drivers/infiniband/hw/hfi1/sdma.c

   1 /*
   2  * Copyright(c) 2015, 2016 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47
  48 #include <linux/spinlock.h>
  49 #include <linux/seqlock.h>
  50 #include <linux/netdevice.h>
  51 #include <linux/moduleparam.h>
  52 #include <linux/bitops.h>
  53 #include <linux/timer.h>
  54 #include <linux/vmalloc.h>
  55 #include <linux/highmem.h>
  56
  57 #include "hfi.h"
  58 #include "common.h"
  59 #include "qp.h"
  60 #include "sdma.h"
  61 #include "iowait.h"
  62 #include "trace.h"
  63
  64 /* must be a power of 2 >= 64 <= 32768 */
  65 #define SDMA_DESCQ_CNT 2048
  66 #define SDMA_DESC_INTR 64
  67 #define INVALID_TAIL 0xffff
  68
  69 static uint sdma_descq_cnt = SDMA_DESCQ_CNT;
  70 module_param(sdma_descq_cnt, uint, S_IRUGO);
  71 MODULE_PARM_DESC(sdma_descq_cnt, "Number of SDMA descq entries");
  72
  73 static uint sdma_idle_cnt = 250;
  74 module_param(sdma_idle_cnt, uint, S_IRUGO);
  75 MODULE_PARM_DESC(sdma_idle_cnt, "sdma interrupt idle delay (ns,default 250)");
  76
  77 uint mod_num_sdma;
  78 module_param_named(num_sdma, mod_num_sdma, uint, S_IRUGO);
  79 MODULE_PARM_DESC(num_sdma, "Set max number SDMA engines to use");
  80
  81 static uint sdma_desct_intr = SDMA_DESC_INTR;
  82 module_param_named(desct_intr, sdma_desct_intr, uint, S_IRUGO | S_IWUSR);
  83 MODULE_PARM_DESC(desct_intr, "Number of SDMA descriptor before interrupt");
  84
  85 #define SDMA_WAIT_BATCH_SIZE 20
  86 /* max wait time for a SDMA engine to indicate it has halted */
  87 #define SDMA_ERR_HALT_TIMEOUT 10 /* ms */
  88 /* all SDMA engine errors that cause a halt */
  89
  90 #define SD(name) SEND_DMA_##name
  91 #define ALL_SDMA_ENG_HALT_ERRS \
  92         (SD(ENG_ERR_STATUS_SDMA_WRONG_DW_ERR_SMASK) \
  93         | SD(ENG_ERR_STATUS_SDMA_GEN_MISMATCH_ERR_SMASK) \
  94         | SD(ENG_ERR_STATUS_SDMA_TOO_LONG_ERR_SMASK) \
  95         | SD(ENG_ERR_STATUS_SDMA_TAIL_OUT_OF_BOUNDS_ERR_SMASK) \
  96         | SD(ENG_ERR_STATUS_SDMA_FIRST_DESC_ERR_SMASK) \
  97         | SD(ENG_ERR_STATUS_SDMA_MEM_READ_ERR_SMASK) \
  98         | SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK) \
  99         | SD(ENG_ERR_STATUS_SDMA_LENGTH_MISMATCH_ERR_SMASK) \
 100         | SD(ENG_ERR_STATUS_SDMA_PACKET_DESC_OVERFLOW_ERR_SMASK) \
 101         | SD(ENG_ERR_STATUS_SDMA_HEADER_SELECT_ERR_SMASK) \
 102         | SD(ENG_ERR_STATUS_SDMA_HEADER_ADDRESS_ERR_SMASK) \
 103         | SD(ENG_ERR_STATUS_SDMA_HEADER_LENGTH_ERR_SMASK) \
 104         | SD(ENG_ERR_STATUS_SDMA_TIMEOUT_ERR_SMASK) \
 105         | SD(ENG_ERR_STATUS_SDMA_DESC_TABLE_UNC_ERR_SMASK) \
 106         | SD(ENG_ERR_STATUS_SDMA_ASSEMBLY_UNC_ERR_SMASK) \
 107         | SD(ENG_ERR_STATUS_SDMA_PACKET_TRACKING_UNC_ERR_SMASK) \
 108         | SD(ENG_ERR_STATUS_SDMA_HEADER_STORAGE_UNC_ERR_SMASK) \
 109         | SD(ENG_ERR_STATUS_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SMASK))
 110
 111 /* sdma_sendctrl operations */
 112 #define SDMA_SENDCTRL_OP_ENABLE    BIT(0)
 113 #define SDMA_SENDCTRL_OP_INTENABLE BIT(1)
 114 #define SDMA_SENDCTRL_OP_HALT      BIT(2)
 115 #define SDMA_SENDCTRL_OP_CLEANUP   BIT(3)
 116
 117 /* handle long defines */
 118 #define SDMA_EGRESS_PACKET_OCCUPANCY_SMASK \
 119 SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SMASK
 120 #define SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT \
 121 SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT
 122
 123 static const char * const sdma_state_names[] = {
 124         [sdma_state_s00_hw_down]                = "s00_HwDown",
 125         [sdma_state_s10_hw_start_up_halt_wait]  = "s10_HwStartUpHaltWait",
 126         [sdma_state_s15_hw_start_up_clean_wait] = "s15_HwStartUpCleanWait",
 127         [sdma_state_s20_idle]                   = "s20_Idle",
 128         [sdma_state_s30_sw_clean_up_wait]       = "s30_SwCleanUpWait",
 129         [sdma_state_s40_hw_clean_up_wait]       = "s40_HwCleanUpWait",
 130         [sdma_state_s50_hw_halt_wait]           = "s50_HwHaltWait",
 131         [sdma_state_s60_idle_halt_wait]         = "s60_IdleHaltWait",
 132         [sdma_state_s80_hw_freeze]              = "s80_HwFreeze",
 133         [sdma_state_s82_freeze_sw_clean]        = "s82_FreezeSwClean",
 134         [sdma_state_s99_running]                = "s99_Running",
 135 };
 136
 137 #ifdef CONFIG_SDMA_VERBOSITY
 138 static const char * const sdma_event_names[] = {
 139         [sdma_event_e00_go_hw_down]   = "e00_GoHwDown",
 140         [sdma_event_e10_go_hw_start]  = "e10_GoHwStart",
 141         [sdma_event_e15_hw_halt_done] = "e15_HwHaltDone",
 142         [sdma_event_e25_hw_clean_up_done] = "e25_HwCleanUpDone",
 143         [sdma_event_e30_go_running]   = "e30_GoRunning",
 144         [sdma_event_e40_sw_cleaned]   = "e40_SwCleaned",
 145         [sdma_event_e50_hw_cleaned]   = "e50_HwCleaned",
 146         [sdma_event_e60_hw_halted]    = "e60_HwHalted",
 147         [sdma_event_e70_go_idle]      = "e70_GoIdle",
 148         [sdma_event_e80_hw_freeze]    = "e80_HwFreeze",
 149         [sdma_event_e81_hw_frozen]    = "e81_HwFrozen",
 150         [sdma_event_e82_hw_unfreeze]  = "e82_HwUnfreeze",
 151         [sdma_event_e85_link_down]    = "e85_LinkDown",
 152         [sdma_event_e90_sw_halted]    = "e90_SwHalted",
 153 };
 154 #endif
 155
 156 static const struct sdma_set_state_action sdma_action_table[] = {
 157         [sdma_state_s00_hw_down] = {
 158                 .go_s99_running_tofalse = 1,
 159                 .op_enable = 0,
 160                 .op_intenable = 0,
 161                 .op_halt = 0,
 162                 .op_cleanup = 0,
 163         },
 164         [sdma_state_s10_hw_start_up_halt_wait] = {
 165                 .op_enable = 0,
 166                 .op_intenable = 0,
 167                 .op_halt = 1,
 168                 .op_cleanup = 0,
 169         },
 170         [sdma_state_s15_hw_start_up_clean_wait] = {
 171                 .op_enable = 0,
 172                 .op_intenable = 1,
 173                 .op_halt = 0,
 174                 .op_cleanup = 1,
 175         },
 176         [sdma_state_s20_idle] = {
 177                 .op_enable = 0,
 178                 .op_intenable = 1,
 179                 .op_halt = 0,
 180                 .op_cleanup = 0,
 181         },
 182         [sdma_state_s30_sw_clean_up_wait] = {
 183                 .op_enable = 0,
 184                 .op_intenable = 0,
 185                 .op_halt = 0,
 186                 .op_cleanup = 0,
 187         },
 188         [sdma_state_s40_hw_clean_up_wait] = {
 189                 .op_enable = 0,
 190                 .op_intenable = 0,
 191                 .op_halt = 0,
 192                 .op_cleanup = 1,
 193         },
 194         [sdma_state_s50_hw_halt_wait] = {
 195                 .op_enable = 0,
 196                 .op_intenable = 0,
 197                 .op_halt = 0,
 198                 .op_cleanup = 0,
 199         },
 200         [sdma_state_s60_idle_halt_wait] = {
 201                 .go_s99_running_tofalse = 1,
 202                 .op_enable = 0,
 203                 .op_intenable = 0,
 204                 .op_halt = 1,
 205                 .op_cleanup = 0,
 206         },
 207         [sdma_state_s80_hw_freeze] = {
 208                 .op_enable = 0,
 209                 .op_intenable = 0,
 210                 .op_halt = 0,
 211                 .op_cleanup = 0,
 212         },
 213         [sdma_state_s82_freeze_sw_clean] = {
 214                 .op_enable = 0,
 215                 .op_intenable = 0,
 216                 .op_halt = 0,
 217                 .op_cleanup = 0,
 218         },
 219         [sdma_state_s99_running] = {
 220                 .op_enable = 1,
 221                 .op_intenable = 1,
 222                 .op_halt = 0,
 223                 .op_cleanup = 0,
 224                 .go_s99_running_totrue = 1,
 225         },
 226 };
 227
 228 #define SDMA_TAIL_UPDATE_THRESH 0x1F
 229
 230 /* declare all statics here rather than keep sorting */
 231 static void sdma_complete(struct kref *);
 232 static void sdma_finalput(struct sdma_state *);
 233 static void sdma_get(struct sdma_state *);
 234 static void sdma_hw_clean_up_task(unsigned long);
 235 static void sdma_put(struct sdma_state *);
 236 static void sdma_set_state(struct sdma_engine *, enum sdma_states);
 237 static void sdma_start_hw_clean_up(struct sdma_engine *);
 238 static void sdma_sw_clean_up_task(unsigned long);
 239 static void sdma_sendctrl(struct sdma_engine *, unsigned);
 240 static void init_sdma_regs(struct sdma_engine *, u32, uint);
 241 static void sdma_process_event(
 242         struct sdma_engine *sde,
 243         enum sdma_events event);
 244 static void __sdma_process_event(
 245         struct sdma_engine *sde,
 246         enum sdma_events event);
 247 static void dump_sdma_state(struct sdma_engine *sde);
 248 static void sdma_make_progress(struct sdma_engine *sde, u64 status);
 249 static void sdma_desc_avail(struct sdma_engine *sde, uint avail);
 250 static void sdma_flush_descq(struct sdma_engine *sde);
 251
 252 /**
 253  * sdma_state_name() - return state string from enum
 254  * @state: state
 255  */
 256 static const char *sdma_state_name(enum sdma_states state)
 257 {
 258         return sdma_state_names[state];
 259 }
 260
 261 static void sdma_get(struct sdma_state *ss)
 262 {
 263         kref_get(&ss->kref);
 264 }
 265
 266 static void sdma_complete(struct kref *kref)
 267 {
 268         struct sdma_state *ss =
 269                 container_of(kref, struct sdma_state, kref);
 270
 271         complete(&ss->comp);
 272 }
 273
 274 static void sdma_put(struct sdma_state *ss)
 275 {
 276         kref_put(&ss->kref, sdma_complete);
 277 }
 278
 279 static void sdma_finalput(struct sdma_state *ss)
 280 {
 281         sdma_put(ss);
 282         wait_for_completion(&ss->comp);
 283 }
 284
 285 static inline void write_sde_csr(
 286         struct sdma_engine *sde,
 287         u32 offset0,
 288         u64 value)
 289 {
 290         write_kctxt_csr(sde->dd, sde->this_idx, offset0, value);
 291 }
 292
 293 static inline u64 read_sde_csr(
 294         struct sdma_engine *sde,
 295         u32 offset0)
 296 {
 297         return read_kctxt_csr(sde->dd, sde->this_idx, offset0);
 298 }
 299
 300 /*
 301  * sdma_wait_for_packet_egress() - wait for the VL FIFO occupancy for
 302  * sdma engine 'sde' to drop to 0.
 303  */
 304 static void sdma_wait_for_packet_egress(struct sdma_engine *sde,
 305                                         int pause)
 306 {
 307         u64 off = 8 * sde->this_idx;
 308         struct hfi1_devdata *dd = sde->dd;
 309         int lcnt = 0;
 310         u64 reg_prev;
 311         u64 reg = 0;
 312
 313         while (1) {
 314                 reg_prev = reg;
 315                 reg = read_csr(dd, off + SEND_EGRESS_SEND_DMA_STATUS);
 316
 317                 reg &= SDMA_EGRESS_PACKET_OCCUPANCY_SMASK;
 318                 reg >>= SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT;
 319                 if (reg == 0)
 320                         break;
 321                 /* counter is reest if accupancy count changes */
 322                 if (reg != reg_prev)
 323                         lcnt = 0;
 324                 if (lcnt++ > 500) {
 325                         /* timed out - bounce the link */
 326                         dd_dev_err(dd, "%s: engine %u timeout waiting for packets to egress, remaining count %u, bouncing link\n",
 327                                    __func__, sde->this_idx, (u32)reg);
 328                         queue_work(dd->pport->link_wq,
 329                                    &dd->pport->link_bounce_work);
 330                         break;
 331                 }
 332                 udelay(1);
 333         }
 334 }
 335
 336 /*
 337  * sdma_wait() - wait for packet egress to complete for all SDMA engines,
 338  * and pause for credit return.
 339  */
 340 void sdma_wait(struct hfi1_devdata *dd)
 341 {
 342         int i;
 343
 344         for (i = 0; i < dd->num_sdma; i++) {
 345                 struct sdma_engine *sde = &dd->per_sdma[i];
 346
 347                 sdma_wait_for_packet_egress(sde, 0);
 348         }
 349 }
 350
 351 static inline void sdma_set_desc_cnt(struct sdma_engine *sde, unsigned cnt)
 352 {
 353         u64 reg;
 354
 355         if (!(sde->dd->flags & HFI1_HAS_SDMA_TIMEOUT))
 356                 return;
 357         reg = cnt;
 358         reg &= SD(DESC_CNT_CNT_MASK);
 359         reg <<= SD(DESC_CNT_CNT_SHIFT);
 360         write_sde_csr(sde, SD(DESC_CNT), reg);
 361 }
 362
 363 static inline void complete_tx(struct sdma_engine *sde,
 364                                struct sdma_txreq *tx,
 365                                int res)
 366 {
 367         /* protect against complete modifying */
 368         struct iowait *wait = tx->wait;
 369         callback_t complete = tx->complete;
 370
 371 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
 372         trace_hfi1_sdma_out_sn(sde, tx->sn);
 373         if (WARN_ON_ONCE(sde->head_sn != tx->sn))
 374                 dd_dev_err(sde->dd, "expected %llu got %llu\n",
 375                            sde->head_sn, tx->sn);
 376         sde->head_sn++;
 377 #endif
 378         __sdma_txclean(sde->dd, tx);
 379         if (complete)
 380                 (*complete)(tx, res);
 381         if (wait && iowait_sdma_dec(wait))
 382                 iowait_drain_wakeup(wait);
 383 }
 384
 385 /*
 386  * Complete all the sdma requests with a SDMA_TXREQ_S_ABORTED status
 387  *
 388  * Depending on timing there can be txreqs in two places:
 389  * - in the descq ring
 390  * - in the flush list
 391  *
 392  * To avoid ordering issues the descq ring needs to be flushed
 393  * first followed by the flush list.
 394  *
 395  * This routine is called from two places
 396  * - From a work queue item
 397  * - Directly from the state machine just before setting the
 398  *   state to running
 399  *
 400  * Must be called with head_lock held
 401  *
 402  */
 403 static void sdma_flush(struct sdma_engine *sde)
 404 {
 405         struct sdma_txreq *txp, *txp_next;
 406         LIST_HEAD(flushlist);
 407         unsigned long flags;
 408
 409         /* flush from head to tail */
 410         sdma_flush_descq(sde);
 411         spin_lock_irqsave(&sde->flushlist_lock, flags);
 412         /* copy flush list */
 413         list_for_each_entry_safe(txp, txp_next, &sde->flushlist, list) {
 414                 list_del_init(&txp->list);
 415                 list_add_tail(&txp->list, &flushlist);
 416         }
 417         spin_unlock_irqrestore(&sde->flushlist_lock, flags);
 418         /* flush from flush list */
 419         list_for_each_entry_safe(txp, txp_next, &flushlist, list)
 420                 complete_tx(sde, txp, SDMA_TXREQ_S_ABORTED);
 421 }
 422
 423 /*
 424  * Fields a work request for flushing the descq ring
 425  * and the flush list
 426  *
 427  * If the engine has been brought to running during
 428  * the scheduling delay, the flush is ignored, assuming
 429  * that the process of bringing the engine to running
 430  * would have done this flush prior to going to running.
