drivers/misc/habanalabs/common/hw_queue.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /*
   4  * Copyright 2016-2019 HabanaLabs, Ltd.
   5  * All Rights Reserved.
   6  */
   7
   8 #include "habanalabs.h"
   9
  10 #include <linux/slab.h>
  11
  12 /*
  13  * hl_queue_add_ptr - add to pi or ci and checks if it wraps around
  14  *
  15  * @ptr: the current pi/ci value
  16  * @val: the amount to add
  17  *
  18  * Add val to ptr. It can go until twice the queue length.
  19  */
  20 inline u32 hl_hw_queue_add_ptr(u32 ptr, u16 val)
  21 {
  22         ptr += val;
  23         ptr &= ((HL_QUEUE_LENGTH << 1) - 1);
  24         return ptr;
  25 }
  26 static inline int queue_ci_get(atomic_t *ci, u32 queue_len)
  27 {
  28         return atomic_read(ci) & ((queue_len << 1) - 1);
  29 }
  30
  31 static inline int queue_free_slots(struct hl_hw_queue *q, u32 queue_len)
  32 {
  33         int delta = (q->pi - queue_ci_get(&q->ci, queue_len));
  34
  35         if (delta >= 0)
  36                 return (queue_len - delta);
  37         else
  38                 return (abs(delta) - queue_len);
  39 }
  40
  41 void hl_int_hw_queue_update_ci(struct hl_cs *cs)
  42 {
  43         struct hl_device *hdev = cs->ctx->hdev;
  44         struct hl_hw_queue *q;
  45         int i;
  46
  47         if (hdev->disabled)
  48                 return;
  49
  50         q = &hdev->kernel_queues[0];
  51         for (i = 0 ; i < hdev->asic_prop.max_queues ; i++, q++) {
  52                 if (q->queue_type == QUEUE_TYPE_INT)
  53                         atomic_add(cs->jobs_in_queue_cnt[i], &q->ci);
  54         }
  55 }
  56
  57 /*
  58  * ext_and_hw_queue_submit_bd() - Submit a buffer descriptor to an external or a
  59  *                                H/W queue.
  60  * @hdev: pointer to habanalabs device structure
  61  * @q: pointer to habanalabs queue structure
  62  * @ctl: BD's control word
  63  * @len: BD's length
  64  * @ptr: BD's pointer
  65  *
  66  * This function assumes there is enough space on the queue to submit a new
  67  * BD to it. It initializes the next BD and calls the device specific
  68  * function to set the pi (and doorbell)
  69  *
  70  * This function must be called when the scheduler mutex is taken
  71  *
  72  */
  73 static void ext_and_hw_queue_submit_bd(struct hl_device *hdev,
  74                         struct hl_hw_queue *q, u32 ctl, u32 len, u64 ptr)
  75 {
  76         struct hl_bd *bd;
  77
  78         bd = q->kernel_address;
  79         bd += hl_pi_2_offset(q->pi);
  80         bd->ctl = cpu_to_le32(ctl);
  81         bd->len = cpu_to_le32(len);
  82         bd->ptr = cpu_to_le64(ptr);
  83
  84         q->pi = hl_queue_inc_ptr(q->pi);
  85         hdev->asic_funcs->ring_doorbell(hdev, q->hw_queue_id, q->pi);
  86 }
  87
  88 /*
  89  * ext_queue_sanity_checks - perform some sanity checks on external queue
  90  *
  91  * @hdev              : pointer to hl_device structure
  92  * @q                 : pointer to hl_hw_queue structure
  93  * @num_of_entries    : how many entries to check for space
  94  * @reserve_cq_entry  : whether to reserve an entry in the cq
  95  *
  96  * H/W queues spinlock should be taken before calling this function
  97  *
  98  * Perform the following:
  99  * - Make sure we have enough space in the h/w queue
 100  * - Make sure we have enough space in the completion queue
 101  * - Reserve space in the completion queue (needs to be reversed if there
 102  *   is a failure down the road before the actual submission of work). Only
 103  *   do this action if reserve_cq_entry is true
 104  *
 105  */
 106 static int ext_queue_sanity_checks(struct hl_device *hdev,
 107                                 struct hl_hw_queue *q, int num_of_entries,
 108                                 bool reserve_cq_entry)
 109 {
 110         atomic_t *free_slots =
 111                         &hdev->completion_queue[q->cq_id].free_slots_cnt;
 112         int free_slots_cnt;
 113
 114         /* Check we have enough space in the queue */
 115         free_slots_cnt = queue_free_slots(q, HL_QUEUE_LENGTH);
 116
 117         if (free_slots_cnt < num_of_entries) {
 118                 dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
 119                         q->hw_queue_id, num_of_entries);
 120                 return -EAGAIN;
 121         }
 122
 123         if (reserve_cq_entry) {
 124                 /*
 125                  * Check we have enough space in the completion queue
 126                  * Add -1 to counter (decrement) unless counter was already 0
 127                  * In that case, CQ is full so we can't submit a new CB because
