", dd->instance);
4435
4436         dd->isr_workq = create_workqueue(dd->workq_name);
4437         if (!dd->isr_workq) {
4438                 dev_warn(&pdev->dev, "Can't create wq %d\n", dd->instance);
4439                 rv = -ENOMEM;
4440                 goto block_initialize_err;
4441         }
4442
4443         memset(cpu_list, 0, sizeof(cpu_list));
4444
4445         node_mask = cpumask_of_node(dd->numa_node);
4446         if (!cpumask_empty(node_mask)) {
4447                 for_each_cpu(cpu, node_mask)
4448                 {
4449                         snprintf(&cpu_list[j], 256 - j, "%d ", cpu);
4450                         j = strlen(cpu_list);
4451                 }
4452
4453                 dev_info(&pdev->dev, "Node %d on package %d has %d cpu(s): %s\n",
4454                         dd->numa_node,
4455                         topology_physical_package_id(cpumask_first(node_mask)),
4456                         nr_cpus_node(dd->numa_node),
4457                         cpu_list);
4458         } else
4459                 dev_dbg(&pdev->dev, "mtip32xx: node_mask empty\n");
4460
4461         dd->isr_binding = get_least_used_cpu_on_node(dd->numa_node);
4462         dev_info(&pdev->dev, "Initial IRQ binding node:cpu %d:%d\n",
4463                 cpu_to_node(dd->isr_binding), dd->isr_binding);
4464
4465         /* first worker context always runs in ISR */
4466         dd->work[0].cpu_binding = dd->isr_binding;
4467         dd->work[1].cpu_binding = get_least_used_cpu_on_node(dd->numa_node);
4468         dd->work[2].cpu_binding = get_least_used_cpu_on_node(dd->numa_node);
4469         dd->work[3].cpu_binding = dd->work[0].cpu_binding;
4470         dd->work[4].cpu_binding = dd->work[1].cpu_binding;
4471         dd->work[5].cpu_binding = dd->work[2].cpu_binding;
4472         dd->work[6].cpu_binding = dd->work[2].cpu_binding;
4473         dd->work[7].cpu_binding = dd->work[1].cpu_binding;
4474
4475         /* Log the bindings */
4476         for_each_present_cpu(cpu) {
4477                 memset(cpu_list, 0, sizeof(cpu_list));
4478                 for (i = 0, j = 0; i < MTIP_MAX_SLOT_GROUPS; i++) {
4479                         if (dd->work[i].cpu_binding == cpu) {
4480                                 snprintf(&cpu_list[j], 256 - j, "%d ", i);
4481                                 j = strlen(cpu_list);
4482                         }
4483                 }
4484                 if (j)
4485                         dev_info(&pdev->dev, "CPU %d: WQs %s\n", cpu, cpu_list);
4486         }
4487
4488         INIT_WORK(&dd->work[0].work, mtip_workq_sdbf0);
4489         INIT_WORK(&dd->work[1].work, mtip_workq_sdbf1);
4490         INIT_WORK(&dd->work[2].work, mtip_workq_sdbf2);
4491         INIT_WORK(&dd->work[3].work, mtip_workq_sdbf3);
4492         INIT_WORK(&dd->work[4].work, mtip_workq_sdbf4);
4493         INIT_WORK(&dd->work[5].work, mtip_workq_sdbf5);
4494         INIT_WORK(&dd->work[6].work, mtip_workq_sdbf6);
4495         INIT_WORK(&dd->work[7].work, mtip_workq_sdbf7);
4496
4497         pci_set_master(pdev);
4498         rv = pci_enable_msi(pdev);
4499         if (rv) {
4500                 dev_warn(&pdev->dev,
4501                         "Unable to enable MSI interrupt.\n");
4502                 goto msi_initialize_err;
4503         }
4504
4505         mtip_fix_ero_nosnoop(dd, pdev);
4506
4507         /* Initialize the block layer. */
4508         rv = mtip_block_initialize(dd);
4509         if (rv < 0) {
4510                 dev_err(&pdev->dev,
4511                         "Unable to initialize block layer\n");
4512                 goto block_initialize_err;
4513         }
4514
4515         /*
4516          * Increment the instance count so that each device has a unique
4517          * instance number.
4518          */
4519         instance++;
4520         if (rv != MTIP_FTL_REBUILD_MAGIC)
4521                 set_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag);
4522         else
4523                 rv = 0; /* device in rebuild state, return 0 from probe */
4524
4525         /* Add to online list even if in ftl rebuild */
4526         spin_lock_irqsave(&dev_lock, flags);
4527         list_add(&dd->online_list, &online_list);
4528         spin_unlock_irqrestore(&dev_lock, flags);
4529
4530         goto done;
4531
4532 block_initialize_err:
4533         pci_disable_msi(pdev);
4534
4535 msi_initialize_err:
4536         if (dd->isr_workq) {
4537                 flush_workqueue(dd->isr_workq);
4538                 destroy_workqueue(dd->isr_workq);
4539                 drop_cpu(dd->work[0].cpu_binding);
4540                 drop_cpu(dd->work[1].cpu_binding);
4541                 drop_cpu(dd->work[2].cpu_binding);
4542         }
4543 setmask_err:
4544         pcim_iounmap_regions(pdev, 1 << MTIP_ABAR);
4545
4546 iomap_err:
4547         kfree(dd);
4548         pci_set_drvdata(pdev, NULL);
4549         return rv;
4550 done:
4551         return rv;
4552 }
4553
4554 /*
4555  * Called for each probed device when the device is removed or the
4556  * driver is unloaded.