 431  *
 432  */
 433 static void sdma_field_flush(struct work_struct *work)
 434 {
 435         unsigned long flags;
 436         struct sdma_engine *sde =
 437                 container_of(work, struct sdma_engine, flush_worker);
 438
 439         write_seqlock_irqsave(&sde->head_lock, flags);
 440         if (!__sdma_running(sde))
 441                 sdma_flush(sde);
 442         write_sequnlock_irqrestore(&sde->head_lock, flags);
 443 }
 444
 445 static void sdma_err_halt_wait(struct work_struct *work)
 446 {
 447         struct sdma_engine *sde = container_of(work, struct sdma_engine,
 448                                                 err_halt_worker);
 449         u64 statuscsr;
 450         unsigned long timeout;
 451
 452         timeout = jiffies + msecs_to_jiffies(SDMA_ERR_HALT_TIMEOUT);
 453         while (1) {
 454                 statuscsr = read_sde_csr(sde, SD(STATUS));
 455                 statuscsr &= SD(STATUS_ENG_HALTED_SMASK);
 456                 if (statuscsr)
 457                         break;
 458                 if (time_after(jiffies, timeout)) {
 459                         dd_dev_err(sde->dd,
 460                                    "SDMA engine %d - timeout waiting for engine to halt\n",
 461                                    sde->this_idx);
 462                         /*
 463                          * Continue anyway.  This could happen if there was
 464                          * an uncorrectable error in the wrong spot.
 465                          */
 466                         break;
 467                 }
 468                 usleep_range(80, 120);
 469         }
 470
 471         sdma_process_event(sde, sdma_event_e15_hw_halt_done);
 472 }
 473
 474 static void sdma_err_progress_check_schedule(struct sdma_engine *sde)
 475 {
 476         if (!is_bx(sde->dd) && HFI1_CAP_IS_KSET(SDMA_AHG)) {
 477                 unsigned index;
 478                 struct hfi1_devdata *dd = sde->dd;
 479
 480                 for (index = 0; index < dd->num_sdma; index++) {
 481                         struct sdma_engine *curr_sdma = &dd->per_sdma[index];
 482
 483                         if (curr_sdma != sde)
 484                                 curr_sdma->progress_check_head =
 485                                                         curr_sdma->descq_head;
 486                 }
 487                 dd_dev_err(sde->dd,
 488                            "SDMA engine %d - check scheduled\n",
 489                                 sde->this_idx);
 490                 mod_timer(&sde->err_progress_check_timer, jiffies + 10);
 491         }
 492 }
 493
 494 static void sdma_err_progress_check(struct timer_list *t)
 495 {
 496         unsigned index;
 497         struct sdma_engine *sde = from_timer(sde, t, err_progress_check_timer);
 498
 499         dd_dev_err(sde->dd, "SDE progress check event\n");
 500         for (index = 0; index < sde->dd->num_sdma; index++) {
 501                 struct sdma_engine *curr_sde = &sde->dd->per_sdma[index];
 502                 unsigned long flags;
 503
 504                 /* check progress on each engine except the current one */
 505                 if (curr_sde == sde)
 506                         continue;
 507                 /*
 508                  * We must lock interrupts when acquiring sde->lock,
 509                  * to avoid a deadlock if interrupt triggers and spins on
 510                  * the same lock on same CPU
 511                  */
 512                 spin_lock_irqsave(&curr_sde->tail_lock, flags);
 513                 write_seqlock(&curr_sde->head_lock);
 514
 515                 /* skip non-running queues */
 516                 if (curr_sde->state.current_state != sdma_state_s99_running) {
 517                         write_sequnlock(&curr_sde->head_lock);
 518                         spin_unlock_irqrestore(&curr_sde->tail_lock, flags);
 519                         continue;
 520                 }
 521
 522                 if ((curr_sde->descq_head != curr_sde->descq_tail) &&
 523                     (curr_sde->descq_head ==
 524                                 curr_sde->progress_check_head))
 525                         __sdma_process_event(curr_sde,
 526                                              sdma_event_e90_sw_halted);
 527                 write_sequnlock(&curr_sde->head_lock);
 528                 spin_unlock_irqrestore(&curr_sde->tail_lock, flags);
 529         }
 530         schedule_work(&sde->err_halt_worker);
 531 }
 532
 533 static void sdma_hw_clean_up_task(unsigned long opaque)
 534 {
 535         struct sdma_engine *sde = (struct sdma_engine *)opaque;
 536         u64 statuscsr;
 537
 538         while (1) {
 539 #ifdef CONFIG_SDMA_VERBOSITY
 540                 dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
 541                            sde->this_idx, slashstrip(__FILE__), __LINE__,
 542                         __func__);
 543 #endif
 544                 statuscsr = read_sde_csr(sde, SD(STATUS));
 545                 statuscsr &= SD(STATUS_ENG_CLEANED_UP_SMASK);
 546                 if (statuscsr)
 547                         break;
 548                 udelay(10);
 549         }
 550
 551         sdma_process_event(sde, sdma_event_e25_hw_clean_up_done);
 552 }
 553
 554 static inline struct sdma_txreq *get_txhead(struct sdma_engine *sde)
 555 {
 556         return sde->tx_ring[sde->tx_head & sde->sdma_mask];
 557 }
 558
 559 /*
 560  * flush ring for recovery
 561  */
 562 static void sdma_flush_descq(struct sdma_engine *sde)
 563 {
 564         u16 head, tail;
 565         int progress = 0;
 566         struct sdma_txreq *txp = get_txhead(sde);
 567
 568         /* The reason for some of the complexity of this code is that
 569          * not all descriptors have corresponding txps.  So, we have to
 570          * be able to skip over descs until we wander into the range of
 571          * the next txp on the list.
 572          */
 573         head = sde->descq_head & sde->sdma_mask;
 574         tail = sde->descq_tail & sde->sdma_mask;
 575         while (head != tail) {
 576                 /* advance head, wrap if needed */
 577                 head = ++sde->descq_head & sde->sdma_mask;
 578                 /* if now past this txp's descs, do the callback */
 579                 if (txp && txp->next_descq_idx == head) {
 580                         /* remove from list */
 581                         sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL;
 582                         complete_tx(sde, txp, SDMA_TXREQ_S_ABORTED);
 583                         trace_hfi1_sdma_progress(sde, head, tail, txp);
 584                         txp = get_txhead(sde);
 585                 }
 586                 progress++;
 587         }
 588         if (progress)
 589                 sdma_desc_avail(sde, sdma_descq_freecnt(sde));
 590 }
 591
 592 static void sdma_sw_clean_up_task(unsigned long opaque)
 593 {
 594         struct sdma_engine *sde = (struct sdma_engine *)opaque;
 595         unsigned long flags;
 596
 597         spin_lock_irqsave(&sde->tail_lock, flags);
 598         write_seqlock(&sde->head_lock);
 599
 600         /*
 601          * At this point, the following should always be true:
 602          * - We are halted, so no more descriptors are getting retired.
 603          * - We are not running, so no one is submitting new work.
 604          * - Only we can send the e40_sw_cleaned, so we can't start
 605          *   running again until we say so.  So, the active list and
 606          *   descq are ours to play with.
 607          */
 608
 609         /*
 610          * In the error clean up sequence, software clean must be called
 611          * before the hardware clean so we can use the hardware head in
 612          * the progress routine.  A hardware clean or SPC unfreeze will
 613          * reset the hardware head.
 614          *
 615          * Process all retired requests. The progress routine will use the
 616          * latest physical hardware head - we are not running so speed does
 617          * not matter.
 618          */
 619         sdma_make_progress(sde, 0);
 620
 621         sdma_flush(sde);
 622
 623         /*
 624          * Reset our notion of head and tail.
 625          * Note that the HW registers have been reset via an earlier
 626          * clean up.
 627          */
 628         sde->descq_tail = 0;
 629         sde->descq_head = 0;
 630         sde->desc_avail = sdma_descq_freecnt(sde);
 631         *sde->head_dma = 0;
 632
 633         __sdma_process_event(sde, sdma_event_e40_sw_cleaned);
 634
 635         write_sequnlock(&sde->head_lock);
 636         spin_unlock_irqrestore(&sde->tail_lock, flags);
 637 }
 638
 639 static void sdma_sw_tear_down(struct sdma_engine *sde)
 640 {
 641         struct sdma_state *ss = &sde->state;
 642
 643         /* Releasing this reference means the state machine has stopped. */
 644         sdma_put(ss);
 645
 646         /* stop waiting for all unfreeze events to complete */
 647         atomic_set(&sde->dd->sdma_unfreeze_count, -1);
 648         wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
 649 }
 650
 651 static void sdma_start_hw_clean_up(struct sdma_engine *sde)
 652 {
 653         tasklet_hi_schedule(&sde->sdma_hw_clean_up_task);
 654 }
 655
 656 static void sdma_set_state(struct sdma_engine *sde,
 657                            enum sdma_states next_state)
 658 {
 659         struct sdma_state *ss = &sde->state;
 660         const struct sdma_set_state_action *action = sdma_action_table;
 661         unsigned op = 0;
 662
 663         trace_hfi1_sdma_state(
 664                 sde,
 665                 sdma_state_names[ss->current_state],
 666                 sdma_state_names[next_state]);
 667
 668         /* debugging bookkeeping */
 669         ss->previous_state = ss->current_state;
 670         ss->previous_op = ss->current_op;
 671         ss->current_state = next_state;
 672
 673         if (ss->previous_state != sdma_state_s99_running &&
 674             next_state == sdma_state_s99_running)
 675                 sdma_flush(sde);
 676
 677         if (action[next_state].op_enable)
 678                 op |= SDMA_SENDCTRL_OP_ENABLE;
 679
 680         if (action[next_state].op_intenable)
 681                 op |= SDMA_SENDCTRL_OP_INTENABLE;
 682
 683         if (action[next_state].op_halt)
 684                 op |= SDMA_SENDCTRL_OP_HALT;
 685
 686         if (action[next_state].op_cleanup)
 687                 op |= SDMA_SENDCTRL_OP_CLEANUP;
 688
 689         if (action[next_state].go_s99_running_tofalse)
 690                 ss->go_s99_running = 0;
 691
 692         if (action[next_state].go_s99_running_totrue)
 693                 ss->go_s99_running = 1;
 694
 695         ss->current_op = op;
 696         sdma_sendctrl(sde, ss->current_op);
 697 }
 698
 699 /**
 700  * sdma_get_descq_cnt() - called when device probed
 701  *
 702  * Return a validated descq count.
 703  *
 704  * This is currently only used in the verbs initialization to build the tx
 705  * list.
 706  *
 707  * This will probably be deleted in favor of a more scalable approach to
 708  * alloc tx's.
 709  *
 710  */
 711 u16 sdma_get_descq_cnt(void)
 712 {
 713         u16 count = sdma_descq_cnt;
 714
 715         if (!count)
 716                 return SDMA_DESCQ_CNT;
 717         /* count must be a power of 2 greater than 64 and less than
 718          * 32768.   Otherwise return default.
 719          */
 720         if (!is_power_of_2(count))
 721                 return SDMA_DESCQ_CNT;
 722         if (count < 64 || count > 32768)
 723                 return SDMA_DESCQ_CNT;
 724         return count;
 725 }
 726
 727 /**
 728  * sdma_engine_get_vl() - return vl for a given sdma engine
 729  * @sde: sdma engine
 730  *
 731  * This function returns the vl mapped to a given engine, or an error if
 732  * the mapping can't be found. The mapping fields are protected by RCU.
 733  */
 734 int sdma_engine_get_vl(struct sdma_engine *sde)
 735 {
 736         struct hfi1_devdata *dd = sde->dd;
 737         struct sdma_vl_map *m;
 738         u8 vl;
 739
 740         if (sde->this_idx >= TXE_NUM_SDMA_ENGINES)
 741                 return -EINVAL;
 742
 743         rcu_read_lock();
 744         m = rcu_dereference(dd->sdma_map);
 745         if (unlikely(!m)) {
 746                 rcu_read_unlock();
 747                 return -EINVAL;
 748         }
 749         vl = m->engine_to_vl[sde->this_idx];
 750         rcu_read_unlock();
 751
 752         return vl;
 753 }
 754
 755 /**
 756  * sdma_select_engine_vl() - select sdma engine
 757  * @dd: devdata
 758  * @selector: a spreading factor
 759  * @vl: this vl
 760  *
 761  *
 762  * This function returns an engine based on the selector and a vl.  The
 763  * mapping fields are protected by RCU.
 764  */
 765 struct sdma_engine *sdma_select_engine_vl(
 766         struct hfi1_devdata *dd,
 767         u32 selector,
 768         u8 vl)
 769 {
 770         struct sdma_vl_map *m;
 771         struct sdma_map_elem *e;
 772         struct sdma_engine *rval;
 773
 774         /* NOTE This should only happen if SC->VL changed after the initial
 775          *      checks on the QP/AH
 776          *      Default will return engine 0 below
 777          */
 778         if (vl >= num_vls) {
 779                 rval = NULL;
 780                 goto done;
 781         }
 782
 783         rcu_read_lock();
 784         m = rcu_dereference(dd->sdma_map);
 785         if (unlikely(!m)) {
 786                 rcu_read_unlock();
 787                 return &dd->per_sdma[0];
 788         }
 789         e = m->map[vl & m->mask];
 790         rval = e->sde[selector & e->mask];
 791         rcu_read_unlock();
 792
 793 done:
 794         rval =  !rval ? &dd->per_sdma[0] : rval;
 795         trace_hfi1_sdma_engine_select(dd, selector, vl, rval->this_idx);
 796         return rval;
 797 }
 798
 799 /**
 800  * sdma_select_engine_sc() - select sdma engine
 801  * @dd: devdata
 802  * @selector: a spreading factor
 803  * @sc5: the 5 bit sc
 804  *
 805  *
 806  * This function returns an engine based on the selector and an sc.
 807  */
 808 struct sdma_engine *sdma_select_engine_sc(
 809         struct hfi1_devdata *dd,
 810         u32 selector,
 811         u8 sc5)
 812 {
 813         u8 vl = sc_to_vlt(dd, sc5);
 814
 815         return sdma_select_engine_vl(dd, selector, vl);
 816 }
 817
 818 struct sdma_rht_map_elem {
 819         u32 mask;
 820         u8 ctr;
 821         struct sdma_engine *sde[0];
 822 };
 823
 824 struct sdma_rht_node {
 825         unsigned long cpu_id;
 826         struct sdma_rht_map_elem *map[HFI1_MAX_VLS_SUPPORTED];
 827         struct rhash_head node;
 828 };
 829
 830 #define NR_CPUS_HINT 192
 831
 832 static const struct rhashtable_params sdma_rht_params = {
 833         .nelem_hint = NR_CPUS_HINT,
 834         .head_offset = offsetof(struct sdma_rht_node, node),
 835         .key_offset = offsetof(struct sdma_rht_node, cpu_id),
 836         .key_len = FIELD_SIZEOF(struct sdma_rht_node, cpu_id),
 837         .max_size = NR_CPUS,
 838         .min_size = 8,
 839         .automatic_shrinking = true,
 840 };
 841
 842 /*
 843  * sdma_select_user_engine() - select sdma engine based on user setup
 844  * @dd: devdata
 845  * @selector: a spreading factor
 846  * @vl: this vl
 847  *
 848  * This function returns an sdma engine for a user sdma request.
 849  * User defined sdma engine affinity setting is honored when applicable,
 850  * otherwise system default sdma engine mapping is used. To ensure correct
 851  * ordering, the mapping from <selector, vl> to sde must remain unchanged.
 852  */
 853 struct sdma_engine *sdma_select_user_engine(struct hfi1_devdata *dd,
 854                                             u32 selector, u8 vl)
 855 {
 856         struct sdma_rht_node *rht_node;
 857         struct sdma_engine *sde = NULL;
 858         const struct cpumask *current_mask = &current->cpus_allowed;
 859         unsigned long cpu_id;
 860
 861         /*
 862          * To ensure that always the same sdma engine(s) will be
 863          * selected make sure the process is pinned to this CPU only.