 128                  * we won't get ack on its completion
 129                  * atomic_add_unless will return 0 if counter was already 0
 130                  */
 131                 if (atomic_add_negative(num_of_entries * -1, free_slots)) {
 132                         dev_dbg(hdev->dev, "No space for %d on CQ %d\n",
 133                                 num_of_entries, q->hw_queue_id);
 134                         atomic_add(num_of_entries, free_slots);
 135                         return -EAGAIN;
 136                 }
 137         }
 138
 139         return 0;
 140 }
 141
 142 /*
 143  * int_queue_sanity_checks - perform some sanity checks on internal queue
 144  *
 145  * @hdev              : pointer to hl_device structure
 146  * @q                 : pointer to hl_hw_queue structure
 147  * @num_of_entries    : how many entries to check for space
 148  *
 149  * H/W queues spinlock should be taken before calling this function
 150  *
 151  * Perform the following:
 152  * - Make sure we have enough space in the h/w queue
 153  *
 154  */
 155 static int int_queue_sanity_checks(struct hl_device *hdev,
 156                                         struct hl_hw_queue *q,
 157                                         int num_of_entries)
 158 {
 159         int free_slots_cnt;
 160
 161         if (num_of_entries > q->int_queue_len) {
 162                 dev_err(hdev->dev,
 163                         "Cannot populate queue %u with %u jobs\n",
 164                         q->hw_queue_id, num_of_entries);
 165                 return -ENOMEM;
 166         }
 167
 168         /* Check we have enough space in the queue */
 169         free_slots_cnt = queue_free_slots(q, q->int_queue_len);
 170
 171         if (free_slots_cnt < num_of_entries) {
 172                 dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
 173                         q->hw_queue_id, num_of_entries);
 174                 return -EAGAIN;
 175         }
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * hw_queue_sanity_checks() - Make sure we have enough space in the h/w queue
 182  * @hdev: Pointer to hl_device structure.
 183  * @q: Pointer to hl_hw_queue structure.
 184  * @num_of_entries: How many entries to check for space.
 185  *
 186  * Notice: We do not reserve queue entries so this function mustn't be called
 187  *         more than once per CS for the same queue
 188  *
 189  */
 190 static int hw_queue_sanity_checks(struct hl_device *hdev, struct hl_hw_queue *q,
 191                                         int num_of_entries)
 192 {
 193         int free_slots_cnt;
 194
 195         /* Check we have enough space in the queue */
 196         free_slots_cnt = queue_free_slots(q, HL_QUEUE_LENGTH);
 197
 198         if (free_slots_cnt < num_of_entries) {
 199                 dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
 200                         q->hw_queue_id, num_of_entries);
 201                 return -EAGAIN;
 202         }
 203
 204         return 0;
 205 }
 206
 207 /*
 208  * hl_hw_queue_send_cb_no_cmpl - send a single CB (not a JOB) without completion
 209  *
 210  * @hdev: pointer to hl_device structure
 211  * @hw_queue_id: Queue's type
 212  * @cb_size: size of CB
 213  * @cb_ptr: pointer to CB location
 214  *
 215  * This function sends a single CB, that must NOT generate a completion entry
 216  *
 217  */
 218 int hl_hw_queue_send_cb_no_cmpl(struct hl_device *hdev, u32 hw_queue_id,
 219                                 u32 cb_size, u64 cb_ptr)
 220 {
 221         struct hl_hw_queue *q = &hdev->kernel_queues[hw_queue_id];
 222         int rc = 0;
 223
 224         /*
 225          * The CPU queue is a synchronous queue with an effective depth of
 226          * a single entry (although it is allocated with room for multiple
 227          * entries). Therefore, there is a different lock, called
 228          * send_cpu_message_lock, that serializes accesses to the CPU queue.
 229          * As a result, we don't need to lock the access to the entire H/W
 230          * queues module when submitting a JOB to the CPU queue
 231          */
 232         if (q->queue_type != QUEUE_TYPE_CPU)
 233                 hdev->asic_funcs->hw_queues_lock(hdev);
 234
 235         if (hdev->disabled) {
 236                 rc = -EPERM;
 237                 goto out;
 238         }
 239
 240         /*
 241          * hl_hw_queue_send_cb_no_cmpl() is called for queues of a H/W queue
 242          * type only on init phase, when the queues are empty and being tested,
 243          * so there is no need for sanity checks.
