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_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
  52         /* There are no internal queues if H/W queues are being used */
  53         if (!hdev->asic_prop.max_queues || q->queue_type == QUEUE_TYPE_HW)
  54                 return;
  55
  56         /* We must increment CI for every queue that will never get a
  57          * completion, there are 2 scenarios this can happen:
  58          * 1. All queues of a non completion CS will never get a completion.
  59          * 2. Internal queues never gets completion.
  60          */
  61         for (i = 0 ; i < hdev->asic_prop.max_queues ; i++, q++) {
  62                 if (!cs_needs_completion(cs) || q->queue_type == QUEUE_TYPE_INT)
  63                         atomic_add(cs->jobs_in_queue_cnt[i], &q->ci);
  64         }
  65 }
  66
  67 /*
  68  * hl_hw_queue_submit_bd() - Submit a buffer descriptor to an external or a
  69  *                                H/W queue.
  70  * @hdev: pointer to habanalabs device structure
  71  * @q: pointer to habanalabs queue structure
  72  * @ctl: BD's control word
  73  * @len: BD's length
  74  * @ptr: BD's pointer
  75  *
  76  * This function assumes there is enough space on the queue to submit a new
  77  * BD to it. It initializes the next BD and calls the device specific
  78  * function to set the pi (and doorbell)
  79  *
  80  * This function must be called when the scheduler mutex is taken
  81  *
  82  */
  83 void hl_hw_queue_submit_bd(struct hl_device *hdev, struct hl_hw_queue *q,
  84                 u32 ctl, u32 len, u64 ptr)
  85 {
  86         struct hl_bd *bd;
  87
  88         bd = q->kernel_address;
  89         bd += hl_pi_2_offset(q->pi);
  90         bd->ctl = cpu_to_le32(ctl);
  91         bd->len = cpu_to_le32(len);
  92         bd->ptr = cpu_to_le64(ptr);
  93
  94         q->pi = hl_queue_inc_ptr(q->pi);
  95         hdev->asic_funcs->ring_doorbell(hdev, q->hw_queue_id, q->pi);
  96 }
  97
  98 /*
  99  * ext_queue_sanity_checks - perform some sanity checks on external queue
 100  *
 101  * @hdev              : pointer to hl_device structure
 102  * @q                 : pointer to hl_hw_queue structure
 103  * @num_of_entries    : how many entries to check for space
 104  * @reserve_cq_entry  : whether to reserve an entry in the cq
 105  *
 106  * H/W queues spinlock should be taken before calling this function
 107  *
 108  * Perform the following:
 109  * - Make sure we have enough space in the h/w queue
 110  * - Make sure we have enough space in the completion queue
 111  * - Reserve space in the completion queue (needs to be reversed if there
 112  *   is a failure down the road before the actual submission of work). Only
 113  *   do this action if reserve_cq_entry is true
 114  *
 115  */
 116 static int ext_queue_sanity_checks(struct hl_device *hdev,
 117                                 struct hl_hw_queue *q, int num_of_entries,
 118                                 bool reserve_cq_entry)
 119 {
 120         atomic_t *free_slots =
 121                         &hdev->completion_queue[q->cq_id].free_slots_cnt;
 122         int free_slots_cnt;
 123
 124         /* Check we have enough space in the queue */
 125         free_slots_cnt = queue_free_slots(q, HL_QUEUE_LENGTH);
 126
 127         if (free_slots_cnt < num_of_entries) {
 128                 dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
 129                         q->hw_queue_id, num_of_entries);
 130                 return -EAGAIN;
 131         }
 132
 133         if (reserve_cq_entry) {
 134                 /*
 135                  * Check we have enough space in the completion queue
 136                  * Add -1 to counter (decrement) unless counter was already 0
 137                  * In that case, CQ is full so we can't submit a new CB because
 138                  * we won't get ack on its completion
 139                  * atomic_add_unless will return 0 if counter was already 0