 864          */
 865         if (cpumask_weight(current_mask) != 1)
 866                 goto out;
 867
 868         cpu_id = smp_processor_id();
 869         rcu_read_lock();
 870         rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpu_id,
 871                                           sdma_rht_params);
 872
 873         if (rht_node && rht_node->map[vl]) {
 874                 struct sdma_rht_map_elem *map = rht_node->map[vl];
 875
 876                 sde = map->sde[selector & map->mask];
 877         }
 878         rcu_read_unlock();
 879
 880         if (sde)
 881                 return sde;
 882
 883 out:
 884         return sdma_select_engine_vl(dd, selector, vl);
 885 }
 886
 887 static void sdma_populate_sde_map(struct sdma_rht_map_elem *map)
 888 {
 889         int i;
 890
 891         for (i = 0; i < roundup_pow_of_two(map->ctr ? : 1) - map->ctr; i++)
 892                 map->sde[map->ctr + i] = map->sde[i];
 893 }
 894
 895 static void sdma_cleanup_sde_map(struct sdma_rht_map_elem *map,
 896                                  struct sdma_engine *sde)
 897 {
 898         unsigned int i, pow;
 899
 900         /* only need to check the first ctr entries for a match */
 901         for (i = 0; i < map->ctr; i++) {
 902                 if (map->sde[i] == sde) {
 903                         memmove(&map->sde[i], &map->sde[i + 1],
 904                                 (map->ctr - i - 1) * sizeof(map->sde[0]));
 905                         map->ctr--;
 906                         pow = roundup_pow_of_two(map->ctr ? : 1);
 907                         map->mask = pow - 1;
 908                         sdma_populate_sde_map(map);
 909                         break;
 910                 }
 911         }
 912 }
 913
 914 /*
 915  * Prevents concurrent reads and writes of the sdma engine cpu_mask
 916  */
 917 static DEFINE_MUTEX(process_to_sde_mutex);
 918
 919 ssize_t sdma_set_cpu_to_sde_map(struct sdma_engine *sde, const char *buf,
 920                                 size_t count)
 921 {
 922         struct hfi1_devdata *dd = sde->dd;
 923         cpumask_var_t mask, new_mask;
 924         unsigned long cpu;
 925         int ret, vl, sz;
 926
 927         vl = sdma_engine_get_vl(sde);
 928         if (unlikely(vl < 0))
 929                 return -EINVAL;
 930
 931         ret = zalloc_cpumask_var(&mask, GFP_KERNEL);
 932         if (!ret)
 933                 return -ENOMEM;
 934
 935         ret = zalloc_cpumask_var(&new_mask, GFP_KERNEL);
 936         if (!ret) {
 937                 free_cpumask_var(mask);
 938                 return -ENOMEM;
 939         }
 940         ret = cpulist_parse(buf, mask);
 941         if (ret)
 942                 goto out_free;
 943
 944         if (!cpumask_subset(mask, cpu_online_mask)) {
 945                 dd_dev_warn(sde->dd, "Invalid CPU mask\n");
 946                 ret = -EINVAL;
 947                 goto out_free;
 948         }
 949
 950         sz = sizeof(struct sdma_rht_map_elem) +
 951                         (TXE_NUM_SDMA_ENGINES * sizeof(struct sdma_engine *));
 952
 953         mutex_lock(&process_to_sde_mutex);
 954
 955         for_each_cpu(cpu, mask) {
 956                 struct sdma_rht_node *rht_node;
 957
 958                 /* Check if we have this already mapped */
 959                 if (cpumask_test_cpu(cpu, &sde->cpu_mask)) {
 960                         cpumask_set_cpu(cpu, new_mask);
 961                         continue;
 962                 }
 963
 964                 if (vl >= ARRAY_SIZE(rht_node->map)) {
 965                         ret = -EINVAL;
 966                         goto out;
 967                 }
 968
 969                 rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpu,
 970                                                   sdma_rht_params);
 971                 if (!rht_node) {
 972                         rht_node = kzalloc(sizeof(*rht_node), GFP_KERNEL);
 973                         if (!rht_node) {
 974                                 ret = -ENOMEM;
 975                                 goto out;
 976                         }
 977
 978                         rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
 979                         if (!rht_node->map[vl]) {
 980                                 kfree(rht_node);
 981                                 ret = -ENOMEM;
 982                                 goto out;
 983                         }
 984                         rht_node->cpu_id = cpu;
 985                         rht_node->map[vl]->mask = 0;
 986                         rht_node->map[vl]->ctr = 1;
 987                         rht_node->map[vl]->sde[0] = sde;
 988
 989                         ret = rhashtable_insert_fast(dd->sdma_rht,
 990                                                      &rht_node->node,
 991                                                      sdma_rht_params);
 992                         if (ret) {
 993                                 kfree(rht_node->map[vl]);
 994                                 kfree(rht_node);
 995                                 dd_dev_err(sde->dd, "Failed to set process to sde affinity for cpu %lu\n",
 996                                            cpu);
 997                                 goto out;
 998                         }
 999
1000                 } else {
1001                         int ctr, pow;
1002
1003                         /* Add new user mappings */
1004                         if (!rht_node->map[vl])
1005                                 rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
1006
1007                         if (!rht_node->map[vl]) {
1008                                 ret = -ENOMEM;
1009                                 goto out;
1010                         }
1011
1012                         rht_node->map[vl]->ctr++;
1013                         ctr = rht_node->map[vl]->ctr;
1014                         rht_node->map[vl]->sde[ctr - 1] = sde;
1015                         pow = roundup_pow_of_two(ctr);
1016                         rht_node->map[vl]->mask = pow - 1;
1017
1018                         /* Populate the sde map table */
1019                         sdma_populate_sde_map(rht_node->map[vl]);
1020                 }
1021                 cpumask_set_cpu(cpu, new_mask);
1022         }
1023
1024         /* Clean up old mappings */
1025         for_each_cpu(cpu, cpu_online_mask) {
1026                 struct sdma_rht_node *rht_node;
1027
1028                 /* Don't cleanup sdes that are set in the new mask */
1029                 if (cpumask_test_cpu(cpu, mask))
1030                         continue;
1031
1032                 rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpu,
1033                                                   sdma_rht_params);
1034                 if (rht_node) {
1035                         bool empty = true;
1036                         int i;
1037
1038                         /* Remove mappings for old sde */
1039                         for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1040                                 if (rht_node->map[i])
1041                                         sdma_cleanup_sde_map(rht_node->map[i],
1042                                                              sde);
1043
1044                         /* Free empty hash table entries */
1045                         for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
1046                                 if (!rht_node->map[i])
1047                                         continue;
1048
1049                                 if (rht_node->map[i]->ctr) {
1050                                         empty = false;
1051                                         break;
1052                                 }
1053                         }
1054
1055                         if (empty) {
1056                                 ret = rhashtable_remove_fast(dd->sdma_rht,
1057                                                              &rht_node->node,
1058                                                              sdma_rht_params);
1059                                 WARN_ON(ret);
1060
1061                                 for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1062                                         kfree(rht_node->map[i]);
1063
1064                                 kfree(rht_node);
1065                         }
1066                 }
1067         }
1068
1069         cpumask_copy(&sde->cpu_mask, new_mask);
1070 out:
1071         mutex_unlock(&process_to_sde_mutex);
1072 out_free:
1073         free_cpumask_var(mask);
1074         free_cpumask_var(new_mask);
1075         return ret ? : strnlen(buf, PAGE_SIZE);
1076 }
1077
1078 ssize_t sdma_get_cpu_to_sde_map(struct sdma_engine *sde, char *buf)
1079 {
1080         mutex_lock(&process_to_sde_mutex);
1081         if (cpumask_empty(&sde->cpu_mask))
1082                 snprintf(buf, PAGE_SIZE, "%s\n", "empty");
1083         else
1084                 cpumap_print_to_pagebuf(true, buf, &sde->cpu_mask);
1085         mutex_unlock(&process_to_sde_mutex);
1086         return strnlen(buf, PAGE_SIZE);
1087 }
1088
1089 static void sdma_rht_free(void *ptr, void *arg)
1090 {
1091         struct sdma_rht_node *rht_node = ptr;
1092         int i;
1093
1094         for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1095                 kfree(rht_node->map[i]);
1096
1097         kfree(rht_node);
1098 }
1099
1100 /**
1101  * sdma_seqfile_dump_cpu_list() - debugfs dump the cpu to sdma mappings
1102  * @s: seq file
1103  * @dd: hfi1_devdata
1104  * @cpuid: cpu id
1105  *
1106  * This routine dumps the process to sde mappings per cpu
1107  */
1108 void sdma_seqfile_dump_cpu_list(struct seq_file *s,
1109                                 struct hfi1_devdata *dd,
1110                                 unsigned long cpuid)
1111 {
1112         struct sdma_rht_node *rht_node;
1113         int i, j;
1114
1115         rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpuid,
1116                                           sdma_rht_params);
1117         if (!rht_node)
1118                 return;
1119
1120         seq_printf(s, "cpu%3lu: ", cpuid);
1121         for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
1122                 if (!rht_node->map[i] || !rht_node->map[i]->ctr)
1123                         continue;
1124
1125                 seq_printf(s, " vl%d: [", i);
1126
1127                 for (j = 0; j < rht_node->map[i]->ctr; j++) {
1128                         if (!rht_node->map[i]->sde[j])
1129                                 continue;
1130
1131                         if (j > 0)
1132                                 seq_puts(s, ",");
1133
1134                         seq_printf(s, " sdma%2d",
1135                                    rht_node->map[i]->sde[j]->this_idx);
1136                 }
1137                 seq_puts(s, " ]");
1138         }
1139
1140         seq_puts(s, "\n");
1141 }
1142
1143 /*
1144  * Free the indicated map struct
1145  */
1146 static void sdma_map_free(struct sdma_vl_map *m)
1147 {
1148         int i;
1149
1150         for (i = 0; m && i < m->actual_vls; i++)
1151                 kfree(m->map[i]);
1152         kfree(m);
1153 }
1154
1155 /*
1156  * Handle RCU callback
1157  */
1158 static void sdma_map_rcu_callback(struct rcu_head *list)
1159 {
1160         struct sdma_vl_map *m = container_of(list, struct sdma_vl_map, list);
1161
1162         sdma_map_free(m);
1163 }
1164
1165 /**
1166  * sdma_map_init - called when # vls change
1167  * @dd: hfi1_devdata
1168  * @port: port number
1169  * @num_vls: number of vls
1170  * @vl_engines: per vl engine mapping (optional)
1171  *
1172  * This routine changes the mapping based on the number of vls.
1173  *
1174  * vl_engines is used to specify a non-uniform vl/engine loading. NULL
1175  * implies auto computing the loading and giving each VLs a uniform
1176  * distribution of engines per VL.
1177  *
1178  * The auto algorithm computes the sde_per_vl and the number of extra
1179  * engines.  Any extra engines are added from the last VL on down.
1180  *
1181  * rcu locking is used here to control access to the mapping fields.
1182  *
1183  * If either the num_vls or num_sdma are non-power of 2, the array sizes
1184  * in the struct sdma_vl_map and the struct sdma_map_elem are rounded
1185  * up to the next highest power of 2 and the first entry is reused
1186  * in a round robin fashion.
1187  *
1188  * If an error occurs the map change is not done and the mapping is
1189  * not changed.
1190  *
1191  */
1192 int sdma_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_engines)
1193 {
1194         int i, j;
1195         int extra, sde_per_vl;
1196         int engine = 0;
1197         u8 lvl_engines[OPA_MAX_VLS];
1198         struct sdma_vl_map *oldmap, *newmap;
1199
1200         if (!(dd->flags & HFI1_HAS_SEND_DMA))
1201                 return 0;
1202
1203         if (!vl_engines) {
1204                 /* truncate divide */
1205                 sde_per_vl = dd->num_sdma / num_vls;
1206                 /* extras */
1207                 extra = dd->num_sdma % num_vls;
1208                 vl_engines = lvl_engines;
1209                 /* add extras from last vl down */
1210                 for (i = num_vls - 1; i >= 0; i--, extra--)
1211                         vl_engines[i] = sde_per_vl + (extra > 0 ? 1 : 0);
1212         }
1213         /* build new map */
1214         newmap = kzalloc(
1215                 sizeof(struct sdma_vl_map) +
1216                         roundup_pow_of_two(num_vls) *
1217                         sizeof(struct sdma_map_elem *),
1218                 GFP_KERNEL);
1219         if (!newmap)
1220                 goto bail;
1221         newmap->actual_vls = num_vls;
1222         newmap->vls = roundup_pow_of_two(num_vls);
1223         newmap->mask = (1 << ilog2(newmap->vls)) - 1;
1224         /* initialize back-map */
1225         for (i = 0; i < TXE_NUM_SDMA_ENGINES; i++)
1226                 newmap->engine_to_vl[i] = -1;
1227         for (i = 0; i < newmap->vls; i++) {
1228                 /* save for wrap around */
1229                 int first_engine = engine;
1230
1231                 if (i < newmap->actual_vls) {
1232                         int sz = roundup_pow_of_two(vl_engines[i]);
1233
1234                         /* only allocate once */
1235                         newmap->map[i] = kzalloc(
1236                                 sizeof(struct sdma_map_elem) +
1237                                         sz * sizeof(struct sdma_engine *),
1238                                 GFP_KERNEL);
1239                         if (!newmap->map[i])
1240                                 goto bail;
1241                         newmap->map[i]->mask = (1 << ilog2(sz)) - 1;
1242                         /* assign engines */
1243                         for (j = 0; j < sz; j++) {
1244                                 newmap->map[i]->sde[j] =
1245                                         &dd->per_sdma[engine];
1246                                 if (++engine >= first_engine + vl_engines[i])
1247                                         /* wrap back to first engine */
1248                                         engine = first_engine;
1249                         }
1250                         /* assign back-map */
1251                         for (j = 0; j < vl_engines[i]; j++)
1252                                 newmap->engine_to_vl[first_engine + j] = i;
1253                 } else {
1254                         /* just re-use entry without allocating */
1255                         newmap->map[i] = newmap->map[i % num_vls];
1256                 }
1257                 engine = first_engine + vl_engines[i];
1258         }
1259         /* newmap in hand, save old map */
1260         spin_lock_irq(&dd->sde_map_lock);
1261         oldmap = rcu_dereference_protected(dd->sdma_map,
1262                                            lockdep_is_held(&dd->sde_map_lock));
1263
1264         /* publish newmap */
1265         rcu_assign_pointer(dd->sdma_map, newmap);
1266
1267         spin_unlock_irq(&dd->sde_map_lock);
1268         /* success, free any old map after grace period */
1269         if (oldmap)
1270                 call_rcu(&oldmap->list, sdma_map_rcu_callback);
1271         return 0;
1272 bail:
1273         /* free any partial allocation */
1274         sdma_map_free(newmap);
1275         return -ENOMEM;
1276 }
1277
1278 /*
1279  * Clean up allocated memory.
1280  *
1281  * This routine is can be called regardless of the success of sdma_init()
1282  *
1283  */
1284 static void sdma_clean(struct hfi1_devdata *dd, size_t num_engines)
1285 {
1286         size_t i;
1287         struct sdma_engine *sde;
1288
1289         if (dd->sdma_pad_dma) {
1290                 dma_free_coherent(&dd->pcidev->dev, 4,
1291                                   (void *)dd->sdma_pad_dma,
1292                                   dd->sdma_pad_phys);
1293                 dd->sdma_pad_dma = NULL;
1294                 dd->sdma_pad_phys = 0;
1295         }
1296         if (dd->sdma_heads_dma) {
1297                 dma_free_coherent(&dd->pcidev->dev, dd->sdma_heads_size,
1298                                   (void *)dd->sdma_heads_dma,
1299                                   dd->sdma_heads_phys);
1300                 dd->sdma_heads_dma = NULL;
1301                 dd->sdma_heads_phys = 0;
1302         }
1303         for (i = 0; dd->per_sdma && i < num_engines; ++i) {
1304                 sde = &dd->per_sdma[i];
1305
1306                 sde->head_dma = NULL;
1307                 sde->head_phys = 0;
1308
1309                 if (sde->descq) {
1310                         dma_free_coherent(
1311                                 &dd->pcidev->dev,
1312                                 sde->descq_cnt * sizeof(u64[2]),
1313                                 sde->descq,
1314                                 sde->descq_phys
1315                         );
1316                         sde->descq = NULL;
1317                         sde->descq_phys = 0;
1318                 }
1319                 kvfree(sde->tx_ring);
1320                 sde->tx_ring = NULL;
1321         }
1322         spin_lock_irq(&dd->sde_map_lock);
1323         sdma_map_free(rcu_access_pointer(dd->sdma_map));
1324         RCU_INIT_POINTER(dd->sdma_map, NULL);
1325         spin_unlock_irq(&dd->sde_map_lock);
1326         synchronize_rcu();
1327         kfree(dd->per_sdma);
1328         dd->per_sdma = NULL;
1329
1330         if (dd->sdma_rht) {
1331                 rhashtable_free_and_destroy(dd->sdma_rht, sdma_rht_free, NULL);
1332                 kfree(dd->sdma_rht);
1333                 dd->sdma_rht = NULL;
1334         }
1335 }
1336
1337 /**
1338  * sdma_init() - called when device probed
1339  * @dd: hfi1_devdata
1340  * @port: port number (currently only zero)
1341  *
1342  * Initializes each sde and its csrs.