 244          */
 245         if (q->queue_type != QUEUE_TYPE_HW) {
 246                 rc = ext_queue_sanity_checks(hdev, q, 1, false);
 247                 if (rc)
 248                         goto out;
 249         }
 250
 251         ext_and_hw_queue_submit_bd(hdev, q, 0, cb_size, cb_ptr);
 252
 253 out:
 254         if (q->queue_type != QUEUE_TYPE_CPU)
 255                 hdev->asic_funcs->hw_queues_unlock(hdev);
 256
 257         return rc;
 258 }
 259
 260 /*
 261  * ext_queue_schedule_job - submit a JOB to an external queue
 262  *
 263  * @job: pointer to the job that needs to be submitted to the queue
 264  *
 265  * This function must be called when the scheduler mutex is taken
 266  *
 267  */
 268 static void ext_queue_schedule_job(struct hl_cs_job *job)
 269 {
 270         struct hl_device *hdev = job->cs->ctx->hdev;
 271         struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
 272         struct hl_cq_entry cq_pkt;
 273         struct hl_cq *cq;
 274         u64 cq_addr;
 275         struct hl_cb *cb;
 276         u32 ctl;
 277         u32 len;
 278         u64 ptr;
 279
 280         /*
 281          * Update the JOB ID inside the BD CTL so the device would know what
 282          * to write in the completion queue
 283          */
 284         ctl = ((q->pi << BD_CTL_SHADOW_INDEX_SHIFT) & BD_CTL_SHADOW_INDEX_MASK);
 285
 286         cb = job->patched_cb;
 287         len = job->job_cb_size;
 288         ptr = cb->bus_address;
 289
 290         cq_pkt.data = cpu_to_le32(
 291                         ((q->pi << CQ_ENTRY_SHADOW_INDEX_SHIFT)
 292                                 & CQ_ENTRY_SHADOW_INDEX_MASK) |
 293                         FIELD_PREP(CQ_ENTRY_SHADOW_INDEX_VALID_MASK, 1) |
 294                         FIELD_PREP(CQ_ENTRY_READY_MASK, 1));
 295
 296         /*
 297          * No need to protect pi_offset because scheduling to the
 298          * H/W queues is done under the scheduler mutex
 299          *
 300          * No need to check if CQ is full because it was already
 301          * checked in ext_queue_sanity_checks
 302          */
 303         cq = &hdev->completion_queue[q->cq_id];
 304         cq_addr = cq->bus_address + cq->pi * sizeof(struct hl_cq_entry);
 305
 306         hdev->asic_funcs->add_end_of_cb_packets(hdev, cb->kernel_address, len,
 307                                                 cq_addr,
 308                                                 le32_to_cpu(cq_pkt.data),
 309                                                 q->msi_vec,
 310                                                 job->contains_dma_pkt);
 311
 312         q->shadow_queue[hl_pi_2_offset(q->pi)] = job;
 313
 314         cq->pi = hl_cq_inc_ptr(cq->pi);
 315
 316         ext_and_hw_queue_submit_bd(hdev, q, ctl, len, ptr);
 317 }
 318
 319 /*
 320  * int_queue_schedule_job - submit a JOB to an internal queue
 321  *
 322  * @job: pointer to the job that needs to be submitted to the queue
 323  *
 324  * This function must be called when the scheduler mutex is taken
 325  *
 326  */
 327 static void int_queue_schedule_job(struct hl_cs_job *job)
 328 {
 329         struct hl_device *hdev = job->cs->ctx->hdev;
 330         struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
 331         struct hl_bd bd;
 332         __le64 *pi;
 333
 334         bd.ctl = 0;
 335         bd.len = cpu_to_le32(job->job_cb_size);
 336         bd.ptr = cpu_to_le64((u64) (uintptr_t) job->user_cb);
 337
 338         pi = q->kernel_address + (q->pi & (q->int_queue_len - 1)) * sizeof(bd);
 339
 340         q->pi++;
 341         q->pi &= ((q->int_queue_len << 1) - 1);
 342
 343         hdev->asic_funcs->pqe_write(hdev, pi, &bd);
 344
 345         hdev->asic_funcs->ring_doorbell(hdev, q->hw_queue_id, q->pi);
 346 }
 347
 348 /*
 349  * hw_queue_schedule_job - submit a JOB to a H/W queue
 350  *
 351  * @job: pointer to the job that needs to be submitted to the queue
 352  *
 353  * This function must be called when the scheduler mutex is taken
 354  *
 355  */
 356 static void hw_queue_schedule_job(struct hl_cs_job *job)
 357 {
 358         struct hl_device *hdev = job->cs->ctx->hdev;
 359         struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
 360         u64 ptr;
 361         u32 offset, ctl, len;
 362
 363         /*
 364          * Upon PQE completion, COMP_DATA is used as the write data to the
 365          * completion queue (QMAN HBW message), and COMP_OFFSET is used as the
 366          * write address offset in the SM block (QMAN LBW message).