4557  *
4558  * return value
4559  *      None
4560  */
4561 static void mtip_pci_remove(struct pci_dev *pdev)
4562 {
4563         struct driver_data *dd = pci_get_drvdata(pdev);
4564         unsigned long flags, to;
4565
4566         set_bit(MTIP_DDF_REMOVAL_BIT, &dd->dd_flag);
4567
4568         spin_lock_irqsave(&dev_lock, flags);
4569         list_del_init(&dd->online_list);
4570         list_add(&dd->remove_list, &removing_list);
4571         spin_unlock_irqrestore(&dev_lock, flags);
4572
4573         mtip_check_surprise_removal(pdev);
4574         synchronize_irq(dd->pdev->irq);
4575
4576         /* Spin until workers are done */
4577         to = jiffies + msecs_to_jiffies(4000);
4578         do {
4579                 msleep(20);
4580         } while (atomic_read(&dd->irq_workers_active) != 0 &&
4581                 time_before(jiffies, to));
4582
4583         if (!dd->sr)
4584                 fsync_bdev(dd->bdev);
4585
4586         if (atomic_read(&dd->irq_workers_active) != 0) {
4587                 dev_warn(&dd->pdev->dev,
4588                         "Completion workers still active!\n");
4589         }
4590
4591         blk_set_queue_dying(dd->queue);
4592         set_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag);
4593
4594         /* Clean up the block layer. */
4595         mtip_block_remove(dd);
4596
4597         if (dd->isr_workq) {
4598                 flush_workqueue(dd->isr_workq);
4599                 destroy_workqueue(dd->isr_workq);
4600                 drop_cpu(dd->work[0].cpu_binding);
4601                 drop_cpu(dd->work[1].cpu_binding);
4602                 drop_cpu(dd->work[2].cpu_binding);
4603         }
4604
4605         pci_disable_msi(pdev);
4606
4607         spin_lock_irqsave(&dev_lock, flags);
4608         list_del_init(&dd->remove_list);
4609         spin_unlock_irqrestore(&dev_lock, flags);
4610
4611         kfree(dd);
4612
4613         pcim_iounmap_regions(pdev, 1 << MTIP_ABAR);
4614         pci_set_drvdata(pdev, NULL);
4615 }
4616
4617 /*
4618  * Called for each probed device when the device is suspended.
4619  *
4620  * return value
4621  *      0  Success
4622  *      <0 Error
4623  */
4624 static int mtip_pci_suspend(struct pci_dev *pdev, pm_message_t mesg)
4625 {
4626         int rv = 0;
4627         struct driver_data *dd = pci_get_drvdata(pdev);
4628
4629         if (!dd) {
4630                 dev_err(&pdev->dev,
4631                         "Driver private datastructure is NULL\n");
4632                 return -EFAULT;
4633         }
4634
4635         set_bit(MTIP_DDF_RESUME_BIT, &dd->dd_flag);
4636
4637         /* Disable ports & interrupts then send standby immediate */
4638         rv = mtip_block_suspend(dd);
4639         if (rv < 0) {
4640                 dev_err(&pdev->dev,
4641                         "Failed to suspend controller\n");
4642                 return rv;
4643         }
4644
4645         /*
4646          * Save the pci config space to pdev structure &
4647          * disable the device
4648          */
4649         pci_save_state(pdev);
4650         pci_disable_device(pdev);
4651
4652         /* Move to Low power state*/
4653         pci_set_power_state(pdev, PCI_D3hot);
4654
4655         return rv;
4656 }
4657
4658 /*
4659  * Called for each probed device when the device is resumed.