1343  * Interrupts are not required to be enabled.
1344  *
1345  * Returns:
1346  * 0 - success, -errno on failure
1347  */
1348 int sdma_init(struct hfi1_devdata *dd, u8 port)
1349 {
1350         unsigned this_idx;
1351         struct sdma_engine *sde;
1352         struct rhashtable *tmp_sdma_rht;
1353         u16 descq_cnt;
1354         void *curr_head;
1355         struct hfi1_pportdata *ppd = dd->pport + port;
1356         u32 per_sdma_credits;
1357         uint idle_cnt = sdma_idle_cnt;
1358         size_t num_engines = dd->chip_sdma_engines;
1359         int ret = -ENOMEM;
1360
1361         if (!HFI1_CAP_IS_KSET(SDMA)) {
1362                 HFI1_CAP_CLEAR(SDMA_AHG);
1363                 return 0;
1364         }
1365         if (mod_num_sdma &&
1366             /* can't exceed chip support */
1367             mod_num_sdma <= dd->chip_sdma_engines &&
1368             /* count must be >= vls */
1369             mod_num_sdma >= num_vls)
1370                 num_engines = mod_num_sdma;
1371
1372         dd_dev_info(dd, "SDMA mod_num_sdma: %u\n", mod_num_sdma);
1373         dd_dev_info(dd, "SDMA chip_sdma_engines: %u\n", dd->chip_sdma_engines);
1374         dd_dev_info(dd, "SDMA chip_sdma_mem_size: %u\n",
1375                     dd->chip_sdma_mem_size);
1376
1377         per_sdma_credits =
1378                 dd->chip_sdma_mem_size / (num_engines * SDMA_BLOCK_SIZE);
1379
1380         /* set up freeze waitqueue */
1381         init_waitqueue_head(&dd->sdma_unfreeze_wq);
1382         atomic_set(&dd->sdma_unfreeze_count, 0);
1383
1384         descq_cnt = sdma_get_descq_cnt();
1385         dd_dev_info(dd, "SDMA engines %zu descq_cnt %u\n",
1386                     num_engines, descq_cnt);
1387
1388         /* alloc memory for array of send engines */
1389         dd->per_sdma = kcalloc(num_engines, sizeof(*dd->per_sdma), GFP_KERNEL);
1390         if (!dd->per_sdma)
1391                 return ret;
1392
1393         idle_cnt = ns_to_cclock(dd, idle_cnt);
1394         if (idle_cnt)
1395                 dd->default_desc1 =
1396                         SDMA_DESC1_HEAD_TO_HOST_FLAG;
1397         else
1398                 dd->default_desc1 =
1399                         SDMA_DESC1_INT_REQ_FLAG;
1400
1401         if (!sdma_desct_intr)
1402                 sdma_desct_intr = SDMA_DESC_INTR;
1403
1404         /* Allocate memory for SendDMA descriptor FIFOs */
1405         for (this_idx = 0; this_idx < num_engines; ++this_idx) {
1406                 sde = &dd->per_sdma[this_idx];
1407                 sde->dd = dd;
1408                 sde->ppd = ppd;
1409                 sde->this_idx = this_idx;
1410                 sde->descq_cnt = descq_cnt;
1411                 sde->desc_avail = sdma_descq_freecnt(sde);
1412                 sde->sdma_shift = ilog2(descq_cnt);
1413                 sde->sdma_mask = (1 << sde->sdma_shift) - 1;
1414
1415                 /* Create a mask specifically for each interrupt source */
1416                 sde->int_mask = (u64)1 << (0 * TXE_NUM_SDMA_ENGINES +
1417                                            this_idx);
1418                 sde->progress_mask = (u64)1 << (1 * TXE_NUM_SDMA_ENGINES +
1419                                                 this_idx);
1420                 sde->idle_mask = (u64)1 << (2 * TXE_NUM_SDMA_ENGINES +
1421                                             this_idx);
1422                 /* Create a combined mask to cover all 3 interrupt sources */
1423                 sde->imask = sde->int_mask | sde->progress_mask |
1424                              sde->idle_mask;
1425
1426                 spin_lock_init(&sde->tail_lock);
1427                 seqlock_init(&sde->head_lock);
1428                 spin_lock_init(&sde->senddmactrl_lock);
1429                 spin_lock_init(&sde->flushlist_lock);
1430                 /* insure there is always a zero bit */
1431                 sde->ahg_bits = 0xfffffffe00000000ULL;
1432
1433                 sdma_set_state(sde, sdma_state_s00_hw_down);
1434
1435                 /* set up reference counting */
1436                 kref_init(&sde->state.kref);
1437                 init_completion(&sde->state.comp);
1438
1439                 INIT_LIST_HEAD(&sde->flushlist);
1440                 INIT_LIST_HEAD(&sde->dmawait);
1441
1442                 sde->tail_csr =
1443                         get_kctxt_csr_addr(dd, this_idx, SD(TAIL));
1444
1445                 tasklet_init(&sde->sdma_hw_clean_up_task, sdma_hw_clean_up_task,
1446                              (unsigned long)sde);
1447
1448                 tasklet_init(&sde->sdma_sw_clean_up_task, sdma_sw_clean_up_task,
1449                              (unsigned long)sde);
1450                 INIT_WORK(&sde->err_halt_worker, sdma_err_halt_wait);
1451                 INIT_WORK(&sde->flush_worker, sdma_field_flush);
1452
1453                 sde->progress_check_head = 0;
1454
1455                 timer_setup(&sde->err_progress_check_timer,
1456                             sdma_err_progress_check, 0);
1457
1458                 sde->descq = dma_zalloc_coherent(
1459                         &dd->pcidev->dev,
1460                         descq_cnt * sizeof(u64[2]),
1461                         &sde->descq_phys,
1462                         GFP_KERNEL
1463                 );
1464                 if (!sde->descq)
1465                         goto bail;
1466                 sde->tx_ring =
1467                         kvzalloc_node(sizeof(struct sdma_txreq *) * descq_cnt,
1468                                       GFP_KERNEL, dd->node);
1469                 if (!sde->tx_ring)
1470                         goto bail;
1471         }
1472
1473         dd->sdma_heads_size = L1_CACHE_BYTES * num_engines;
1474         /* Allocate memory for DMA of head registers to memory */
1475         dd->sdma_heads_dma = dma_zalloc_coherent(
1476                 &dd->pcidev->dev,
1477                 dd->sdma_heads_size,
1478                 &dd->sdma_heads_phys,
1479                 GFP_KERNEL
1480         );
1481         if (!dd->sdma_heads_dma) {
1482                 dd_dev_err(dd, "failed to allocate SendDMA head memory\n");
1483                 goto bail;
1484         }
1485
1486         /* Allocate memory for pad */
1487         dd->sdma_pad_dma = dma_zalloc_coherent(
1488                 &dd->pcidev->dev,
1489                 sizeof(u32),
1490                 &dd->sdma_pad_phys,
1491                 GFP_KERNEL
1492         );
1493         if (!dd->sdma_pad_dma) {
1494                 dd_dev_err(dd, "failed to allocate SendDMA pad memory\n");
1495                 goto bail;
1496         }
1497
1498         /* assign each engine to different cacheline and init registers */
1499         curr_head = (void *)dd->sdma_heads_dma;
1500         for (this_idx = 0; this_idx < num_engines; ++this_idx) {
1501                 unsigned long phys_offset;
1502
1503                 sde = &dd->per_sdma[this_idx];
1504
1505                 sde->head_dma = curr_head;
1506                 curr_head += L1_CACHE_BYTES;
1507                 phys_offset = (unsigned long)sde->head_dma -
1508                               (unsigned long)dd->sdma_heads_dma;
1509                 sde->head_phys = dd->sdma_heads_phys + phys_offset;
1510                 init_sdma_regs(sde, per_sdma_credits, idle_cnt);
1511         }
1512         dd->flags |= HFI1_HAS_SEND_DMA;
1513         dd->flags |= idle_cnt ? HFI1_HAS_SDMA_TIMEOUT : 0;
1514         dd->num_sdma = num_engines;
1515         ret = sdma_map_init(dd, port, ppd->vls_operational, NULL);
1516         if (ret < 0)
1517                 goto bail;
1518
1519         tmp_sdma_rht = kzalloc(sizeof(*tmp_sdma_rht), GFP_KERNEL);
1520         if (!tmp_sdma_rht) {
1521                 ret = -ENOMEM;
1522                 goto bail;
1523         }
1524
1525         ret = rhashtable_init(tmp_sdma_rht, &sdma_rht_params);
1526         if (ret < 0)
1527                 goto bail;
1528         dd->sdma_rht = tmp_sdma_rht;
1529
1530         dd_dev_info(dd, "SDMA num_sdma: %u\n", dd->num_sdma);
1531         return 0;
1532
1533 bail:
1534         sdma_clean(dd, num_engines);
1535         return ret;
1536 }
1537
1538 /**
1539  * sdma_all_running() - called when the link goes up
1540  * @dd: hfi1_devdata
1541  *
1542  * This routine moves all engines to the running state.
1543  */
1544 void sdma_all_running(struct hfi1_devdata *dd)
1545 {
1546         struct sdma_engine *sde;
1547         unsigned int i;
1548
1549         /* move all engines to running */
1550         for (i = 0; i < dd->num_sdma; ++i) {
1551                 sde = &dd->per_sdma[i];
1552                 sdma_process_event(sde, sdma_event_e30_go_running);
1553         }
1554 }
1555
1556 /**
1557  * sdma_all_idle() - called when the link goes down
1558  * @dd: hfi1_devdata
1559  *
1560  * This routine moves all engines to the idle state.
1561  */
1562 void sdma_all_idle(struct hfi1_devdata *dd)
1563 {
1564         struct sdma_engine *sde;
1565         unsigned int i;
1566
1567         /* idle all engines */
1568         for (i = 0; i < dd->num_sdma; ++i) {
1569                 sde = &dd->per_sdma[i];
1570                 sdma_process_event(sde, sdma_event_e70_go_idle);
1571         }
1572 }
1573
1574 /**
1575  * sdma_start() - called to kick off state processing for all engines
1576  * @dd: hfi1_devdata
1577  *
1578  * This routine is for kicking off the state processing for all required
1579  * sdma engines.  Interrupts need to be working at this point.
1580  *
1581  */
1582 void sdma_start(struct hfi1_devdata *dd)
1583 {
1584         unsigned i;
1585         struct sdma_engine *sde;
1586
1587         /* kick off the engines state processing */
1588         for (i = 0; i < dd->num_sdma; ++i) {
1589                 sde = &dd->per_sdma[i];
1590                 sdma_process_event(sde, sdma_event_e10_go_hw_start);
1591         }
1592 }
1593
1594 /**
1595  * sdma_exit() - used when module is removed
1596  * @dd: hfi1_devdata
1597  */
1598 void sdma_exit(struct hfi1_devdata *dd)
1599 {
1600         unsigned this_idx;
1601         struct sdma_engine *sde;
1602
1603         for (this_idx = 0; dd->per_sdma && this_idx < dd->num_sdma;
1604                         ++this_idx) {
1605                 sde = &dd->per_sdma[this_idx];
1606                 if (!list_empty(&sde->dmawait))
1607                         dd_dev_err(dd, "sde %u: dmawait list not empty!\n",
1608                                    sde->this_idx);
1609                 sdma_process_event(sde, sdma_event_e00_go_hw_down);
1610
1611                 del_timer_sync(&sde->err_progress_check_timer);
1612
1613                 /*
1614                  * This waits for the state machine to exit so it is not
1615                  * necessary to kill the sdma_sw_clean_up_task to make sure
1616                  * it is not running.
1617                  */
1618                 sdma_finalput(&sde->state);
1619         }
1620         sdma_clean(dd, dd->num_sdma);
1621 }
1622
1623 /*
1624  * unmap the indicated descriptor
1625  */
1626 static inline void sdma_unmap_desc(
1627         struct hfi1_devdata *dd,
1628         struct sdma_desc *descp)
1629 {
1630         switch (sdma_mapping_type(descp)) {
1631         case SDMA_MAP_SINGLE:
1632                 dma_unmap_single(
1633                         &dd->pcidev->dev,
1634                         sdma_mapping_addr(descp),
1635                         sdma_mapping_len(descp),
1636                         DMA_TO_DEVICE);
1637                 break;
1638         case SDMA_MAP_PAGE:
1639                 dma_unmap_page(
1640                         &dd->pcidev->dev,
1641                         sdma_mapping_addr(descp),
1642                         sdma_mapping_len(descp),
1643                         DMA_TO_DEVICE);
1644                 break;
1645         }
1646 }
1647
1648 /*
1649  * return the mode as indicated by the first
1650  * descriptor in the tx.
1651  */
1652 static inline u8 ahg_mode(struct sdma_txreq *tx)
1653 {
1654         return (tx->descp[0].qw[1] & SDMA_DESC1_HEADER_MODE_SMASK)
1655                 >> SDMA_DESC1_HEADER_MODE_SHIFT;
1656 }
1657
1658 /**
1659  * __sdma_txclean() - clean tx of mappings, descp *kmalloc's
1660  * @dd: hfi1_devdata for unmapping
1661  * @tx: tx request to clean
1662  *
1663  * This is used in the progress routine to clean the tx or
1664  * by the ULP to toss an in-process tx build.
1665  *
1666  * The code can be called multiple times without issue.
1667  *
1668  */
1669 void __sdma_txclean(
1670         struct hfi1_devdata *dd,
1671         struct sdma_txreq *tx)
1672 {
1673         u16 i;
1674
1675         if (tx->num_desc) {
1676                 u8 skip = 0, mode = ahg_mode(tx);
1677
1678                 /* unmap first */
1679                 sdma_unmap_desc(dd, &tx->descp[0]);
1680                 /* determine number of AHG descriptors to skip */
1681                 if (mode > SDMA_AHG_APPLY_UPDATE1)
1682                         skip = mode >> 1;
1683                 for (i = 1 + skip; i < tx->num_desc; i++)
1684                         sdma_unmap_desc(dd, &tx->descp[i]);
1685                 tx->num_desc = 0;
1686         }
1687         kfree(tx->coalesce_buf);
1688         tx->coalesce_buf = NULL;
1689         /* kmalloc'ed descp */
1690         if (unlikely(tx->desc_limit > ARRAY_SIZE(tx->descs))) {
1691                 tx->desc_limit = ARRAY_SIZE(tx->descs);
1692                 kfree(tx->descp);
1693         }
1694 }
1695
1696 static inline u16 sdma_gethead(struct sdma_engine *sde)
1697 {
1698         struct hfi1_devdata *dd = sde->dd;
1699         int use_dmahead;
1700         u16 hwhead;
1701
1702 #ifdef CONFIG_SDMA_VERBOSITY
1703         dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
1704                    sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
1705 #endif
1706
1707 retry:
1708         use_dmahead = HFI1_CAP_IS_KSET(USE_SDMA_HEAD) && __sdma_running(sde) &&
1709                                         (dd->flags & HFI1_HAS_SDMA_TIMEOUT);
1710         hwhead = use_dmahead ?
1711                 (u16)le64_to_cpu(*sde->head_dma) :
1712                 (u16)read_sde_csr(sde, SD(HEAD));
1713
1714         if (unlikely(HFI1_CAP_IS_KSET(SDMA_HEAD_CHECK))) {
1715                 u16 cnt;
1716                 u16 swtail;
1717                 u16 swhead;
1718                 int sane;
1719
1720                 swhead = sde->descq_head & sde->sdma_mask;
1721                 /* this code is really bad for cache line trading */
1722                 swtail = READ_ONCE(sde->descq_tail) & sde->sdma_mask;
1723                 cnt = sde->descq_cnt;
1724
1725                 if (swhead < swtail)
1726                         /* not wrapped */
1727                         sane = (hwhead >= swhead) & (hwhead <= swtail);
1728                 else if (swhead > swtail)
1729                         /* wrapped around */
1730                         sane = ((hwhead >= swhead) && (hwhead < cnt)) ||
1731                                 (hwhead <= swtail);
1732                 else
1733                         /* empty */
1734                         sane = (hwhead == swhead);
1735
1736                 if (unlikely(!sane)) {
1737                         dd_dev_err(dd, "SDMA(%u) bad head (%s) hwhd=%hu swhd=%hu swtl=%hu cnt=%hu\n",
1738                                    sde->this_idx,
1739                                    use_dmahead ? "dma" : "kreg",
1740                                    hwhead, swhead, swtail, cnt);
1741                         if (use_dmahead) {
1742                                 /* try one more time, using csr */
1743                                 use_dmahead = 0;
1744                                 goto retry;
1745                         }
1746                         /* proceed as if no progress */
1747                         hwhead = swhead;
1748                 }
1749         }
1750         return hwhead;
1751 }
1752
1753 /*
1754  * This is called when there are send DMA descriptors that might be
1755  * available.