 367          * The write address offset is calculated as "COMP_OFFSET << 2".
 368          */
 369         offset = job->cs->sequence & (hdev->asic_prop.max_pending_cs - 1);
 370         ctl = ((offset << BD_CTL_COMP_OFFSET_SHIFT) & BD_CTL_COMP_OFFSET_MASK) |
 371                 ((q->pi << BD_CTL_COMP_DATA_SHIFT) & BD_CTL_COMP_DATA_MASK);
 372
 373         len = job->job_cb_size;
 374
 375         /*
 376          * A patched CB is created only if a user CB was allocated by driver and
 377          * MMU is disabled. If MMU is enabled, the user CB should be used
 378          * instead. If the user CB wasn't allocated by driver, assume that it
 379          * holds an address.
 380          */
 381         if (job->patched_cb)
 382                 ptr = job->patched_cb->bus_address;
 383         else if (job->is_kernel_allocated_cb)
 384                 ptr = job->user_cb->bus_address;
 385         else
 386                 ptr = (u64) (uintptr_t) job->user_cb;
 387
 388         ext_and_hw_queue_submit_bd(hdev, q, ctl, len, ptr);
 389 }
 390
 391 /*
 392  * init_signal_wait_cs - initialize a signal/wait CS
 393  * @cs: pointer to the signal/wait CS
 394  *
 395  * H/W queues spinlock should be taken before calling this function
 396  */
 397 static void init_signal_wait_cs(struct hl_cs *cs)
 398 {
 399         struct hl_ctx *ctx = cs->ctx;
 400         struct hl_device *hdev = ctx->hdev;
 401         struct hl_hw_queue *hw_queue;
 402         struct hl_cs_compl *cs_cmpl =
 403                         container_of(cs->fence, struct hl_cs_compl, base_fence);
 404
 405         struct hl_hw_sob *hw_sob;
 406         struct hl_cs_job *job;
 407         u32 q_idx;
 408
 409         /* There is only one job in a signal/wait CS */
 410         job = list_first_entry(&cs->job_list, struct hl_cs_job,
 411                                 cs_node);
 412         q_idx = job->hw_queue_id;
 413         hw_queue = &hdev->kernel_queues[q_idx];
 414
 415         if (cs->type & CS_TYPE_SIGNAL) {
 416                 hw_sob = &hw_queue->hw_sob[hw_queue->curr_sob_offset];
 417
 418                 cs_cmpl->hw_sob = hw_sob;
 419                 cs_cmpl->sob_val = hw_queue->next_sob_val++;
 420
 421                 dev_dbg(hdev->dev,
 422                         "generate signal CB, sob_id: %d, sob val: 0x%x, q_idx: %d\n",
 423                         cs_cmpl->hw_sob->sob_id, cs_cmpl->sob_val, q_idx);
 424
 425                 hdev->asic_funcs->gen_signal_cb(hdev, job->patched_cb,
 426                                         cs_cmpl->hw_sob->sob_id);
 427
 428                 kref_get(&hw_sob->kref);
 429
 430                 /* check for wraparound */
 431                 if (hw_queue->next_sob_val == HL_MAX_SOB_VAL) {
 432                         /*
 433                          * Decrement as we reached the max value.
 434                          * The release function won't be called here as we've
 435                          * just incremented the refcount.