 140                  */
 141                 if (atomic_add_negative(num_of_entries * -1, free_slots)) {
 142                         dev_dbg(hdev->dev, "No space for %d on CQ %d\n",
 143                                 num_of_entries, q->hw_queue_id);
 144                         atomic_add(num_of_entries, free_slots);
 145                         return -EAGAIN;
 146                 }
 147         }
 148
 149         return 0;
 150 }
 151
 152 /*
 153  * int_queue_sanity_checks - perform some sanity checks on internal queue
 154  *
 155  * @hdev              : pointer to hl_device structure
 156  * @q                 : pointer to hl_hw_queue structure
 157  * @num_of_entries    : how many entries to check for space
 158  *
 159  * H/W queues spinlock should be taken before calling this function
 160  *
 161  * Perform the following:
 162  * - Make sure we have enough space in the h/w queue
 163  *
 164  */
 165 static int int_queue_sanity_checks(struct hl_device *hdev,
 166                                         struct hl_hw_queue *q,
 167                                         int num_of_entries)
 168 {
 169         int free_slots_cnt;
 170
 171         if (num_of_entries > q->int_queue_len) {
 172                 dev_err(hdev->dev,
 173                         "Cannot populate queue %u with %u jobs\n",
 174                         q->hw_queue_id, num_of_entries);
 175                 return -ENOMEM;
 176         }
 177
 178         /* Check we have enough space in the queue */
 179         free_slots_cnt = queue_free_slots(q, q->int_queue_len);
 180
 181         if (free_slots_cnt < num_of_entries) {
 182                 dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
 183                         q->hw_queue_id, num_of_entries);
 184                 return -EAGAIN;
 185         }
 186
 187         return 0;
 188 }
 189
 190 /*
 191  * hw_queue_sanity_checks() - Make sure we have enough space in the h/w queue
 192  * @hdev: Pointer to hl_device structure.
 193  * @q: Pointer to hl_hw_queue structure.
 194  * @num_of_entries: How many entries to check for space.
 195  *
 196  * Notice: We do not reserve queue entries so this function mustn't be called
 197  *         more than once per CS for the same queue
 198  *
 199  */
 200 static int hw_queue_sanity_checks(struct hl_device *hdev, struct hl_hw_queue *q,
 201                                         int num_of_entries)
 202 {
 203         int free_slots_cnt;
 204
 205         /* Check we have enough space in the queue */
 206         free_slots_cnt = queue_free_slots(q, HL_QUEUE_LENGTH);
 207
 208         if (free_slots_cnt < num_of_entries) {
 209                 dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
 210                         q->hw_queue_id, num_of_entries);
 211                 return -EAGAIN;
 212         }
 213
 214         return 0;
 215 }
 216
 217 /*
 218  * hl_hw_queue_send_cb_no_cmpl - send a single CB (not a JOB) without completion
 219  *
 220  * @hdev: pointer to hl_device structure
 221  * @hw_queue_id: Queue's type
 222  * @cb_size: size of CB
 223  * @cb_ptr: pointer to CB location
 224  *
 225  * This function sends a single CB, that must NOT generate a completion entry.
 226  * Sending CPU messages can be done instead via 'hl_hw_queue_submit_bd()'
 227  */
 228 int hl_hw_queue_send_cb_no_cmpl(struct hl_device *hdev, u32 hw_queue_id,
 229                                 u32 cb_size, u64 cb_ptr)
 230 {
 231         struct hl_hw_queue *q = &hdev->kernel_queues[hw_queue_id];
 232         int rc = 0;
 233
 234         hdev->asic_funcs->hw_queues_lock(hdev);
 235
 236         if (hdev->disabled) {
 237                 rc = -EPERM;
 238                 goto out;
 239         }
 240
 241         /*
 242          * hl_hw_queue_send_cb_no_cmpl() is called for queues of a H/W queue
 243          * type only on init phase, when the queues are empty and being tested,
 244          * so there is no need for sanity checks.