1756  *
1757  * This is called with head_lock held.
1758  */
1759 static void sdma_desc_avail(struct sdma_engine *sde, uint avail)
1760 {
1761         struct iowait *wait, *nw;
1762         struct iowait *waits[SDMA_WAIT_BATCH_SIZE];
1763         uint i, n = 0, seq, max_idx = 0;
1764         struct sdma_txreq *stx;
1765         struct hfi1_ibdev *dev = &sde->dd->verbs_dev;
1766         u8 max_starved_cnt = 0;
1767
1768 #ifdef CONFIG_SDMA_VERBOSITY
1769         dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", sde->this_idx,
1770                    slashstrip(__FILE__), __LINE__, __func__);
1771         dd_dev_err(sde->dd, "avail: %u\n", avail);
1772 #endif
1773
1774         do {
1775                 seq = read_seqbegin(&dev->iowait_lock);
1776                 if (!list_empty(&sde->dmawait)) {
1777                         /* at least one item */
1778                         write_seqlock(&dev->iowait_lock);
1779                         /* Harvest waiters wanting DMA descriptors */
1780                         list_for_each_entry_safe(
1781                                         wait,
1782                                         nw,
1783                                         &sde->dmawait,
1784                                         list) {
1785                                 u16 num_desc = 0;
1786
1787                                 if (!wait->wakeup)
1788                                         continue;
1789                                 if (n == ARRAY_SIZE(waits))
1790                                         break;
1791                                 if (!list_empty(&wait->tx_head)) {
1792                                         stx = list_first_entry(
1793                                                 &wait->tx_head,
1794                                                 struct sdma_txreq,
1795                                                 list);
1796                                         num_desc = stx->num_desc;
1797                                 }
1798                                 if (num_desc > avail)
1799                                         break;
1800                                 avail -= num_desc;
1801                                 /* Find the most starved wait memeber */
1802                                 iowait_starve_find_max(wait, &max_starved_cnt,
1803                                                        n, &max_idx);
1804                                 list_del_init(&wait->list);
1805                                 waits[n++] = wait;
1806                         }
1807                         write_sequnlock(&dev->iowait_lock);
1808                         break;
1809                 }
1810         } while (read_seqretry(&dev->iowait_lock, seq));
1811
1812         /* Schedule the most starved one first */
1813         if (n)
1814                 waits[max_idx]->wakeup(waits[max_idx], SDMA_AVAIL_REASON);
1815
1816         for (i = 0; i < n; i++)
1817                 if (i != max_idx)
1818                         waits[i]->wakeup(waits[i], SDMA_AVAIL_REASON);
1819 }
1820
1821 /* head_lock must be held */
1822 static void sdma_make_progress(struct sdma_engine *sde, u64 status)
1823 {
1824         struct sdma_txreq *txp = NULL;
1825         int progress = 0;
1826         u16 hwhead, swhead;
1827         int idle_check_done = 0;
1828
1829         hwhead = sdma_gethead(sde);
1830
1831         /* The reason for some of the complexity of this code is that
1832          * not all descriptors have corresponding txps.  So, we have to
1833          * be able to skip over descs until we wander into the range of
1834          * the next txp on the list.
1835          */
1836
1837 retry:
1838         txp = get_txhead(sde);
1839         swhead = sde->descq_head & sde->sdma_mask;
1840         trace_hfi1_sdma_progress(sde, hwhead, swhead, txp);
1841         while (swhead != hwhead) {
1842                 /* advance head, wrap if needed */
1843                 swhead = ++sde->descq_head & sde->sdma_mask;
1844
1845                 /* if now past this txp's descs, do the callback */
1846                 if (txp && txp->next_descq_idx == swhead) {
1847                         /* remove from list */
1848                         sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL;
1849                         complete_tx(sde, txp, SDMA_TXREQ_S_OK);
1850                         /* see if there is another txp */
1851                         txp = get_txhead(sde);
1852                 }
1853                 trace_hfi1_sdma_progress(sde, hwhead, swhead, txp);
1854                 progress++;
1855         }
1856
1857         /*
1858          * The SDMA idle interrupt is not guaranteed to be ordered with respect
1859          * to updates to the the dma_head location in host memory. The head
1860          * value read might not be fully up to date. If there are pending
1861          * descriptors and the SDMA idle interrupt fired then read from the
1862          * CSR SDMA head instead to get the latest value from the hardware.
1863          * The hardware SDMA head should be read at most once in this invocation
1864          * of sdma_make_progress(..) which is ensured by idle_check_done flag
1865          */
1866         if ((status & sde->idle_mask) && !idle_check_done) {
1867                 u16 swtail;
1868
1869                 swtail = READ_ONCE(sde->descq_tail) & sde->sdma_mask;
1870                 if (swtail != hwhead) {
1871                         hwhead = (u16)read_sde_csr(sde, SD(HEAD));
1872                         idle_check_done = 1;
1873                         goto retry;
1874                 }
1875         }
1876
1877         sde->last_status = status;
1878         if (progress)
1879                 sdma_desc_avail(sde, sdma_descq_freecnt(sde));
1880 }
1881
1882 /*
1883  * sdma_engine_interrupt() - interrupt handler for engine
1884  * @sde: sdma engine
1885  * @status: sdma interrupt reason
1886  *
1887  * Status is a mask of the 3 possible interrupts for this engine.  It will
1888  * contain bits _only_ for this SDMA engine.  It will contain at least one
1889  * bit, it may contain more.
1890  */
1891 void sdma_engine_interrupt(struct sdma_engine *sde, u64 status)
1892 {
1893         trace_hfi1_sdma_engine_interrupt(sde, status);
1894         write_seqlock(&sde->head_lock);
1895         sdma_set_desc_cnt(sde, sdma_desct_intr);
1896         if (status & sde->idle_mask)
1897                 sde->idle_int_cnt++;
1898         else if (status & sde->progress_mask)
1899                 sde->progress_int_cnt++;
1900         else if (status & sde->int_mask)
1901                 sde->sdma_int_cnt++;
1902         sdma_make_progress(sde, status);
1903         write_sequnlock(&sde->head_lock);
1904 }
1905
1906 /**
1907  * sdma_engine_error() - error handler for engine
1908  * @sde: sdma engine
1909  * @status: sdma interrupt reason
1910  */
1911 void sdma_engine_error(struct sdma_engine *sde, u64 status)
1912 {
1913         unsigned long flags;
1914
1915 #ifdef CONFIG_SDMA_VERBOSITY
1916         dd_dev_err(sde->dd, "CONFIG SDMA(%u) error status 0x%llx state %s\n",
1917                    sde->this_idx,
1918                    (unsigned long long)status,
1919                    sdma_state_names[sde->state.current_state]);
1920 #endif
1921         spin_lock_irqsave(&sde->tail_lock, flags);
1922         write_seqlock(&sde->head_lock);
1923         if (status & ALL_SDMA_ENG_HALT_ERRS)
1924                 __sdma_process_event(sde, sdma_event_e60_hw_halted);
1925         if (status & ~SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK)) {
1926                 dd_dev_err(sde->dd,
1927                            "SDMA (%u) engine error: 0x%llx state %s\n",
1928                            sde->this_idx,
1929                            (unsigned long long)status,
1930                            sdma_state_names[sde->state.current_state]);
1931                 dump_sdma_state(sde);
1932         }
1933         write_sequnlock(&sde->head_lock);
1934         spin_unlock_irqrestore(&sde->tail_lock, flags);
1935 }
1936
1937 static void sdma_sendctrl(struct sdma_engine *sde, unsigned op)
1938 {
1939         u64 set_senddmactrl = 0;
1940         u64 clr_senddmactrl = 0;
1941         unsigned long flags;
1942
1943 #ifdef CONFIG_SDMA_VERBOSITY
1944         dd_dev_err(sde->dd, "CONFIG SDMA(%u) senddmactrl E=%d I=%d H=%d C=%d\n",
1945                    sde->this_idx,
1946                    (op & SDMA_SENDCTRL_OP_ENABLE) ? 1 : 0,
1947                    (op & SDMA_SENDCTRL_OP_INTENABLE) ? 1 : 0,
1948                    (op & SDMA_SENDCTRL_OP_HALT) ? 1 : 0,
1949                    (op & SDMA_SENDCTRL_OP_CLEANUP) ? 1 : 0);
1950 #endif
1951
1952         if (op & SDMA_SENDCTRL_OP_ENABLE)
1953                 set_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK);
1954         else
1955                 clr_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK);
1956
1957         if (op & SDMA_SENDCTRL_OP_INTENABLE)
1958                 set_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK);
1959         else
1960                 clr_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK);
1961
1962         if (op & SDMA_SENDCTRL_OP_HALT)
1963                 set_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK);
1964         else
1965                 clr_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK);
1966
1967         spin_lock_irqsave(&sde->senddmactrl_lock, flags);
1968
1969         sde->p_senddmactrl |= set_senddmactrl;
1970         sde->p_senddmactrl &= ~clr_senddmactrl;
1971
1972         if (op & SDMA_SENDCTRL_OP_CLEANUP)
1973                 write_sde_csr(sde, SD(CTRL),
1974                               sde->p_senddmactrl |
1975                               SD(CTRL_SDMA_CLEANUP_SMASK));
1976         else
1977                 write_sde_csr(sde, SD(CTRL), sde->p_senddmactrl);
1978
1979         spin_unlock_irqrestore(&sde->senddmactrl_lock, flags);
1980
1981 #ifdef CONFIG_SDMA_VERBOSITY
1982         sdma_dumpstate(sde);
1983 #endif
1984 }
1985
1986 static void sdma_setlengen(struct sdma_engine *sde)
1987 {
1988 #ifdef CONFIG_SDMA_VERBOSITY
1989         dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
1990                    sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
1991 #endif
1992
1993         /*
1994          * Set SendDmaLenGen and clear-then-set the MSB of the generation
1995          * count to enable generation checking and load the internal
1996          * generation counter.
1997          */
1998         write_sde_csr(sde, SD(LEN_GEN),
1999                       (sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT));
2000         write_sde_csr(sde, SD(LEN_GEN),
2001                       ((sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT)) |
2002                       (4ULL << SD(LEN_GEN_GENERATION_SHIFT)));
2003 }
2004
2005 static inline void sdma_update_tail(struct sdma_engine *sde, u16 tail)
2006 {
2007         /* Commit writes to memory and advance the tail on the chip */
2008         smp_wmb(); /* see get_txhead() */
2009         writeq(tail, sde->tail_csr);
2010 }
2011
2012 /*
2013  * This is called when changing to state s10_hw_start_up_halt_wait as
2014  * a result of send buffer errors or send DMA descriptor errors.
2015  */
2016 static void sdma_hw_start_up(struct sdma_engine *sde)
2017 {
2018         u64 reg;
2019
2020 #ifdef CONFIG_SDMA_VERBOSITY
2021         dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
2022                    sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
2023 #endif
2024
2025         sdma_setlengen(sde);
2026         sdma_update_tail(sde, 0); /* Set SendDmaTail */
2027         *sde->head_dma = 0;
2028
2029         reg = SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_MASK) <<
2030               SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SHIFT);
2031         write_sde_csr(sde, SD(ENG_ERR_CLEAR), reg);
2032 }
2033
2034 /*
2035  * set_sdma_integrity
2036  *
2037  * Set the SEND_DMA_CHECK_ENABLE register for send DMA engine 'sde'.
2038  */
2039 static void set_sdma_integrity(struct sdma_engine *sde)
2040 {
2041         struct hfi1_devdata *dd = sde->dd;
2042
2043         write_sde_csr(sde, SD(CHECK_ENABLE),
2044                       hfi1_pkt_base_sdma_integrity(dd));
2045 }
2046
2047 static void init_sdma_regs(
2048         struct sdma_engine *sde,
2049         u32 credits,
2050         uint idle_cnt)
2051 {
2052         u8 opval, opmask;
2053 #ifdef CONFIG_SDMA_VERBOSITY
2054         struct hfi1_devdata *dd = sde->dd;
2055
2056         dd_dev_err(dd, "CONFIG SDMA(%u) %s:%d %s()\n",
2057                    sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
2058 #endif
2059
2060         write_sde_csr(sde, SD(BASE_ADDR), sde->descq_phys);
2061         sdma_setlengen(sde);
2062         sdma_update_tail(sde, 0); /* Set SendDmaTail */
2063         write_sde_csr(sde, SD(RELOAD_CNT), idle_cnt);
2064         write_sde_csr(sde, SD(DESC_CNT), 0);
2065         write_sde_csr(sde, SD(HEAD_ADDR), sde->head_phys);
2066         write_sde_csr(sde, SD(MEMORY),
2067                       ((u64)credits << SD(MEMORY_SDMA_MEMORY_CNT_SHIFT)) |
2068                       ((u64)(credits * sde->this_idx) <<
2069                        SD(MEMORY_SDMA_MEMORY_INDEX_SHIFT)));
2070         write_sde_csr(sde, SD(ENG_ERR_MASK), ~0ull);
2071         set_sdma_integrity(sde);
2072         opmask = OPCODE_CHECK_MASK_DISABLED;
2073         opval = OPCODE_CHECK_VAL_DISABLED;
2074         write_sde_csr(sde, SD(CHECK_OPCODE),
2075                       (opmask << SEND_CTXT_CHECK_OPCODE_MASK_SHIFT) |
2076                       (opval << SEND_CTXT_CHECK_OPCODE_VALUE_SHIFT));
2077 }
2078
2079 #ifdef CONFIG_SDMA_VERBOSITY
2080
2081 #define sdma_dumpstate_helper0(reg) do { \
2082                 csr = read_csr(sde->dd, reg); \
2083                 dd_dev_err(sde->dd, "%36s     0x%016llx\n", #reg, csr); \
2084         } while (0)
2085
2086 #define sdma_dumpstate_helper(reg) do { \
2087                 csr = read_sde_csr(sde, reg); \
2088                 dd_dev_err(sde->dd, "%36s[%02u] 0x%016llx\n", \
2089                         #reg, sde->this_idx, csr); \
2090         } while (0)
2091
2092 #define sdma_dumpstate_helper2(reg) do { \
2093                 csr = read_csr(sde->dd, reg + (8 * i)); \
2094                 dd_dev_err(sde->dd, "%33s_%02u     0x%016llx\n", \
2095                                 #reg, i, csr); \
2096         } while (0)
2097
2098 void sdma_dumpstate(struct sdma_engine *sde)
2099 {
2100         u64 csr;
2101         unsigned i;
2102
2103         sdma_dumpstate_helper(SD(CTRL));
2104         sdma_dumpstate_helper(SD(STATUS));
2105         sdma_dumpstate_helper0(SD(ERR_STATUS));
2106         sdma_dumpstate_helper0(SD(ERR_MASK));
2107         sdma_dumpstate_helper(SD(ENG_ERR_STATUS));
2108         sdma_dumpstate_helper(SD(ENG_ERR_MASK));
2109
2110         for (i = 0; i < CCE_NUM_INT_CSRS; ++i) {
2111                 sdma_dumpstate_helper2(CCE_INT_STATUS);
2112                 sdma_dumpstate_helper2(CCE_INT_MASK);
2113                 sdma_dumpstate_helper2(CCE_INT_BLOCKED);
2114         }
2115
2116         sdma_dumpstate_helper(SD(TAIL));
2117         sdma_dumpstate_helper(SD(HEAD));
2118         sdma_dumpstate_helper(SD(PRIORITY_THLD));
2119         sdma_dumpstate_helper(SD(IDLE_CNT));
2120         sdma_dumpstate_helper(SD(RELOAD_CNT));
2121         sdma_dumpstate_helper(SD(DESC_CNT));
2122         sdma_dumpstate_helper(SD(DESC_FETCHED_CNT));
2123         sdma_dumpstate_helper(SD(MEMORY));
2124         sdma_dumpstate_helper0(SD(ENGINES));
2125         sdma_dumpstate_helper0(SD(MEM_SIZE));
2126         /* sdma_dumpstate_helper(SEND_EGRESS_SEND_DMA_STATUS);  */
2127         sdma_dumpstate_helper(SD(BASE_ADDR));
2128         sdma_dumpstate_helper(SD(LEN_GEN));
2129         sdma_dumpstate_helper(SD(HEAD_ADDR));
2130         sdma_dumpstate_helper(SD(CHECK_ENABLE));
2131         sdma_dumpstate_helper(SD(CHECK_VL));
2132         sdma_dumpstate_helper(SD(CHECK_JOB_KEY));
2133         sdma_dumpstate_helper(SD(CHECK_PARTITION_KEY));
2134         sdma_dumpstate_helper(SD(CHECK_SLID));
2135         sdma_dumpstate_helper(SD(CHECK_OPCODE));
2136 }
2137 #endif
2138
2139 static void dump_sdma_state(struct sdma_engine *sde)
2140 {
2141         struct hw_sdma_desc *descqp;
2142         u64 desc[2];
2143         u64 addr;
2144         u8 gen;
2145         u16 len;
2146         u16 head, tail, cnt;
2147
2148         head = sde->descq_head & sde->sdma_mask;
2149         tail = sde->descq_tail & sde->sdma_mask;
2150         cnt = sdma_descq_freecnt(sde);
2151
2152         dd_dev_err(sde->dd,
2153                    "SDMA (%u) descq_head: %u descq_tail: %u freecnt: %u FLE %d\n",
2154                    sde->this_idx, head, tail, cnt,
2155                    !list_empty(&sde->flushlist));
2156
2157         /* print info for each entry in the descriptor queue */
2158         while (head != tail) {
2159                 char flags[6] = { 'x', 'x', 'x', 'x', 0 };
2160
2161                 descqp = &sde->descq[head];
2162                 desc[0] = le64_to_cpu(descqp->qw[0]);
2163                 desc[1] = le64_to_cpu(descqp->qw[1]);
2164                 flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-';
2165                 flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ?