 436                          */
 437                         kref_put(&hw_sob->kref, hl_sob_reset_error);
 438                         hw_queue->next_sob_val = 1;
 439                         /* only two SOBs are currently in use */
 440                         hw_queue->curr_sob_offset =
 441                                         (hw_queue->curr_sob_offset + 1) %
 442                                                 HL_RSVD_SOBS_IN_USE;
 443
 444                         dev_dbg(hdev->dev, "switched to SOB %d, q_idx: %d\n",
 445                                         hw_queue->curr_sob_offset, q_idx);
 446                 }
 447         } else if (cs->type & CS_TYPE_WAIT) {
 448                 struct hl_cs_compl *signal_cs_cmpl;
 449
 450                 signal_cs_cmpl = container_of(cs->signal_fence,
 451                                                 struct hl_cs_compl,
 452                                                 base_fence);
 453
 454                 /* copy the the SOB id and value of the signal CS */
 455                 cs_cmpl->hw_sob = signal_cs_cmpl->hw_sob;
 456                 cs_cmpl->sob_val = signal_cs_cmpl->sob_val;
 457
 458                 dev_dbg(hdev->dev,
 459                         "generate wait CB, sob_id: %d, sob_val: 0x%x, mon_id: %d, q_idx: %d\n",
 460                         cs_cmpl->hw_sob->sob_id, cs_cmpl->sob_val,
 461                         hw_queue->base_mon_id, q_idx);
 462
 463                 hdev->asic_funcs->gen_wait_cb(hdev, job->patched_cb,
 464                                                 cs_cmpl->hw_sob->sob_id,
 465                                                 cs_cmpl->sob_val,
 466                                                 hw_queue->base_mon_id,
 467                                                 q_idx);
 468
 469                 kref_get(&cs_cmpl->hw_sob->kref);
 470                 /*
 471                  * Must put the signal fence after the SOB refcnt increment so
 472                  * the SOB refcnt won't turn 0 and reset the SOB before the
 473                  * wait CS was submitted.
 474                  */
 475                 mb();
 476                 hl_fence_put(cs->signal_fence);
 477                 cs->signal_fence = NULL;
 478         }
 479 }
 480
 481 /*
 482  * hl_hw_queue_schedule_cs - schedule a command submission
 483  * @cs: pointer to the CS
 484  */
 485 int hl_hw_queue_schedule_cs(struct hl_cs *cs)
 486 {
 487         struct hl_ctx *ctx = cs->ctx;
 488         struct hl_device *hdev = ctx->hdev;
 489         struct hl_cs_job *job, *tmp;
 490         struct hl_hw_queue *q;
 491         u32 max_queues;
 492         int rc = 0, i, cq_cnt;
 493
 494         hdev->asic_funcs->hw_queues_lock(hdev);
 495
 496         if (hl_device_disabled_or_in_reset(hdev)) {
 497                 ctx->cs_counters.device_in_reset_drop_cnt++;
 498                 dev_err(hdev->dev,
 499                         "device is disabled or in reset, CS rejected!\n");
 500                 rc = -EPERM;
 501                 goto out;
 502         }
 503
 504         max_queues = hdev->asic_prop.max_queues;
 505
 506         q = &hdev->kernel_queues[0];
 507         for (i = 0, cq_cnt = 0 ; i < max_queues ; i++, q++) {
 508                 if (cs->jobs_in_queue_cnt[i]) {
 509                         switch (q->queue_type) {
 510                         case QUEUE_TYPE_EXT:
 511                                 rc = ext_queue_sanity_checks(hdev, q,
 512                                                 cs->jobs_in_queue_cnt[i], true);
 513                                 break;
 514                         case QUEUE_TYPE_INT:
 515                                 rc = int_queue_sanity_checks(hdev, q,
 516                                                 cs->jobs_in_queue_cnt[i]);
 517                                 break;
 518                         case QUEUE_TYPE_HW:
 519                                 rc = hw_queue_sanity_checks(hdev, q,
 520                                                 cs->jobs_in_queue_cnt[i]);
 521                                 break;
 522                         default:
 523                                 dev_err(hdev->dev, "Queue type %d is invalid\n",
 524                                         q->queue_type);
 525                                 rc = -EINVAL;
 526                                 break;
 527                         }
 528
 529                         if (rc) {
 530                                 ctx->cs_counters.queue_full_drop_cnt++;
 531                                 goto unroll_cq_resv;
 532                         }
 533
 534                         if (q->queue_type == QUEUE_TYPE_EXT)
 535                                 cq_cnt++;
 536                 }
 537         }
 538
 539         if ((cs->type == CS_TYPE_SIGNAL) || (cs->type == CS_TYPE_WAIT))
 540                 init_signal_wait_cs(cs);
 541
 542         spin_lock(&hdev->hw_queues_mirror_lock);
 543         list_add_tail(&cs->mirror_node, &hdev->hw_queues_mirror_list);