 245          */
 246         if (q->queue_type != QUEUE_TYPE_HW) {
 247                 rc = ext_queue_sanity_checks(hdev, q, 1, false);
 248                 if (rc)
 249                         goto out;
 250         }
 251
 252         hl_hw_queue_submit_bd(hdev, q, 0, cb_size, cb_ptr);
 253
 254 out:
 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         /* Skip completion flow in case this is a non completion CS */
 291         if (!cs_needs_completion(job->cs))
 292                 goto submit_bd;
 293
 294         cq_pkt.data = cpu_to_le32(
 295                         ((q->pi << CQ_ENTRY_SHADOW_INDEX_SHIFT)
 296                                 & CQ_ENTRY_SHADOW_INDEX_MASK) |
 297                         FIELD_PREP(CQ_ENTRY_SHADOW_INDEX_VALID_MASK, 1) |
 298                         FIELD_PREP(CQ_ENTRY_READY_MASK, 1));
 299
 300         /*
 301          * No need to protect pi_offset because scheduling to the
 302          * H/W queues is done under the scheduler mutex
 303          *
 304          * No need to check if CQ is full because it was already
 305          * checked in ext_queue_sanity_checks
 306          */
 307         cq = &hdev->completion_queue[q->cq_id];
 308         cq_addr = cq->bus_address + cq->pi * sizeof(struct hl_cq_entry);
 309
 310         hdev->asic_funcs->add_end_of_cb_packets(hdev, cb->kernel_address, len,
 311                                                 cq_addr,
 312                                                 le32_to_cpu(cq_pkt.data),
 313                                                 q->msi_vec,
 314                                                 job->contains_dma_pkt);
 315
 316         q->shadow_queue[hl_pi_2_offset(q->pi)] = job;
 317
 318         cq->pi = hl_cq_inc_ptr(cq->pi);
 319
 320 submit_bd:
 321         hl_hw_queue_submit_bd(hdev, q, ctl, len, ptr);
 322 }
 323
 324 /*
 325  * int_queue_schedule_job - submit a JOB to an internal queue
 326  *
 327  * @job: pointer to the job that needs to be submitted to the queue
 328  *
 329  * This function must be called when the scheduler mutex is taken
 330  *
 331  */
 332 static void int_queue_schedule_job(struct hl_cs_job *job)
 333 {
 334         struct hl_device *hdev = job->cs->ctx->hdev;
 335         struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
 336         struct hl_bd bd;
 337         __le64 *pi;
 338
 339         bd.ctl = 0;
 340         bd.len = cpu_to_le32(job->job_cb_size);
 341
 342         if (job->is_kernel_allocated_cb)
 343                 /* bus_address is actually a mmu mapped address
 344                  * allocated from an internal pool
 345                  */
 346                 bd.ptr = cpu_to_le64(job->user_cb->bus_address);
 347         else
 348                 bd.ptr = cpu_to_le64((u64) (uintptr_t) job->user_cb);
 349
 350         pi = q->kernel_address + (q->pi & (q->int_queue_len - 1)) * sizeof(bd);
 351
 352         q->pi++;
 353         q->pi &= ((q->int_queue_len << 1) - 1);
 354
 355         hdev->asic_funcs->pqe_write(hdev, pi, &bd);
 356
 357         hdev->asic_funcs->ring_doorbell(hdev, q->hw_queue_id, q->pi);
 358 }
 359
 360 /*
 361  * hw_queue_schedule_job - submit a JOB to a H/W queue
 362  *
 363  * @job: pointer to the job that needs to be submitted to the queue
 364  *
 365  * This function must be called when the scheduler mutex is taken
 366  *
 367  */
 368 static void hw_queue_schedule_job(struct hl_cs_job *job)
 369 {
 370         struct hl_device *hdev = job->cs->ctx->hdev;
 371         struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
 372         u64 ptr;
 373         u32 offset, ctl, len;
 374
 375         /*
 376          * Upon PQE completion, COMP_DATA is used as the write data to the
 377          * completion queue (QMAN HBW message), and COMP_OFFSET is used as the
 378          * write address offset in the SM block (QMAN LBW message).