2166                                 'H' : '-';
2167                 flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-';
2168                 flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-';
2169                 addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT)
2170                         & SDMA_DESC0_PHY_ADDR_MASK;
2171                 gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT)
2172                         & SDMA_DESC1_GENERATION_MASK;
2173                 len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT)
2174                         & SDMA_DESC0_BYTE_COUNT_MASK;
2175                 dd_dev_err(sde->dd,
2176                            "SDMA sdmadesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n",
2177                            head, flags, addr, gen, len);
2178                 dd_dev_err(sde->dd,
2179                            "\tdesc0:0x%016llx desc1 0x%016llx\n",
2180                            desc[0], desc[1]);
2181                 if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG)
2182                         dd_dev_err(sde->dd,
2183                                    "\taidx: %u amode: %u alen: %u\n",
2184                                    (u8)((desc[1] &
2185                                          SDMA_DESC1_HEADER_INDEX_SMASK) >>
2186                                         SDMA_DESC1_HEADER_INDEX_SHIFT),
2187                                    (u8)((desc[1] &
2188                                          SDMA_DESC1_HEADER_MODE_SMASK) >>
2189                                         SDMA_DESC1_HEADER_MODE_SHIFT),
2190                                    (u8)((desc[1] &
2191                                          SDMA_DESC1_HEADER_DWS_SMASK) >>
2192                                         SDMA_DESC1_HEADER_DWS_SHIFT));
2193                 head++;
2194                 head &= sde->sdma_mask;
2195         }
2196 }
2197
2198 #define SDE_FMT \
2199         "SDE %u CPU %d STE %s C 0x%llx S 0x%016llx E 0x%llx T(HW) 0x%llx T(SW) 0x%x H(HW) 0x%llx H(SW) 0x%x H(D) 0x%llx DM 0x%llx GL 0x%llx R 0x%llx LIS 0x%llx AHGI 0x%llx TXT %u TXH %u DT %u DH %u FLNE %d DQF %u SLC 0x%llx\n"
2200 /**
2201  * sdma_seqfile_dump_sde() - debugfs dump of sde
2202  * @s: seq file
2203  * @sde: send dma engine to dump
2204  *
2205  * This routine dumps the sde to the indicated seq file.
2206  */
2207 void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *sde)
2208 {
2209         u16 head, tail;
2210         struct hw_sdma_desc *descqp;
2211         u64 desc[2];
2212         u64 addr;
2213         u8 gen;
2214         u16 len;
2215
2216         head = sde->descq_head & sde->sdma_mask;
2217         tail = READ_ONCE(sde->descq_tail) & sde->sdma_mask;
2218         seq_printf(s, SDE_FMT, sde->this_idx,
2219                    sde->cpu,
2220                    sdma_state_name(sde->state.current_state),
2221                    (unsigned long long)read_sde_csr(sde, SD(CTRL)),
2222                    (unsigned long long)read_sde_csr(sde, SD(STATUS)),
2223                    (unsigned long long)read_sde_csr(sde, SD(ENG_ERR_STATUS)),
2224                    (unsigned long long)read_sde_csr(sde, SD(TAIL)), tail,
2225                    (unsigned long long)read_sde_csr(sde, SD(HEAD)), head,
2226                    (unsigned long long)le64_to_cpu(*sde->head_dma),
2227                    (unsigned long long)read_sde_csr(sde, SD(MEMORY)),
2228                    (unsigned long long)read_sde_csr(sde, SD(LEN_GEN)),
2229                    (unsigned long long)read_sde_csr(sde, SD(RELOAD_CNT)),
2230                    (unsigned long long)sde->last_status,
2231                    (unsigned long long)sde->ahg_bits,
2232                    sde->tx_tail,
2233                    sde->tx_head,
2234                    sde->descq_tail,
2235                    sde->descq_head,
2236                    !list_empty(&sde->flushlist),
2237                    sde->descq_full_count,
2238                    (unsigned long long)read_sde_csr(sde, SEND_DMA_CHECK_SLID));
2239
2240         /* print info for each entry in the descriptor queue */
2241         while (head != tail) {
2242                 char flags[6] = { 'x', 'x', 'x', 'x', 0 };
2243
2244                 descqp = &sde->descq[head];
2245                 desc[0] = le64_to_cpu(descqp->qw[0]);
2246                 desc[1] = le64_to_cpu(descqp->qw[1]);
2247                 flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-';
2248                 flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ?
2249                                 'H' : '-';
2250                 flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-';
2251                 flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-';
2252                 addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT)
2253                         & SDMA_DESC0_PHY_ADDR_MASK;
2254                 gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT)
2255                         & SDMA_DESC1_GENERATION_MASK;
2256                 len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT)
2257                         & SDMA_DESC0_BYTE_COUNT_MASK;
2258                 seq_printf(s,
2259                            "\tdesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n",
2260                            head, flags, addr, gen, len);
2261                 if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG)
2262                         seq_printf(s, "\t\tahgidx: %u ahgmode: %u\n",
2263                                    (u8)((desc[1] &
2264                                          SDMA_DESC1_HEADER_INDEX_SMASK) >>
2265                                         SDMA_DESC1_HEADER_INDEX_SHIFT),
2266                                    (u8)((desc[1] &
2267                                          SDMA_DESC1_HEADER_MODE_SMASK) >>
2268                                         SDMA_DESC1_HEADER_MODE_SHIFT));
2269                 head = (head + 1) & sde->sdma_mask;
2270         }
2271 }
2272
2273 /*
2274  * add the generation number into
2275  * the qw1 and return
2276  */
2277 static inline u64 add_gen(struct sdma_engine *sde, u64 qw1)
2278 {
2279         u8 generation = (sde->descq_tail >> sde->sdma_shift) & 3;
2280
2281         qw1 &= ~SDMA_DESC1_GENERATION_SMASK;
2282         qw1 |= ((u64)generation & SDMA_DESC1_GENERATION_MASK)
2283                         << SDMA_DESC1_GENERATION_SHIFT;
2284         return qw1;
2285 }
2286
2287 /*
2288  * This routine submits the indicated tx
2289  *
2290  * Space has already been guaranteed and
2291  * tail side of ring is locked.
2292  *
2293  * The hardware tail update is done
2294  * in the caller and that is facilitated
2295  * by returning the new tail.
2296  *
2297  * There is special case logic for ahg
2298  * to not add the generation number for
2299  * up to 2 descriptors that follow the
2300  * first descriptor.
2301  *
2302  */
2303 static inline u16 submit_tx(struct sdma_engine *sde, struct sdma_txreq *tx)
2304 {
2305         int i;
2306         u16 tail;
2307         struct sdma_desc *descp = tx->descp;
2308         u8 skip = 0, mode = ahg_mode(tx);
2309
2310         tail = sde->descq_tail & sde->sdma_mask;
2311         sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]);
2312         sde->descq[tail].qw[1] = cpu_to_le64(add_gen(sde, descp->qw[1]));
2313         trace_hfi1_sdma_descriptor(sde, descp->qw[0], descp->qw[1],
2314                                    tail, &sde->descq[tail]);
2315         tail = ++sde->descq_tail & sde->sdma_mask;
2316         descp++;
2317         if (mode > SDMA_AHG_APPLY_UPDATE1)
2318                 skip = mode >> 1;
2319         for (i = 1; i < tx->num_desc; i++, descp++) {
2320                 u64 qw1;
2321
2322                 sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]);
2323                 if (skip) {
2324                         /* edits don't have generation */
2325                         qw1 = descp->qw[1];
2326                         skip--;
2327                 } else {
2328                         /* replace generation with real one for non-edits */
2329                         qw1 = add_gen(sde, descp->qw[1]);
2330                 }
2331                 sde->descq[tail].qw[1] = cpu_to_le64(qw1);
2332                 trace_hfi1_sdma_descriptor(sde, descp->qw[0], qw1,
2333                                            tail, &sde->descq[tail]);
2334                 tail = ++sde->descq_tail & sde->sdma_mask;
2335         }
2336         tx->next_descq_idx = tail;
2337 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2338         tx->sn = sde->tail_sn++;
2339         trace_hfi1_sdma_in_sn(sde, tx->sn);
2340         WARN_ON_ONCE(sde->tx_ring[sde->tx_tail & sde->sdma_mask]);
2341 #endif
2342         sde->tx_ring[sde->tx_tail++ & sde->sdma_mask] = tx;
2343         sde->desc_avail -= tx->num_desc;
2344         return tail;
2345 }
2346
2347 /*
2348  * Check for progress
2349  */
2350 static int sdma_check_progress(
2351         struct sdma_engine *sde,
2352         struct iowait *wait,
2353         struct sdma_txreq *tx,
2354         bool pkts_sent)
2355 {
2356         int ret;
2357
2358         sde->desc_avail = sdma_descq_freecnt(sde);
2359         if (tx->num_desc <= sde->desc_avail)
2360                 return -EAGAIN;
2361         /* pulse the head_lock */
2362         if (wait && wait->sleep) {
2363                 unsigned seq;
2364
2365                 seq = raw_seqcount_begin(
2366                         (const seqcount_t *)&sde->head_lock.seqcount);
2367                 ret = wait->sleep(sde, wait, tx, seq, pkts_sent);
2368                 if (ret == -EAGAIN)
2369                         sde->desc_avail = sdma_descq_freecnt(sde);
2370         } else {
2371                 ret = -EBUSY;
2372         }
2373         return ret;
2374 }
2375
2376 /**
2377  * sdma_send_txreq() - submit a tx req to ring
2378  * @sde: sdma engine to use
2379  * @wait: wait structure to use when full (may be NULL)
2380  * @tx: sdma_txreq to submit
2381  * @pkts_sent: has any packet been sent yet?
2382  *
2383  * The call submits the tx into the ring.  If a iowait structure is non-NULL
2384  * the packet will be queued to the list in wait.
2385  *
2386  * Return:
2387  * 0 - Success, -EINVAL - sdma_txreq incomplete, -EBUSY - no space in
2388  * ring (wait == NULL)
2389  * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state
2390  */
2391 int sdma_send_txreq(struct sdma_engine *sde,
2392                     struct iowait *wait,
2393                     struct sdma_txreq *tx,
2394                     bool pkts_sent)
2395 {
2396         int ret = 0;
2397         u16 tail;
2398         unsigned long flags;
2399
2400         /* user should have supplied entire packet */
2401         if (unlikely(tx->tlen))
2402                 return -EINVAL;
2403         tx->wait = wait;
2404         spin_lock_irqsave(&sde->tail_lock, flags);
2405 retry:
2406         if (unlikely(!__sdma_running(sde)))
2407                 goto unlock_noconn;
2408         if (unlikely(tx->num_desc > sde->desc_avail))
2409                 goto nodesc;
2410         tail = submit_tx(sde, tx);
2411         if (wait)
2412                 iowait_sdma_inc(wait);
2413         sdma_update_tail(sde, tail);
2414 unlock:
2415         spin_unlock_irqrestore(&sde->tail_lock, flags);
2416         return ret;
2417 unlock_noconn:
2418         if (wait)
2419                 iowait_sdma_inc(wait);
2420         tx->next_descq_idx = 0;
2421 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2422         tx->sn = sde->tail_sn++;
2423         trace_hfi1_sdma_in_sn(sde, tx->sn);
2424 #endif
2425         spin_lock(&sde->flushlist_lock);
2426         list_add_tail(&tx->list, &sde->flushlist);
2427         spin_unlock(&sde->flushlist_lock);
2428         if (wait) {
2429                 wait->tx_count++;
2430                 wait->count += tx->num_desc;
2431         }
2432         schedule_work(&sde->flush_worker);
2433         ret = -ECOMM;
2434         goto unlock;
2435 nodesc:
2436         ret = sdma_check_progress(sde, wait, tx, pkts_sent);
2437         if (ret == -EAGAIN) {
2438                 ret = 0;
2439                 goto retry;
2440         }
2441         sde->descq_full_count++;
2442         goto unlock;
2443 }
2444
2445 /**
2446  * sdma_send_txlist() - submit a list of tx req to ring
2447  * @sde: sdma engine to use
2448  * @wait: wait structure to use when full (may be NULL)
2449  * @tx_list: list of sdma_txreqs to submit
2450  * @count: pointer to a u32 which, after return will contain the total number of
2451  *         sdma_txreqs removed from the tx_list. This will include sdma_txreqs
2452  *         whose SDMA descriptors are submitted to the ring and the sdma_txreqs
2453  *         which are added to SDMA engine flush list if the SDMA engine state is
2454  *         not running.
2455  *
2456  * The call submits the list into the ring.
2457  *
2458  * If the iowait structure is non-NULL and not equal to the iowait list
2459  * the unprocessed part of the list  will be appended to the list in wait.
2460  *
2461  * In all cases, the tx_list will be updated so the head of the tx_list is
2462  * the list of descriptors that have yet to be transmitted.
2463  *
2464  * The intent of this call is to provide a more efficient
2465  * way of submitting multiple packets to SDMA while holding the tail
2466  * side locking.
2467  *
2468  * Return:
2469  * 0 - Success,
2470  * -EINVAL - sdma_txreq incomplete, -EBUSY - no space in ring (wait == NULL)
2471  * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state
2472  */
2473 int sdma_send_txlist(struct sdma_engine *sde, struct iowait *wait,
2474                      struct list_head *tx_list, u32 *count_out)
2475 {
2476         struct sdma_txreq *tx, *tx_next;
2477         int ret = 0;
2478         unsigned long flags;
2479         u16 tail = INVALID_TAIL;
2480         u32 submit_count = 0, flush_count = 0, total_count;
2481
2482         spin_lock_irqsave(&sde->tail_lock, flags);
2483 retry:
2484         list_for_each_entry_safe(tx, tx_next, tx_list, list) {
2485                 tx->wait = wait;
2486                 if (unlikely(!__sdma_running(sde)))
2487                         goto unlock_noconn;
2488                 if (unlikely(tx->num_desc > sde->desc_avail))
2489                         goto nodesc;
2490                 if (unlikely(tx->tlen)) {
2491                         ret = -EINVAL;
2492                         goto update_tail;
2493                 }
2494                 list_del_init(&tx->list);
2495                 tail = submit_tx(sde, tx);
2496                 submit_count++;
2497                 if (tail != INVALID_TAIL &&
2498                     (submit_count & SDMA_TAIL_UPDATE_THRESH) == 0) {
2499                         sdma_update_tail(sde, tail);
2500                         tail = INVALID_TAIL;
2501                 }
2502         }
2503 update_tail:
2504         total_count = submit_count + flush_count;
2505         if (wait) {
2506                 iowait_sdma_add(wait, total_count);
2507                 iowait_starve_clear(submit_count > 0, wait);
2508         }
2509         if (tail != INVALID_TAIL)
2510                 sdma_update_tail(sde, tail);
2511         spin_unlock_irqrestore(&sde->tail_lock, flags);
2512         *count_out = total_count;
2513         return ret;
2514 unlock_noconn:
2515         spin_lock(&sde->flushlist_lock);
2516         list_for_each_entry_safe(tx, tx_next, tx_list, list) {
2517                 tx->wait = wait;
2518                 list_del_init(&tx->list);
2519                 tx->next_descq_idx = 0;
2520 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2521                 tx->sn = sde->tail_sn++;
2522                 trace_hfi1_sdma_in_sn(sde, tx->sn);
2523 #endif
2524                 list_add_tail(&tx->list, &sde->flushlist);
2525                 flush_count++;
2526                 if (wait) {
2527                         wait->tx_count++;
2528                         wait->count += tx->num_desc;
2529                 }
2530         }
2531         spin_unlock(&sde->flushlist_lock);
2532         schedule_work(&sde->flush_worker);
2533         ret = -ECOMM;
2534         goto update_tail;
2535 nodesc:
2536         ret = sdma_check_progress(sde, wait, tx, submit_count > 0);
2537         if (ret == -EAGAIN) {
2538                 ret = 0;
2539                 goto retry;
2540         }
2541         sde->descq_full_count++;
2542         goto update_tail;
2543 }
2544
2545 static void sdma_process_event(struct sdma_engine *sde, enum sdma_events event)
2546 {
2547         unsigned long flags;
2548
2549         spin_lock_irqsave(&sde->tail_lock, flags);
2550         write_seqlock(&sde->head_lock);
2551
2552         __sdma_process_event(sde, event);
2553
2554         if (sde->state.current_state == sdma_state_s99_running)
2555                 sdma_desc_avail(sde, sdma_descq_freecnt(sde));
2556
2557         write_sequnlock(&sde->head_lock);
2558         spin_unlock_irqrestore(&sde->tail_lock, flags);
2559 }
2560
2561 static void __sdma_process_event(struct sdma_engine *sde,
2562                                  enum sdma_events event)
2563 {
2564         struct sdma_state *ss = &sde->state;
2565         int need_progress = 0;
2566
2567         /* CONFIG SDMA temporary */
2568 #ifdef CONFIG_SDMA_VERBOSITY
2569         dd_dev_err(sde->dd, "CONFIG SDMA(%u) [%s] %s\n", sde->this_idx,
2570                    sdma_state_names[ss->current_state],
2571                    sdma_event_names[event]);
2572 #endif
2573
2574         switch (ss->current_state) {
2575         case sdma_state_s00_hw_down:
2576                 switch (event) {
2577                 case sdma_event_e00_go_hw_down:
2578                         break;
2579                 case sdma_event_e30_go_running:
2580                         /*
2581                          * If down, but running requested (usually result
2582                          * of link up, then we need to start up.