 544
 545         /* Queue TDR if the CS is the first entry and if timeout is wanted */
 546         if ((hdev->timeout_jiffies != MAX_SCHEDULE_TIMEOUT) &&
 547                         (list_first_entry(&hdev->hw_queues_mirror_list,
 548                                         struct hl_cs, mirror_node) == cs)) {
 549                 cs->tdr_active = true;
 550                 schedule_delayed_work(&cs->work_tdr, hdev->timeout_jiffies);
 551                 spin_unlock(&hdev->hw_queues_mirror_lock);
 552         } else {
 553                 spin_unlock(&hdev->hw_queues_mirror_lock);
 554         }
 555
 556         if (!hdev->cs_active_cnt++) {
 557                 struct hl_device_idle_busy_ts *ts;
 558
 559                 ts = &hdev->idle_busy_ts_arr[hdev->idle_busy_ts_idx];
 560                 ts->busy_to_idle_ts = ktime_set(0, 0);
 561                 ts->idle_to_busy_ts = ktime_get();
 562         }
 563
 564         list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
 565                 switch (job->queue_type) {
 566                 case QUEUE_TYPE_EXT:
 567                         ext_queue_schedule_job(job);
 568                         break;
 569                 case QUEUE_TYPE_INT:
 570                         int_queue_schedule_job(job);
 571                         break;
 572                 case QUEUE_TYPE_HW:
 573                         hw_queue_schedule_job(job);
 574                         break;
 575                 default:
 576                         break;
 577                 }
 578
 579         cs->submitted = true;
 580
 581         goto out;
 582
 583 unroll_cq_resv:
 584         q = &hdev->kernel_queues[0];
 585         for (i = 0 ; (i < max_queues) && (cq_cnt > 0) ; i++, q++) {
 586                 if ((q->queue_type == QUEUE_TYPE_EXT) &&
 587                                                 (cs->jobs_in_queue_cnt[i])) {
 588                         atomic_t *free_slots =
 589                                 &hdev->completion_queue[i].free_slots_cnt;
 590                         atomic_add(cs->jobs_in_queue_cnt[i], free_slots);
 591                         cq_cnt--;
 592                 }
 593         }
 594
 595 out:
 596         hdev->asic_funcs->hw_queues_unlock(hdev);
 597
 598         return rc;
 599 }
 600
 601 /*
 602  * hl_hw_queue_inc_ci_kernel - increment ci for kernel's queue
 603  *
 604  * @hdev: pointer to hl_device structure
 605  * @hw_queue_id: which queue to increment its ci
 606  */
 607 void hl_hw_queue_inc_ci_kernel(struct hl_device *hdev, u32 hw_queue_id)
 608 {
 609         struct hl_hw_queue *q = &hdev->kernel_queues[hw_queue_id];
 610
 611         atomic_inc(&q->ci);
 612 }
 613
 614 static int ext_and_cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
 615                                         bool is_cpu_queue)
 616 {
 617         void *p;
 618         int rc;
 619
 620         if (is_cpu_queue)
 621                 p = hdev->asic_funcs->cpu_accessible_dma_pool_alloc(hdev,
 622                                                         HL_QUEUE_SIZE_IN_BYTES,
 623                                                         &q->bus_address);
 624         else
 625                 p = hdev->asic_funcs->asic_dma_alloc_coherent(hdev,
 626                                                 HL_QUEUE_SIZE_IN_BYTES,
 627                                                 &q->bus_address,
 628                                                 GFP_KERNEL | __GFP_ZERO);
 629         if (!p)
 630                 return -ENOMEM;
 631
 632         q->kernel_address = p;
 633
 634         q->shadow_queue = kmalloc_array(HL_QUEUE_LENGTH,
 635                                         sizeof(*q->shadow_queue),
 636                                         GFP_KERNEL);
 637         if (!q->shadow_queue) {
 638                 dev_err(hdev->dev,
 639                         "Failed to allocate shadow queue for H/W queue %d\n",
 640                         q->hw_queue_id);
 641                 rc = -ENOMEM;
 642                 goto free_queue;
 643         }
 644
 645         /* Make sure read/write pointers are initialized to start of queue */
 646         atomic_set(&q->ci, 0);
 647         q->pi = 0;
 648
 649         return 0;
 650
 651 free_queue:
 652         if (is_cpu_queue)
 653                 hdev->asic_funcs->cpu_accessible_dma_pool_free(hdev,
 654                                         HL_QUEUE_SIZE_IN_BYTES,
 655                                         q->kernel_address);
 656         else
 657                 hdev->asic_funcs->asic_dma_free_coherent(hdev,
 658                                         HL_QUEUE_SIZE_IN_BYTES,
 659                                         q->kernel_address,
 660                                         q->bus_address);
 661
 662         return rc;
 663 }
 664
 665 static int int_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 666 {
 667         void *p;
 668
 669         p = hdev->asic_funcs->get_int_queue_base(hdev, q->hw_queue_id,