 379          * The write address offset is calculated as "COMP_OFFSET << 2".
 380          */
 381         offset = job->cs->sequence & (hdev->asic_prop.max_pending_cs - 1);
 382         ctl = ((offset << BD_CTL_COMP_OFFSET_SHIFT) & BD_CTL_COMP_OFFSET_MASK) |
 383                 ((q->pi << BD_CTL_COMP_DATA_SHIFT) & BD_CTL_COMP_DATA_MASK);
 384
 385         len = job->job_cb_size;
 386
 387         /*
 388          * A patched CB is created only if a user CB was allocated by driver and
 389          * MMU is disabled. If MMU is enabled, the user CB should be used
 390          * instead. If the user CB wasn't allocated by driver, assume that it
 391          * holds an address.
 392          */
 393         if (job->patched_cb)
 394                 ptr = job->patched_cb->bus_address;
 395         else if (job->is_kernel_allocated_cb)
 396                 ptr = job->user_cb->bus_address;
 397         else
 398                 ptr = (u64) (uintptr_t) job->user_cb;
 399
 400         hl_hw_queue_submit_bd(hdev, q, ctl, len, ptr);
 401 }
 402
 403 static int init_signal_cs(struct hl_device *hdev,
 404                 struct hl_cs_job *job, struct hl_cs_compl *cs_cmpl)
 405 {
 406         struct hl_sync_stream_properties *prop;
 407         struct hl_hw_sob *hw_sob;
 408         u32 q_idx;
 409         int rc = 0;
 410
 411         q_idx = job->hw_queue_id;
 412         prop = &hdev->kernel_queues[q_idx].sync_stream_prop;
 413         hw_sob = &prop->hw_sob[prop->curr_sob_offset];
 414
 415         cs_cmpl->hw_sob = hw_sob;
 416         cs_cmpl->sob_val = prop->next_sob_val;
 417
 418         dev_dbg(hdev->dev,
 419                 "generate signal CB, sob_id: %d, sob val: 0x%x, q_idx: %d\n",
 420                 cs_cmpl->hw_sob->sob_id, cs_cmpl->sob_val, q_idx);
 421
 422         /* we set an EB since we must make sure all oeprations are done
 423          * when sending the signal
 424          */
 425         hdev->asic_funcs->gen_signal_cb(hdev, job->patched_cb,
 426                                 cs_cmpl->hw_sob->sob_id, 0, true);
 427
 428         rc = hl_cs_signal_sob_wraparound_handler(hdev, q_idx, &hw_sob, 1);
 429
 430         return rc;
 431 }
 432
 433 static void init_wait_cs(struct hl_device *hdev, struct hl_cs *cs,
 434                 struct hl_cs_job *job, struct hl_cs_compl *cs_cmpl)
 435 {
 436         struct hl_cs_compl *signal_cs_cmpl;
 437         struct hl_sync_stream_properties *prop;
 438         struct hl_gen_wait_properties wait_prop;
 439         u32 q_idx;
 440
 441         q_idx = job->hw_queue_id;
 442         prop = &hdev->kernel_queues[q_idx].sync_stream_prop;
 443
 444         signal_cs_cmpl = container_of(cs->signal_fence,
 445                                         struct hl_cs_compl,
 446                                         base_fence);
 447
 448         /* copy the SOB id and value of the signal CS */
 449         cs_cmpl->hw_sob = signal_cs_cmpl->hw_sob;
 450         cs_cmpl->sob_val = signal_cs_cmpl->sob_val;
 451
 452         dev_dbg(hdev->dev,
 453                 "generate wait CB, sob_id: %d, sob_val: 0x%x, mon_id: %d, q_idx: %d\n",
 454                 cs_cmpl->hw_sob->sob_id, cs_cmpl->sob_val,
 455                 prop->base_mon_id, q_idx);
 456
 457         wait_prop.data = (void *) job->patched_cb;
 458         wait_prop.sob_base = cs_cmpl->hw_sob->sob_id;
 459         wait_prop.sob_mask = 0x1;
 460         wait_prop.sob_val = cs_cmpl->sob_val;
 461         wait_prop.mon_id = prop->base_mon_id;
 462         wait_prop.q_idx = q_idx;
 463         wait_prop.size = 0;
 464         hdev->asic_funcs->gen_wait_cb(hdev, &wait_prop);
 465
 466         kref_get(&cs_cmpl->hw_sob->kref);
 467         /*
 468          * Must put the signal fence after the SOB refcnt increment so
 469          * the SOB refcnt won't turn 0 and reset the SOB before the
 470          * wait CS was submitted.