2583                          * This can happen when hw down is requested while
2584                          * bringing the link up with traffic active on
2585                          * 7220, e.g.
2586                          */
2587                         ss->go_s99_running = 1;
2588                         /* fall through -- and start dma engine */
2589                 case sdma_event_e10_go_hw_start:
2590                         /* This reference means the state machine is started */
2591                         sdma_get(&sde->state);
2592                         sdma_set_state(sde,
2593                                        sdma_state_s10_hw_start_up_halt_wait);
2594                         break;
2595                 case sdma_event_e15_hw_halt_done:
2596                         break;
2597                 case sdma_event_e25_hw_clean_up_done:
2598                         break;
2599                 case sdma_event_e40_sw_cleaned:
2600                         sdma_sw_tear_down(sde);
2601                         break;
2602                 case sdma_event_e50_hw_cleaned:
2603                         break;
2604                 case sdma_event_e60_hw_halted:
2605                         break;
2606                 case sdma_event_e70_go_idle:
2607                         break;
2608                 case sdma_event_e80_hw_freeze:
2609                         break;
2610                 case sdma_event_e81_hw_frozen:
2611                         break;
2612                 case sdma_event_e82_hw_unfreeze:
2613                         break;
2614                 case sdma_event_e85_link_down:
2615                         break;
2616                 case sdma_event_e90_sw_halted:
2617                         break;
2618                 }
2619                 break;
2620
2621         case sdma_state_s10_hw_start_up_halt_wait:
2622                 switch (event) {
2623                 case sdma_event_e00_go_hw_down:
2624                         sdma_set_state(sde, sdma_state_s00_hw_down);
2625                         sdma_sw_tear_down(sde);
2626                         break;
2627                 case sdma_event_e10_go_hw_start:
2628                         break;
2629                 case sdma_event_e15_hw_halt_done:
2630                         sdma_set_state(sde,
2631                                        sdma_state_s15_hw_start_up_clean_wait);
2632                         sdma_start_hw_clean_up(sde);
2633                         break;
2634                 case sdma_event_e25_hw_clean_up_done:
2635                         break;
2636                 case sdma_event_e30_go_running:
2637                         ss->go_s99_running = 1;
2638                         break;
2639                 case sdma_event_e40_sw_cleaned:
2640                         break;
2641                 case sdma_event_e50_hw_cleaned:
2642                         break;
2643                 case sdma_event_e60_hw_halted:
2644                         schedule_work(&sde->err_halt_worker);
2645                         break;
2646                 case sdma_event_e70_go_idle:
2647                         ss->go_s99_running = 0;
2648                         break;
2649                 case sdma_event_e80_hw_freeze:
2650                         break;
2651                 case sdma_event_e81_hw_frozen:
2652                         break;
2653                 case sdma_event_e82_hw_unfreeze:
2654                         break;
2655                 case sdma_event_e85_link_down:
2656                         break;
2657                 case sdma_event_e90_sw_halted:
2658                         break;
2659                 }
2660                 break;
2661
2662         case sdma_state_s15_hw_start_up_clean_wait:
2663                 switch (event) {
2664                 case sdma_event_e00_go_hw_down:
2665                         sdma_set_state(sde, sdma_state_s00_hw_down);
2666                         sdma_sw_tear_down(sde);
2667                         break;
2668                 case sdma_event_e10_go_hw_start:
2669                         break;
2670                 case sdma_event_e15_hw_halt_done:
2671                         break;
2672                 case sdma_event_e25_hw_clean_up_done:
2673                         sdma_hw_start_up(sde);
2674                         sdma_set_state(sde, ss->go_s99_running ?
2675                                        sdma_state_s99_running :
2676                                        sdma_state_s20_idle);
2677                         break;
2678                 case sdma_event_e30_go_running:
2679                         ss->go_s99_running = 1;
2680                         break;
2681                 case sdma_event_e40_sw_cleaned:
2682                         break;
2683                 case sdma_event_e50_hw_cleaned:
2684                         break;
2685                 case sdma_event_e60_hw_halted:
2686                         break;
2687                 case sdma_event_e70_go_idle:
2688                         ss->go_s99_running = 0;
2689                         break;
2690                 case sdma_event_e80_hw_freeze:
2691                         break;
2692                 case sdma_event_e81_hw_frozen:
2693                         break;
2694                 case sdma_event_e82_hw_unfreeze:
2695                         break;
2696                 case sdma_event_e85_link_down:
2697                         break;
2698                 case sdma_event_e90_sw_halted:
2699                         break;
2700                 }
2701                 break;
2702
2703         case sdma_state_s20_idle:
2704                 switch (event) {
2705                 case sdma_event_e00_go_hw_down:
2706                         sdma_set_state(sde, sdma_state_s00_hw_down);
2707                         sdma_sw_tear_down(sde);
2708                         break;
2709                 case sdma_event_e10_go_hw_start:
2710                         break;
2711                 case sdma_event_e15_hw_halt_done:
2712                         break;
2713                 case sdma_event_e25_hw_clean_up_done:
2714                         break;
2715                 case sdma_event_e30_go_running:
2716                         sdma_set_state(sde, sdma_state_s99_running);
2717                         ss->go_s99_running = 1;
2718                         break;
2719                 case sdma_event_e40_sw_cleaned:
2720                         break;
2721                 case sdma_event_e50_hw_cleaned:
2722                         break;
2723                 case sdma_event_e60_hw_halted:
2724                         sdma_set_state(sde, sdma_state_s50_hw_halt_wait);
2725                         schedule_work(&sde->err_halt_worker);
2726                         break;
2727                 case sdma_event_e70_go_idle:
2728                         break;
2729                 case sdma_event_e85_link_down:
2730                         /* fall through */
2731                 case sdma_event_e80_hw_freeze:
2732                         sdma_set_state(sde, sdma_state_s80_hw_freeze);
2733                         atomic_dec(&sde->dd->sdma_unfreeze_count);
2734                         wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
2735                         break;
2736                 case sdma_event_e81_hw_frozen:
2737                         break;
2738                 case sdma_event_e82_hw_unfreeze:
2739                         break;
2740                 case sdma_event_e90_sw_halted:
2741                         break;
2742                 }
2743                 break;
2744
2745         case sdma_state_s30_sw_clean_up_wait:
2746                 switch (event) {
2747                 case sdma_event_e00_go_hw_down:
2748                         sdma_set_state(sde, sdma_state_s00_hw_down);
2749                         break;
2750                 case sdma_event_e10_go_hw_start:
2751                         break;
2752                 case sdma_event_e15_hw_halt_done:
2753                         break;
2754                 case sdma_event_e25_hw_clean_up_done:
2755                         break;
2756                 case sdma_event_e30_go_running:
2757                         ss->go_s99_running = 1;
2758                         break;
2759                 case sdma_event_e40_sw_cleaned:
2760                         sdma_set_state(sde, sdma_state_s40_hw_clean_up_wait);
2761                         sdma_start_hw_clean_up(sde);
2762                         break;
2763                 case sdma_event_e50_hw_cleaned:
2764                         break;
2765                 case sdma_event_e60_hw_halted:
2766                         break;
2767                 case sdma_event_e70_go_idle:
2768                         ss->go_s99_running = 0;
2769                         break;
2770                 case sdma_event_e80_hw_freeze:
2771                         break;
2772                 case sdma_event_e81_hw_frozen:
2773                         break;
2774                 case sdma_event_e82_hw_unfreeze:
2775                         break;
2776                 case sdma_event_e85_link_down:
2777                         ss->go_s99_running = 0;
2778                         break;
2779                 case sdma_event_e90_sw_halted:
2780                         break;
2781                 }
2782                 break;
2783
2784         case sdma_state_s40_hw_clean_up_wait:
2785                 switch (event) {
2786                 case sdma_event_e00_go_hw_down:
2787                         sdma_set_state(sde, sdma_state_s00_hw_down);
2788                         tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2789                         break;
2790                 case sdma_event_e10_go_hw_start:
2791                         break;
2792                 case sdma_event_e15_hw_halt_done:
2793                         break;
2794                 case sdma_event_e25_hw_clean_up_done:
2795                         sdma_hw_start_up(sde);
2796                         sdma_set_state(sde, ss->go_s99_running ?
2797                                        sdma_state_s99_running :
2798                                        sdma_state_s20_idle);
2799                         break;
2800                 case sdma_event_e30_go_running:
2801                         ss->go_s99_running = 1;
2802                         break;
2803                 case sdma_event_e40_sw_cleaned:
2804                         break;
2805                 case sdma_event_e50_hw_cleaned:
2806                         break;
2807                 case sdma_event_e60_hw_halted:
2808                         break;
2809                 case sdma_event_e70_go_idle:
2810                         ss->go_s99_running = 0;
2811                         break;
2812                 case sdma_event_e80_hw_freeze:
2813                         break;
2814                 case sdma_event_e81_hw_frozen:
2815                         break;
2816                 case sdma_event_e82_hw_unfreeze:
2817                         break;
2818                 case sdma_event_e85_link_down:
2819                         ss->go_s99_running = 0;
2820                         break;
2821                 case sdma_event_e90_sw_halted:
2822                         break;
2823                 }
2824                 break;
2825
2826         case sdma_state_s50_hw_halt_wait:
2827                 switch (event) {
2828                 case sdma_event_e00_go_hw_down:
2829                         sdma_set_state(sde, sdma_state_s00_hw_down);
2830                         tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2831                         break;
2832                 case sdma_event_e10_go_hw_start:
2833                         break;
2834                 case sdma_event_e15_hw_halt_done:
2835                         sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait);
2836                         tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2837                         break;
2838                 case sdma_event_e25_hw_clean_up_done:
2839                         break;
2840                 case sdma_event_e30_go_running:
2841                         ss->go_s99_running = 1;
2842                         break;
2843                 case sdma_event_e40_sw_cleaned:
2844                         break;
2845                 case sdma_event_e50_hw_cleaned:
2846                         break;
2847                 case sdma_event_e60_hw_halted:
2848                         schedule_work(&sde->err_halt_worker);
2849                         break;
2850                 case sdma_event_e70_go_idle:
2851                         ss->go_s99_running = 0;
2852                         break;
2853                 case sdma_event_e80_hw_freeze:
2854                         break;
2855                 case sdma_event_e81_hw_frozen:
2856                         break;
2857                 case sdma_event_e82_hw_unfreeze:
2858                         break;
2859                 case sdma_event_e85_link_down:
2860                         ss->go_s99_running = 0;
2861                         break;
2862                 case sdma_event_e90_sw_halted:
2863                         break;
2864                 }
2865                 break;
2866
2867         case sdma_state_s60_idle_halt_wait:
2868                 switch (event) {
2869                 case sdma_event_e00_go_hw_down:
2870                         sdma_set_state(sde, sdma_state_s00_hw_down);
2871                         tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2872                         break;
2873                 case sdma_event_e10_go_hw_start:
2874                         break;
2875                 case sdma_event_e15_hw_halt_done:
2876                         sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait);
2877                         tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2878                         break;
2879                 case sdma_event_e25_hw_clean_up_done:
2880                         break;
2881                 case sdma_event_e30_go_running:
2882                         ss->go_s99_running = 1;
2883                         break;
2884                 case sdma_event_e40_sw_cleaned:
2885                         break;
2886                 case sdma_event_e50_hw_cleaned:
2887                         break;
2888                 case sdma_event_e60_hw_halted:
2889                         schedule_work(&sde->err_halt_worker);
2890                         break;
2891                 case sdma_event_e70_go_idle:
2892                         ss->go_s99_running = 0;
2893                         break;
2894                 case sdma_event_e80_hw_freeze:
2895                         break;
2896                 case sdma_event_e81_hw_frozen:
2897                         break;
2898                 case sdma_event_e82_hw_unfreeze:
2899                         break;
2900                 case sdma_event_e85_link_down:
2901                         break;
2902                 case sdma_event_e90_sw_halted:
2903                         break;
2904                 }
2905                 break;
2906
2907         case sdma_state_s80_hw_freeze:
2908                 switch (event) {
2909                 case sdma_event_e00_go_hw_down:
2910                         sdma_set_state(sde, sdma_state_s00_hw_down);
2911                         tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2912                         break;
2913                 case sdma_event_e10_go_hw_start:
2914                         break;
2915                 case sdma_event_e15_hw_halt_done:
2916                         break;
2917                 case sdma_event_e25_hw_clean_up_done:
2918                         break;
2919                 case sdma_event_e30_go_running:
2920                         ss->go_s99_running = 1;
2921                         break;
2922                 case sdma_event_e40_sw_cleaned:
2923                         break;
2924                 case sdma_event_e50_hw_cleaned:
2925                         break;
2926                 case sdma_event_e60_hw_halted:
2927                         break;
2928                 case sdma_event_e70_go_idle:
2929                         ss->go_s99_running = 0;
2930                         break;
2931                 case sdma_event_e80_hw_freeze:
2932                         break;
2933                 case sdma_event_e81_hw_frozen:
2934                         sdma_set_state(sde, sdma_state_s82_freeze_sw_clean);
2935                         tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2936                         break;
2937                 case sdma_event_e82_hw_unfreeze:
2938                         break;
2939                 case sdma_event_e85_link_down:
2940                         break;
2941                 case sdma_event_e90_sw_halted:
2942                         break;
2943                 }
2944                 break;
2945
2946         case sdma_state_s82_freeze_sw_clean:
2947                 switch (event) {
2948                 case sdma_event_e00_go_hw_down:
2949                         sdma_set_state(sde, sdma_state_s00_hw_down);
2950                         tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2951                         break;
2952                 case sdma_event_e10_go_hw_start:
2953                         break;
2954                 case sdma_event_e15_hw_halt_done:
2955                         break;
2956                 case sdma_event_e25_hw_clean_up_done:
2957                         break;
2958                 case sdma_event_e30_go_running:
2959                         ss->go_s99_running = 1;
2960                         break;
2961                 case sdma_event_e40_sw_cleaned:
2962                         /* notify caller this engine is done cleaning */
2963                         atomic_dec(&sde->dd->sdma_unfreeze_count);
2964                         wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
2965                         break;
2966                 case sdma_event_e50_hw_cleaned:
2967                         break;
2968                 case sdma_event_e60_hw_halted:
2969                         break;
2970                 case sdma_event_e70_go_idle:
2971                         ss->go_s99_running = 0;
2972                         break;
2973                 case sdma_event_e80_hw_freeze:
2974                         break;
2975                 case sdma_event_e81_hw_frozen:
2976                         break;
2977                 case sdma_event_e82_hw_unfreeze:
2978                         sdma_hw_start_up(sde);
2979                         sdma_set_state(sde, ss->go_s99_running ?