 670                                         &q->bus_address, &q->int_queue_len);
 671         if (!p) {
 672                 dev_err(hdev->dev,
 673                         "Failed to get base address for internal queue %d\n",
 674                         q->hw_queue_id);
 675                 return -EFAULT;
 676         }
 677
 678         q->kernel_address = p;
 679         q->pi = 0;
 680         atomic_set(&q->ci, 0);
 681
 682         return 0;
 683 }
 684
 685 static int cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 686 {
 687         return ext_and_cpu_queue_init(hdev, q, true);
 688 }
 689
 690 static int ext_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 691 {
 692         return ext_and_cpu_queue_init(hdev, q, false);
 693 }
 694
 695 static int hw_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 696 {
 697         void *p;
 698
 699         p = hdev->asic_funcs->asic_dma_alloc_coherent(hdev,
 700                                                 HL_QUEUE_SIZE_IN_BYTES,
 701                                                 &q->bus_address,
 702                                                 GFP_KERNEL | __GFP_ZERO);
 703         if (!p)
 704                 return -ENOMEM;
 705
 706         q->kernel_address = p;
 707
 708         /* Make sure read/write pointers are initialized to start of queue */
 709         atomic_set(&q->ci, 0);
 710         q->pi = 0;
 711
 712         return 0;
 713 }
 714
 715 static void sync_stream_queue_init(struct hl_device *hdev, u32 q_idx)
 716 {
 717         struct hl_hw_queue *hw_queue = &hdev->kernel_queues[q_idx];
 718         struct asic_fixed_properties *prop = &hdev->asic_prop;
 719         struct hl_hw_sob *hw_sob;
 720         int sob, queue_idx = hdev->sync_stream_queue_idx++;
 721
 722         hw_queue->base_sob_id =
 723                 prop->sync_stream_first_sob + queue_idx * HL_RSVD_SOBS;
 724         hw_queue->base_mon_id =
 725                 prop->sync_stream_first_mon + queue_idx * HL_RSVD_MONS;
 726         hw_queue->next_sob_val = 1;
 727         hw_queue->curr_sob_offset = 0;
 728
 729         for (sob = 0 ; sob < HL_RSVD_SOBS ; sob++) {
 730                 hw_sob = &hw_queue->hw_sob[sob];
 731                 hw_sob->hdev = hdev;
 732                 hw_sob->sob_id = hw_queue->base_sob_id + sob;
 733                 hw_sob->q_idx = q_idx;
 734                 kref_init(&hw_sob->kref);
 735         }
 736 }
 737
 738 static void sync_stream_queue_reset(struct hl_device *hdev, u32 q_idx)
 739 {
 740         struct hl_hw_queue *hw_queue = &hdev->kernel_queues[q_idx];
 741
 742         /*
 743          * In case we got here due to a stuck CS, the refcnt might be bigger
 744          * than 1 and therefore we reset it.
 745          */
 746         kref_init(&hw_queue->hw_sob[hw_queue->curr_sob_offset].kref);
 747         hw_queue->curr_sob_offset = 0;
 748         hw_queue->next_sob_val = 1;
 749 }
 750
 751 /*
 752  * queue_init - main initialization function for H/W queue object
 753  *
 754  * @hdev: pointer to hl_device device structure
 755  * @q: pointer to hl_hw_queue queue structure
 756  * @hw_queue_id: The id of the H/W queue
 757  *
 758  * Allocate dma-able memory for the queue and initialize fields
 759  * Returns 0 on success
 760  */
 761 static int queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
 762                         u32 hw_queue_id)
 763 {
 764         int rc;
 765
 766         q->hw_queue_id = hw_queue_id;
 767
 768         switch (q->queue_type) {
 769         case QUEUE_TYPE_EXT:
 770                 rc = ext_queue_init(hdev, q);
 771                 break;
 772         case QUEUE_TYPE_INT:
 773                 rc = int_queue_init(hdev, q);
 774                 break;
 775         case QUEUE_TYPE_CPU:
 776                 rc = cpu_queue_init(hdev, q);
 777                 break;
 778         case QUEUE_TYPE_HW:
 779                 rc = hw_queue_init(hdev, q);
 780                 break;
 781         case QUEUE_TYPE_NA:
 782                 q->valid = 0;
 783                 return 0;
 784         default:
 785                 dev_crit(hdev->dev, "wrong queue type %d during init\n",
 786                         q->queue_type);
 787                 rc = -EINVAL;
 788                 break;
 789         }
 790
 791         if (q->supports_sync_stream)
 792                 sync_stream_queue_init(hdev, q->hw_queue_id);
 793
 794         if (rc)
 795                 return rc;
 796
 797         q->valid = 1;
 798
 799         return 0;
 800 }
 801
 802 /*
 803  * hw_queue_fini - destroy queue
 804  *
 805  * @hdev: pointer to hl_device device structure
 806  * @q: pointer to hl_hw_queue queue structure
 807  *
 808  * Free the queue memory
 809  */
 810 static void queue_fini(struct hl_device *hdev, struct hl_hw_queue *q)
 811 {
 812         if (!q->valid)
 813                 return;
 814
 815         /*
 816          * If we arrived here, there are no jobs waiting on this queue
 817          * so we can safely remove it.