4660  *
4661  * return value
4662  *      0  Success
4663  *      <0 Error
4664  */
4665 static int mtip_pci_resume(struct pci_dev *pdev)
4666 {
4667         int rv = 0;
4668         struct driver_data *dd;
4669
4670         dd = pci_get_drvdata(pdev);
4671         if (!dd) {
4672                 dev_err(&pdev->dev,
4673                         "Driver private datastructure is NULL\n");
4674                 return -EFAULT;
4675         }
4676
4677         /* Move the device to active State */
4678         pci_set_power_state(pdev, PCI_D0);
4679
4680         /* Restore PCI configuration space */
4681         pci_restore_state(pdev);
4682
4683         /* Enable the PCI device*/
4684         rv = pcim_enable_device(pdev);
4685         if (rv < 0) {
4686                 dev_err(&pdev->dev,
4687                         "Failed to enable card during resume\n");
4688                 goto err;
4689         }
4690         pci_set_master(pdev);
4691
4692         /*
4693          * Calls hbaReset, initPort, & startPort function
4694          * then enables interrupts
4695          */
4696         rv = mtip_block_resume(dd);
4697         if (rv < 0)
4698                 dev_err(&pdev->dev, "Unable to resume\n");
4699
4700 err:
4701         clear_bit(MTIP_DDF_RESUME_BIT, &dd->dd_flag);
4702
4703         return rv;
4704 }
4705
4706 /*
4707  * Shutdown routine
4708  *
4709  * return value
4710  *      None
4711  */
4712 static void mtip_pci_shutdown(struct pci_dev *pdev)
4713 {
4714         struct driver_data *dd = pci_get_drvdata(pdev);
4715         if (dd)
4716                 mtip_block_shutdown(dd);
4717 }
4718
4719 /* Table of device ids supported by this driver. */
4720 static const struct pci_device_id mtip_pci_tbl[] = {
4721         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320H_DEVICE_ID) },
4722         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320M_DEVICE_ID) },
4723         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320S_DEVICE_ID) },
4724         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P325M_DEVICE_ID) },
4725         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P420H_DEVICE_ID) },
4726         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P420M_DEVICE_ID) },
4727         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P425M_DEVICE_ID) },
4728         { 0 }
4729 };
4730
4731 /* Structure that describes the PCI driver functions. */
4732 static struct pci_driver mtip_pci_driver = {
4733         .name                   = MTIP_DRV_NAME,
4734         .id_table               = mtip_pci_tbl,
4735         .probe                  = mtip_pci_probe,
4736         .remove                 = mtip_pci_remove,
4737         .suspend                = mtip_pci_suspend,
4738         .resume                 = mtip_pci_resume,
4739         .shutdown               = mtip_pci_shutdown,
4740 };
4741
4742 MODULE_DEVICE_TABLE(pci, mtip_pci_tbl);
4743
4744 /*
4745  * Module initialization function.
4746  *
4747  * Called once when the module is loaded. This function allocates a major
4748  * block device number to the Cyclone devices and registers the PCI layer
4749  * of the driver.
4750  *
4751  * Return value
4752  *      0 on success else error code.
4753  */
4754 static int __init mtip_init(void)
4755 {
4756         int error;
4757
4758         pr_info(MTIP_DRV_NAME " Version " MTIP_DRV_VERSION "\n");
4759
4760         spin_lock_init(&dev_lock);
4761
4762         INIT_LIST_HEAD(&online_list);
4763         INIT_LIST_HEAD(&removing_list);
4764
4765         /* Allocate a major block device number to use with this driver. */
4766         error = register_blkdev(0, MTIP_DRV_NAME);
4767         if (error <= 0) {
4768                 pr_err("Unable to register block device (%d)\n",
4769                 error);
4770                 return -EBUSY;
4771         }
4772         mtip_major = error;
4773
4774         dfs_parent = debugfs_create_dir("rssd", NULL);
4775         if (IS_ERR_OR_NULL(dfs_parent)) {
4776                 pr_warn("Error creating debugfs parent\n");
4777                 dfs_parent = NULL;
4778         }
4779         if (dfs_parent) {
4780                 dfs_device_status = debugfs_create_file("device_status",
4781                                         S_IRUGO, dfs_parent, NULL,
4782                                         &mtip_device_status_fops);
4783                 if (IS_ERR_OR_NULL(dfs_device_status)) {
4784                         pr_err("Error creating device_status node\n");
4785                         dfs_device_status = NULL;
4786                 }
4787         }
4788
4789         /* Register our PCI operations. */
4790         error = pci_register_driver(&mtip_pci_driver);
4791         if (error) {
4792                 debugfs_remove(dfs_parent);
4793                 unregister_blkdev(mtip_major, MTIP_DRV_NAME);
4794         }
4795
4796         return error;
4797 }
4798
4799 /*
4800  * Module de-initialization function.
4801  *
4802  * Called once when the module is unloaded. This function deallocates
4803  * the major block device number allocated by mtip_init() and
4804  * unregisters the PCI layer of the driver.
4805  *
4806  * Return value
4807  *      none
4808  */
4809 static void __exit mtip_exit(void)
4810 {
4811         /* Release the allocated major block device number. */
4812         unregister_blkdev(mtip_major, MTIP_DRV_NAME);
4813
4814         /* Unregister the PCI driver. */
4815         pci_unregister_driver(&mtip_pci_driver);
4816
4817         debugfs_remove_recursive(dfs_parent);
4818 }
4819
4820 MODULE_AUTHOR("Micron Technology, Inc");
4821 MODULE_DESCRIPTION("Micron RealSSD PCIe Block Driver");
4822 MODULE_LICENSE("GPL");
4823 MODULE_VERSION(MTIP_DRV_VERSION);
4824
4825 module_init(mtip_init);
4826 module_exit(mtip_exit);