2980                                        sdma_state_s99_running :
2981                                        sdma_state_s20_idle);
2982                         break;
2983                 case sdma_event_e85_link_down:
2984                         break;
2985                 case sdma_event_e90_sw_halted:
2986                         break;
2987                 }
2988                 break;
2989
2990         case sdma_state_s99_running:
2991                 switch (event) {
2992                 case sdma_event_e00_go_hw_down:
2993                         sdma_set_state(sde, sdma_state_s00_hw_down);
2994                         tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2995                         break;
2996                 case sdma_event_e10_go_hw_start:
2997                         break;
2998                 case sdma_event_e15_hw_halt_done:
2999                         break;
3000                 case sdma_event_e25_hw_clean_up_done:
3001                         break;
3002                 case sdma_event_e30_go_running:
3003                         break;
3004                 case sdma_event_e40_sw_cleaned:
3005                         break;
3006                 case sdma_event_e50_hw_cleaned:
3007                         break;
3008                 case sdma_event_e60_hw_halted:
3009                         need_progress = 1;
3010                         sdma_err_progress_check_schedule(sde);
3011                         /* fall through */
3012                 case sdma_event_e90_sw_halted:
3013                         /*
3014                         * SW initiated halt does not perform engines
3015                         * progress check
3016                         */
3017                         sdma_set_state(sde, sdma_state_s50_hw_halt_wait);
3018                         schedule_work(&sde->err_halt_worker);
3019                         break;
3020                 case sdma_event_e70_go_idle:
3021                         sdma_set_state(sde, sdma_state_s60_idle_halt_wait);
3022                         break;
3023                 case sdma_event_e85_link_down:
3024                         ss->go_s99_running = 0;
3025                         /* fall through */
3026                 case sdma_event_e80_hw_freeze:
3027                         sdma_set_state(sde, sdma_state_s80_hw_freeze);
3028                         atomic_dec(&sde->dd->sdma_unfreeze_count);
3029                         wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
3030                         break;
3031                 case sdma_event_e81_hw_frozen:
3032                         break;
3033                 case sdma_event_e82_hw_unfreeze:
3034                         break;
3035                 }
3036                 break;
3037         }
3038
3039         ss->last_event = event;
3040         if (need_progress)
3041                 sdma_make_progress(sde, 0);
3042 }
3043
3044 /*
3045  * _extend_sdma_tx_descs() - helper to extend txreq
3046  *
3047  * This is called once the initial nominal allocation
3048  * of descriptors in the sdma_txreq is exhausted.
3049  *
3050  * The code will bump the allocation up to the max
3051  * of MAX_DESC (64) descriptors. There doesn't seem
3052  * much point in an interim step. The last descriptor
3053  * is reserved for coalesce buffer in order to support
3054  * cases where input packet has >MAX_DESC iovecs.
3055  *
3056  */
3057 static int _extend_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx)
3058 {
3059         int i;
3060
3061         /* Handle last descriptor */
3062         if (unlikely((tx->num_desc == (MAX_DESC - 1)))) {
3063                 /* if tlen is 0, it is for padding, release last descriptor */
3064                 if (!tx->tlen) {
3065                         tx->desc_limit = MAX_DESC;
3066                 } else if (!tx->coalesce_buf) {
3067                         /* allocate coalesce buffer with space for padding */
3068                         tx->coalesce_buf = kmalloc(tx->tlen + sizeof(u32),
3069                                                    GFP_ATOMIC);
3070                         if (!tx->coalesce_buf)
3071                                 goto enomem;
3072                         tx->coalesce_idx = 0;
3073                 }
3074                 return 0;
3075         }
3076
3077         if (unlikely(tx->num_desc == MAX_DESC))
3078                 goto enomem;
3079
3080         tx->descp = kmalloc_array(
3081                         MAX_DESC,
3082                         sizeof(struct sdma_desc),
3083                         GFP_ATOMIC);
3084         if (!tx->descp)
3085                 goto enomem;
3086
3087         /* reserve last descriptor for coalescing */
3088         tx->desc_limit = MAX_DESC - 1;
3089         /* copy ones already built */
3090         for (i = 0; i < tx->num_desc; i++)
3091                 tx->descp[i] = tx->descs[i];
3092         return 0;
3093 enomem:
3094         __sdma_txclean(dd, tx);
3095         return -ENOMEM;
3096 }
3097
3098 /*
3099  * ext_coal_sdma_tx_descs() - extend or coalesce sdma tx descriptors
3100  *
3101  * This is called once the initial nominal allocation of descriptors
3102  * in the sdma_txreq is exhausted.
3103  *
3104  * This function calls _extend_sdma_tx_descs to extend or allocate
3105  * coalesce buffer. If there is a allocated coalesce buffer, it will
3106  * copy the input packet data into the coalesce buffer. It also adds
3107  * coalesce buffer descriptor once when whole packet is received.
3108  *
3109  * Return:
3110  * <0 - error
3111  * 0 - coalescing, don't populate descriptor
3112  * 1 - continue with populating descriptor
3113  */
3114 int ext_coal_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx,
3115                            int type, void *kvaddr, struct page *page,
3116                            unsigned long offset, u16 len)
3117 {
3118         int pad_len, rval;
3119         dma_addr_t addr;
3120
3121         rval = _extend_sdma_tx_descs(dd, tx);
3122         if (rval) {
3123                 __sdma_txclean(dd, tx);
3124                 return rval;
3125         }
3126
3127         /* If coalesce buffer is allocated, copy data into it */
3128         if (tx->coalesce_buf) {
3129                 if (type == SDMA_MAP_NONE) {
3130                         __sdma_txclean(dd, tx);
3131                         return -EINVAL;
3132                 }
3133
3134                 if (type == SDMA_MAP_PAGE) {
3135                         kvaddr = kmap(page);
3136                         kvaddr += offset;
3137                 } else if (WARN_ON(!kvaddr)) {
3138                         __sdma_txclean(dd, tx);
3139                         return -EINVAL;
3140                 }
3141
3142                 memcpy(tx->coalesce_buf + tx->coalesce_idx, kvaddr, len);
3143                 tx->coalesce_idx += len;
3144                 if (type == SDMA_MAP_PAGE)
3145                         kunmap(page);
3146
3147                 /* If there is more data, return */
3148                 if (tx->tlen - tx->coalesce_idx)
3149                         return 0;
3150
3151                 /* Whole packet is received; add any padding */
3152                 pad_len = tx->packet_len & (sizeof(u32) - 1);
3153                 if (pad_len) {
3154                         pad_len = sizeof(u32) - pad_len;
3155                         memset(tx->coalesce_buf + tx->coalesce_idx, 0, pad_len);
3156                         /* padding is taken care of for coalescing case */
3157                         tx->packet_len += pad_len;
3158                         tx->tlen += pad_len;
3159                 }
3160
3161                 /* dma map the coalesce buffer */
3162                 addr = dma_map_single(&dd->pcidev->dev,
3163                                       tx->coalesce_buf,
3164                                       tx->tlen,
3165                                       DMA_TO_DEVICE);
3166
3167                 if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) {
3168                         __sdma_txclean(dd, tx);
3169                         return -ENOSPC;
3170                 }
3171
3172                 /* Add descriptor for coalesce buffer */
3173                 tx->desc_limit = MAX_DESC;
3174                 return _sdma_txadd_daddr(dd, SDMA_MAP_SINGLE, tx,
3175                                          addr, tx->tlen);
3176         }
3177
3178         return 1;
3179 }
3180
3181 /* Update sdes when the lmc changes */
3182 void sdma_update_lmc(struct hfi1_devdata *dd, u64 mask, u32 lid)
3183 {
3184         struct sdma_engine *sde;
3185         int i;
3186         u64 sreg;
3187
3188         sreg = ((mask & SD(CHECK_SLID_MASK_MASK)) <<
3189                 SD(CHECK_SLID_MASK_SHIFT)) |
3190                 (((lid & mask) & SD(CHECK_SLID_VALUE_MASK)) <<
3191                 SD(CHECK_SLID_VALUE_SHIFT));
3192
3193         for (i = 0; i < dd->num_sdma; i++) {
3194                 hfi1_cdbg(LINKVERB, "SendDmaEngine[%d].SLID_CHECK = 0x%x",
3195                           i, (u32)sreg);
3196                 sde = &dd->per_sdma[i];
3197                 write_sde_csr(sde, SD(CHECK_SLID), sreg);
3198         }
3199 }
3200
3201 /* tx not dword sized - pad */
3202 int _pad_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx)
3203 {
3204         int rval = 0;
3205
3206         tx->num_desc++;
3207         if ((unlikely(tx->num_desc == tx->desc_limit))) {
3208                 rval = _extend_sdma_tx_descs(dd, tx);
3209                 if (rval) {
3210                         __sdma_txclean(dd, tx);
3211                         return rval;
3212                 }
3213         }
3214         /* finish the one just added */
3215         make_tx_sdma_desc(
3216                 tx,
3217                 SDMA_MAP_NONE,
3218                 dd->sdma_pad_phys,
3219                 sizeof(u32) - (tx->packet_len & (sizeof(u32) - 1)));
3220         _sdma_close_tx(dd, tx);
3221         return rval;
3222 }
3223
3224 /*
3225  * Add ahg to the sdma_txreq
3226  *
3227  * The logic will consume up to 3
3228  * descriptors at the beginning of
3229  * sdma_txreq.
3230  */
3231 void _sdma_txreq_ahgadd(
3232         struct sdma_txreq *tx,
3233         u8 num_ahg,
3234         u8 ahg_entry,
3235         u32 *ahg,
3236         u8 ahg_hlen)
3237 {
3238         u32 i, shift = 0, desc = 0;
3239         u8 mode;
3240
3241         WARN_ON_ONCE(num_ahg > 9 || (ahg_hlen & 3) || ahg_hlen == 4);
3242         /* compute mode */
3243         if (num_ahg == 1)
3244                 mode = SDMA_AHG_APPLY_UPDATE1;
3245         else if (num_ahg <= 5)
3246                 mode = SDMA_AHG_APPLY_UPDATE2;
3247         else
3248                 mode = SDMA_AHG_APPLY_UPDATE3;
3249         tx->num_desc++;
3250         /* initialize to consumed descriptors to zero */
3251         switch (mode) {
3252         case SDMA_AHG_APPLY_UPDATE3:
3253                 tx->num_desc++;
3254                 tx->descs[2].qw[0] = 0;
3255                 tx->descs[2].qw[1] = 0;
3256                 /* FALLTHROUGH */
3257         case SDMA_AHG_APPLY_UPDATE2:
3258                 tx->num_desc++;
3259                 tx->descs[1].qw[0] = 0;
3260                 tx->descs[1].qw[1] = 0;
3261                 break;
3262         }
3263         ahg_hlen >>= 2;
3264         tx->descs[0].qw[1] |=
3265                 (((u64)ahg_entry & SDMA_DESC1_HEADER_INDEX_MASK)
3266                         << SDMA_DESC1_HEADER_INDEX_SHIFT) |
3267                 (((u64)ahg_hlen & SDMA_DESC1_HEADER_DWS_MASK)
3268                         << SDMA_DESC1_HEADER_DWS_SHIFT) |
3269                 (((u64)mode & SDMA_DESC1_HEADER_MODE_MASK)
3270                         << SDMA_DESC1_HEADER_MODE_SHIFT) |
3271                 (((u64)ahg[0] & SDMA_DESC1_HEADER_UPDATE1_MASK)
3272                         << SDMA_DESC1_HEADER_UPDATE1_SHIFT);
3273         for (i = 0; i < (num_ahg - 1); i++) {
3274                 if (!shift && !(i & 2))
3275                         desc++;
3276                 tx->descs[desc].qw[!!(i & 2)] |=
3277                         (((u64)ahg[i + 1])
3278                                 << shift);
3279                 shift = (shift + 32) & 63;
3280         }
3281 }
3282
3283 /**
3284  * sdma_ahg_alloc - allocate an AHG entry
3285  * @sde: engine to allocate from
3286  *
3287  * Return:
3288  * 0-31 when successful, -EOPNOTSUPP if AHG is not enabled,
3289  * -ENOSPC if an entry is not available
3290  */
3291 int sdma_ahg_alloc(struct sdma_engine *sde)
3292 {
3293         int nr;
3294         int oldbit;
3295
3296         if (!sde) {
3297                 trace_hfi1_ahg_allocate(sde, -EINVAL);
3298                 return -EINVAL;
3299         }
3300         while (1) {
3301                 nr = ffz(READ_ONCE(sde->ahg_bits));
3302                 if (nr > 31) {
3303                         trace_hfi1_ahg_allocate(sde, -ENOSPC);
3304                         return -ENOSPC;
3305                 }
3306                 oldbit = test_and_set_bit(nr, &sde->ahg_bits);
3307                 if (!oldbit)
3308                         break;
3309                 cpu_relax();
3310         }
3311         trace_hfi1_ahg_allocate(sde, nr);
3312         return nr;
3313 }
3314
3315 /**
3316  * sdma_ahg_free - free an AHG entry
3317  * @sde: engine to return AHG entry
3318  * @ahg_index: index to free
3319  *
3320  * This routine frees the indicate AHG entry.
3321  */
3322 void sdma_ahg_free(struct sdma_engine *sde, int ahg_index)
3323 {
3324         if (!sde)
3325                 return;
3326         trace_hfi1_ahg_deallocate(sde, ahg_index);
3327         if (ahg_index < 0 || ahg_index > 31)
3328                 return;
3329         clear_bit(ahg_index, &sde->ahg_bits);
3330 }
3331
3332 /*
3333  * SPC freeze handling for SDMA engines.  Called when the driver knows
3334  * the SPC is going into a freeze but before the freeze is fully
3335  * settled.  Generally an error interrupt.
3336  *
3337  * This event will pull the engine out of running so no more entries can be
3338  * added to the engine's queue.
3339  */
3340 void sdma_freeze_notify(struct hfi1_devdata *dd, int link_down)
3341 {
3342         int i;
3343         enum sdma_events event = link_down ? sdma_event_e85_link_down :
3344                                              sdma_event_e80_hw_freeze;
3345
3346         /* set up the wait but do not wait here */
3347         atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma);
3348
3349         /* tell all engines to stop running and wait */
3350         for (i = 0; i < dd->num_sdma; i++)
3351                 sdma_process_event(&dd->per_sdma[i], event);
3352
3353         /* sdma_freeze() will wait for all engines to have stopped */
3354 }
3355
3356 /*
3357  * SPC freeze handling for SDMA engines.  Called when the driver knows
3358  * the SPC is fully frozen.
3359  */
3360 void sdma_freeze(struct hfi1_devdata *dd)
3361 {
3362         int i;
3363         int ret;
3364
3365         /*
3366          * Make sure all engines have moved out of the running state before
3367          * continuing.
3368          */
3369         ret = wait_event_interruptible(dd->sdma_unfreeze_wq,
3370                                        atomic_read(&dd->sdma_unfreeze_count) <=
3371                                        0);
3372         /* interrupted or count is negative, then unloading - just exit */
3373         if (ret || atomic_read(&dd->sdma_unfreeze_count) < 0)
3374                 return;
3375
3376         /* set up the count for the next wait */
3377         atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma);
3378
3379         /* tell all engines that the SPC is frozen, they can start cleaning */
3380         for (i = 0; i < dd->num_sdma; i++)
3381                 sdma_process_event(&dd->per_sdma[i], sdma_event_e81_hw_frozen);
3382
3383         /*
3384          * Wait for everyone to finish software clean before exiting.  The
3385          * software clean will read engine CSRs, so must be completed before
3386          * the next step, which will clear the engine CSRs.
3387          */
3388         (void)wait_event_interruptible(dd->sdma_unfreeze_wq,
3389                                 atomic_read(&dd->sdma_unfreeze_count) <= 0);
3390         /* no need to check results - done no matter what */
3391 }
3392
3393 /*
3394  * SPC freeze handling for the SDMA engines.  Called after the SPC is unfrozen.
3395  *
3396  * The SPC freeze acts like a SDMA halt and a hardware clean combined.  All
3397  * that is left is a software clean.  We could do it after the SPC is fully
3398  * frozen, but then we'd have to add another state to wait for the unfreeze.
3399  * Instead, just defer the software clean until the unfreeze step.
3400  */
3401 void sdma_unfreeze(struct hfi1_devdata *dd)
3402 {
3403         int i;
3404
3405         /* tell all engines start freeze clean up */
3406         for (i = 0; i < dd->num_sdma; i++)
3407                 sdma_process_event(&dd->per_sdma[i],
3408                                    sdma_event_e82_hw_unfreeze);
3409 }
3410
3411 /**
3412  * _sdma_engine_progress_schedule() - schedule progress on engine
3413  * @sde: sdma_engine to schedule progress
3414  *
3415  */
3416 void _sdma_engine_progress_schedule(
3417         struct sdma_engine *sde)
3418 {
3419         trace_hfi1_sdma_engine_progress(sde, sde->progress_mask);
3420         /* assume we have selected a good cpu */
3421         write_csr(sde->dd,
3422                   CCE_INT_FORCE + (8 * (IS_SDMA_START / 64)),
3423                   sde->progress_mask);
3424 }