 471          */
 472         mb();
 473         hl_fence_put(cs->signal_fence);
 474         cs->signal_fence = NULL;
 475 }
 476
 477 /*
 478  * init_signal_wait_cs - initialize a signal/wait CS
 479  * @cs: pointer to the signal/wait CS
 480  *
 481  * H/W queues spinlock should be taken before calling this function
 482  */
 483 static int init_signal_wait_cs(struct hl_cs *cs)
 484 {
 485         struct hl_ctx *ctx = cs->ctx;
 486         struct hl_device *hdev = ctx->hdev;
 487         struct hl_cs_job *job;
 488         struct hl_cs_compl *cs_cmpl =
 489                         container_of(cs->fence, struct hl_cs_compl, base_fence);
 490         int rc = 0;
 491
 492         /* There is only one job in a signal/wait CS */
 493         job = list_first_entry(&cs->job_list, struct hl_cs_job,
 494                                 cs_node);
 495
 496         if (cs->type & CS_TYPE_SIGNAL)
 497                 rc = init_signal_cs(hdev, job, cs_cmpl);
 498         else if (cs->type & CS_TYPE_WAIT)
 499                 init_wait_cs(hdev, cs, job, cs_cmpl);
 500
 501         return rc;
 502 }
 503
 504 /*
 505  * hl_hw_queue_schedule_cs - schedule a command submission
 506  * @cs: pointer to the CS
 507  */
 508 int hl_hw_queue_schedule_cs(struct hl_cs *cs)
 509 {
 510         enum hl_device_status status;
 511         struct hl_cs_counters_atomic *cntr;
 512         struct hl_ctx *ctx = cs->ctx;
 513         struct hl_device *hdev = ctx->hdev;
 514         struct hl_cs_job *job, *tmp;
 515         struct hl_hw_queue *q;
 516         int rc = 0, i, cq_cnt;
 517         bool first_entry;
 518         u32 max_queues;
 519
 520         cntr = &hdev->aggregated_cs_counters;
 521
 522         hdev->asic_funcs->hw_queues_lock(hdev);
 523
 524         if (!hl_device_operational(hdev, &status)) {
 525                 atomic64_inc(&cntr->device_in_reset_drop_cnt);
 526                 atomic64_inc(&ctx->cs_counters.device_in_reset_drop_cnt);
 527                 dev_err(hdev->dev,
 528                         "device is %s, CS rejected!\n", hdev->status[status]);
 529                 rc = -EPERM;
 530                 goto out;
 531         }
 532
 533         max_queues = hdev->asic_prop.max_queues;
 534
 535         q = &hdev->kernel_queues[0];
 536         for (i = 0, cq_cnt = 0 ; i < max_queues ; i++, q++) {
 537                 if (cs->jobs_in_queue_cnt[i]) {
 538                         switch (q->queue_type) {
 539                         case QUEUE_TYPE_EXT:
 540                                 rc = ext_queue_sanity_checks(hdev, q,
 541                                                 cs->jobs_in_queue_cnt[i],
 542                                                 cs_needs_completion(cs) ?
