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