 818          * This is because this function can only called when:
 819          * 1. Either a context is deleted, which only can occur if all its
 820          *    jobs were finished
 821          * 2. A context wasn't able to be created due to failure or timeout,
 822          *    which means there are no jobs on the queue yet
 823          *
 824          * The only exception are the queues of the kernel context, but
 825          * if they are being destroyed, it means that the entire module is
 826          * being removed. If the module is removed, it means there is no open
 827          * user context. It also means that if a job was submitted by
 828          * the kernel driver (e.g. context creation), the job itself was
 829          * released by the kernel driver when a timeout occurred on its
 830          * Completion. Thus, we don't need to release it again.
 831          */
 832
 833         if (q->queue_type == QUEUE_TYPE_INT)
 834                 return;
 835
 836         kfree(q->shadow_queue);
 837
 838         if (q->queue_type == QUEUE_TYPE_CPU)
 839                 hdev->asic_funcs->cpu_accessible_dma_pool_free(hdev,
 840                                         HL_QUEUE_SIZE_IN_BYTES,
 841                                         q->kernel_address);
 842         else
 843                 hdev->asic_funcs->asic_dma_free_coherent(hdev,
 844                                         HL_QUEUE_SIZE_IN_BYTES,
 845                                         q->kernel_address,
 846                                         q->bus_address);
 847 }
 848
 849 int hl_hw_queues_create(struct hl_device *hdev)
 850 {
 851         struct asic_fixed_properties *asic = &hdev->asic_prop;
 852         struct hl_hw_queue *q;
 853         int i, rc, q_ready_cnt;
 854
 855         hdev->kernel_queues = kcalloc(asic->max_queues,
 856                                 sizeof(*hdev->kernel_queues), GFP_KERNEL);
 857
 858         if (!hdev->kernel_queues) {
 859                 dev_err(hdev->dev, "Not enough memory for H/W queues\n");
 860                 return -ENOMEM;
 861         }
 862
 863         /* Initialize the H/W queues */
 864         for (i = 0, q_ready_cnt = 0, q = hdev->kernel_queues;
 865                         i < asic->max_queues ; i++, q_ready_cnt++, q++) {
 866
 867                 q->queue_type = asic->hw_queues_props[i].type;
 868                 q->supports_sync_stream =
 869                                 asic->hw_queues_props[i].supports_sync_stream;
 870                 rc = queue_init(hdev, q, i);
 871                 if (rc) {
 872                         dev_err(hdev->dev,
 873                                 "failed to initialize queue %d\n", i);
 874                         goto release_queues;
 875                 }
 876         }
 877
 878         return 0;
 879
 880 release_queues:
 881         for (i = 0, q = hdev->kernel_queues ; i < q_ready_cnt ; i++, q++)
 882                 queue_fini(hdev, q);
 883
 884         kfree(hdev->kernel_queues);
 885
 886         return rc;
 887 }
 888
 889 void hl_hw_queues_destroy(struct hl_device *hdev)
 890 {
 891         struct hl_hw_queue *q;
 892         u32 max_queues = hdev->asic_prop.max_queues;
 893         int i;
 894
 895         for (i = 0, q = hdev->kernel_queues ; i < max_queues ; i++, q++)
 896                 queue_fini(hdev, q);
 897
 898         kfree(hdev->kernel_queues);
 899 }
 900
 901 void hl_hw_queue_reset(struct hl_device *hdev, bool hard_reset)
 902 {
 903         struct hl_hw_queue *q;
 904         u32 max_queues = hdev->asic_prop.max_queues;
 905         int i;
 906
 907         for (i = 0, q = hdev->kernel_queues ; i < max_queues ; i++, q++) {
 908                 if ((!q->valid) ||
 909                         ((!hard_reset) && (q->queue_type == QUEUE_TYPE_CPU)))
 910                         continue;
 911                 q->pi = 0;
 912                 atomic_set(&q->ci, 0);
 913
 914                 if (q->supports_sync_stream)
 915                         sync_stream_queue_reset(hdev, q->hw_queue_id);
 916         }
 917 }