 543                                                                 true : false);
 544                                 break;
 545                         case QUEUE_TYPE_INT:
 546                                 rc = int_queue_sanity_checks(hdev, q,
 547                                                 cs->jobs_in_queue_cnt[i]);
 548                                 break;
 549                         case QUEUE_TYPE_HW:
 550                                 rc = hw_queue_sanity_checks(hdev, q,
 551                                                 cs->jobs_in_queue_cnt[i]);
 552                                 break;
 553                         default:
 554                                 dev_err(hdev->dev, "Queue type %d is invalid\n",
 555                                         q->queue_type);
 556                                 rc = -EINVAL;
 557                                 break;
 558                         }
 559
 560                         if (rc) {
 561                                 atomic64_inc(
 562                                         &ctx->cs_counters.queue_full_drop_cnt);
 563                                 atomic64_inc(&cntr->queue_full_drop_cnt);
 564                                 goto unroll_cq_resv;
 565                         }
 566
 567                         if (q->queue_type == QUEUE_TYPE_EXT)
 568                                 cq_cnt++;
 569                 }
 570         }
 571
 572         if ((cs->type == CS_TYPE_SIGNAL) || (cs->type == CS_TYPE_WAIT)) {
 573                 rc = init_signal_wait_cs(cs);
 574                 if (rc) {
 575                         dev_err(hdev->dev, "Failed to submit signal cs\n");
 576                         goto unroll_cq_resv;
 577                 }
 578         } else if (cs->type == CS_TYPE_COLLECTIVE_WAIT)
 579                 hdev->asic_funcs->collective_wait_init_cs(cs);
 580
 581
 582         spin_lock(&hdev->cs_mirror_lock);
 583
 584         /* Verify staged CS exists and add to the staged list */
 585         if (cs->staged_cs && !cs->staged_first) {
 586                 struct hl_cs *staged_cs;
 587
 588                 staged_cs = hl_staged_cs_find_first(hdev, cs->staged_sequence);
 589                 if (!staged_cs) {
 590                         dev_err(hdev->dev,
 591                                 "Cannot find staged submission sequence %llu",
 592                                 cs->staged_sequence);
 593                         rc = -EINVAL;
 594                         goto unlock_cs_mirror;
 595                 }
 596
 597                 if (is_staged_cs_last_exists(hdev, staged_cs)) {
 598                         dev_err(hdev->dev,
 599                                 "Staged submission sequence %llu already submitted",
 600                                 cs->staged_sequence);
 601                         rc = -EINVAL;
 602                         goto unlock_cs_mirror;
 603                 }
 604
 605                 list_add_tail(&cs->staged_cs_node, &staged_cs->staged_cs_node);
 606
 607                 /* update stream map of the first CS */
 608                 if (hdev->supports_wait_for_multi_cs)
 609                         staged_cs->fence->stream_map |= cs->fence->stream_map;
 610         }
 611
 612         list_add_tail(&cs->mirror_node, &hdev->cs_mirror_list);
 613
 614         /* Queue TDR if the CS is the first entry and if timeout is wanted */
 615         first_entry = list_first_entry(&hdev->cs_mirror_list,
 616                                         struct hl_cs, mirror_node) == cs;
 617         if ((hdev->timeout_jiffies != MAX_SCHEDULE_TIMEOUT) &&
 618                                 first_entry && cs_needs_timeout(cs)) {
 619                 cs->tdr_active = true;
 620                 schedule_delayed_work(&cs->work_tdr, cs->timeout_jiffies);
 621
 622         }
 623
 624         spin_unlock(&hdev->cs_mirror_lock);
 625
 626         list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
 627                 switch (job->queue_type) {
 628                 case QUEUE_TYPE_EXT:
 629                         ext_queue_schedule_job(job);
 630                         break;
 631                 case QUEUE_TYPE_INT:
 632                         int_queue_schedule_job(job);
 633                         break;
 634                 case QUEUE_TYPE_HW:
 635                         hw_queue_schedule_job(job);
 636                         break;
 637                 default:
 638                         break;
 639                 }
 640
 641         cs->submitted = true;
 642
 643         goto out;
 644
 645 unlock_cs_mirror:
 646         spin_unlock(&hdev->cs_mirror_lock);
 647 unroll_cq_resv:
 648         q = &hdev->kernel_queues[0];
 649         for (i = 0 ; (i < max_queues) && (cq_cnt > 0) ; i++, q++) {
 650                 if ((q->queue_type == QUEUE_TYPE_EXT) &&
 651                                                 (cs->jobs_in_queue_cnt[i])) {
 652                         atomic_t *free_slots =
 653                                 &hdev->completion_queue[i].free_slots_cnt;
 654                         atomic_add(cs->jobs_in_queue_cnt[i], free_slots);
 655                         cq_cnt--;
 656                 }
 657         }
 658
 659 out:
 660         hdev->asic_funcs->hw_queues_unlock(hdev);
 661
 662         return rc;
 663 }
 664
 665 /*
 666  * hl_hw_queue_inc_ci_kernel - increment ci for kernel's queue
 667  *
 668  * @hdev: pointer to hl_device structure
 669  * @hw_queue_id: which queue to increment its ci
 670  */
 671 void hl_hw_queue_inc_ci_kernel(struct hl_device *hdev, u32 hw_queue_id)
 672 {
 673         struct hl_hw_queue *q = &hdev->kernel_queues[hw_queue_id];
 674
 675         atomic_inc(&q->ci);
 676 }
 677
 678 static int ext_and_cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
 679                                         bool is_cpu_queue)
 680 {
 681         void *p;
 682         int rc;
 683
 684         if (is_cpu_queue)
 685                 p = hdev->asic_funcs->cpu_accessible_dma_pool_alloc(hdev,
 686                                                         HL_QUEUE_SIZE_IN_BYTES,
 687                                                         &q->bus_address);
 688         else
 689                 p = hdev->asic_funcs->asic_dma_alloc_coherent(hdev,
 690                                                 HL_QUEUE_SIZE_IN_BYTES,
 691                                                 &q->bus_address,
 692                                                 GFP_KERNEL | __GFP_ZERO);
 693         if (!p)
 694                 return -ENOMEM;
 695
 696         q->kernel_address = p;
 697
 698         q->shadow_queue = kmalloc_array(HL_QUEUE_LENGTH,
 699                                         sizeof(*q->shadow_queue),
 700                                         GFP_KERNEL);
 701         if (!q->shadow_queue) {
 702                 dev_err(hdev->dev,
 703                         "Failed to allocate shadow queue for H/W queue %d\n",
 704                         q->hw_queue_id);
 705                 rc = -ENOMEM;
 706                 goto free_queue;
 707         }
 708
 709         /* Make sure read/write pointers are initialized to start of queue */
 710         atomic_set(&q->ci, 0);
 711         q->pi = 0;
 712
 713         return 0;
 714
 715 free_queue:
 716         if (is_cpu_queue)
 717                 hdev->asic_funcs->cpu_accessible_dma_pool_free(hdev,
 718                                         HL_QUEUE_SIZE_IN_BYTES,
 719                                         q->kernel_address);
 720         else
 721                 hdev->asic_funcs->asic_dma_free_coherent(hdev,
 722                                         HL_QUEUE_SIZE_IN_BYTES,
 723                                         q->kernel_address,
 724                                         q->bus_address);
 725
 726         return rc;
 727 }
 728
 729 static int int_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 730 {
 731         void *p;
 732
 733         p = hdev->asic_funcs->get_int_queue_base(hdev, q->hw_queue_id,
 734                                         &q->bus_address, &q->int_queue_len);
 735         if (!p) {
 736                 dev_err(hdev->dev,
 737                         "Failed to get base address for internal queue %d\n",
 738                         q->hw_queue_id);
 739                 return -EFAULT;
 740         }
 741
 742         q->kernel_address = p;
 743         q->pi = 0;
 744         atomic_set(&q->ci, 0);
 745
 746         return 0;
 747 }
 748
 749 static int cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 750 {
 751         return ext_and_cpu_queue_init(hdev, q, true);
 752 }
 753
 754 static int ext_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 755 {
 756         return ext_and_cpu_queue_init(hdev, q, false);
 757 }
 758
 759 static int hw_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 760 {
 761         void *p;
 762
 763         p = hdev->asic_funcs->asic_dma_alloc_coherent(hdev,
 764                                                 HL_QUEUE_SIZE_IN_BYTES,
 765                                                 &q->bus_address,
 766                                                 GFP_KERNEL | __GFP_ZERO);
 767         if (!p)
 768                 return -ENOMEM;
 769
 770         q->kernel_address = p;
 771
 772         /* Make sure read/write pointers are initialized to start of queue */
 773         atomic_set(&q->ci, 0);
 774         q->pi = 0;
 775
 776         return 0;
 777 }
 778
 779 static void sync_stream_queue_init(struct hl_device *hdev, u32 q_idx)
 780 {
 781         struct hl_sync_stream_properties *sync_stream_prop;
 782         struct asic_fixed_properties *prop = &hdev->asic_prop;
 783         struct hl_hw_sob *hw_sob;
 784         int sob, reserved_mon_idx, queue_idx;
 785
 786         sync_stream_prop = &hdev->kernel_queues[q_idx].sync_stream_prop;
 787
 788         /* We use 'collective_mon_idx' as a running index in order to reserve
 789          * monitors for collective master/slave queues.
 790          * collective master queue gets 2 reserved monitors
 791          * collective slave queue gets 1 reserved monitor
 792          */
 793         if (hdev->kernel_queues[q_idx].collective_mode ==
 794                         HL_COLLECTIVE_MASTER) {
 795                 reserved_mon_idx = hdev->collective_mon_idx;
 796
 797                 /* reserve the first monitor for collective master queue */
 798                 sync_stream_prop->collective_mstr_mon_id[0] =
 799                         prop->collective_first_mon + reserved_mon_idx;
 800
 801                 /* reserve the second monitor for collective master queue */
 802                 sync_stream_prop->collective_mstr_mon_id[1] =
 803                         prop->collective_first_mon + reserved_mon_idx + 1;
 804
 805                 hdev->collective_mon_idx += HL_COLLECTIVE_RSVD_MSTR_MONS;
 806         } else if (hdev->kernel_queues[q_idx].collective_mode ==
 807                         HL_COLLECTIVE_SLAVE) {
 808                 reserved_mon_idx = hdev->collective_mon_idx++;
 809
 810                 /* reserve a monitor for collective slave queue */
 811                 sync_stream_prop->collective_slave_mon_id =
 812                         prop->collective_first_mon + reserved_mon_idx;
 813         }
 814
 815         if (!hdev->kernel_queues[q_idx].supports_sync_stream)
 816                 return;
 817
 818         queue_idx = hdev->sync_stream_queue_idx++;
 819
 820         sync_stream_prop->base_sob_id = prop->sync_stream_first_sob +
 821                         (queue_idx * HL_RSVD_SOBS);
 822         sync_stream_prop->base_mon_id = prop->sync_stream_first_mon +
 823                         (queue_idx * HL_RSVD_MONS);
 824         sync_stream_prop->next_sob_val = 1;
 825         sync_stream_prop->curr_sob_offset = 0;
 826
 827         for (sob = 0 ; sob < HL_RSVD_SOBS ; sob++) {
 828                 hw_sob = &sync_stream_prop->hw_sob[sob];
 829                 hw_sob->hdev = hdev;
 830                 hw_sob->sob_id = sync_stream_prop->base_sob_id + sob;
 831                 hw_sob->q_idx = q_idx;
 832                 kref_init(&hw_sob->kref);
 833         }
 834 }
 835
 836 static void sync_stream_queue_reset(struct hl_device *hdev, u32 q_idx)
 837 {
 838         struct hl_sync_stream_properties *prop =
 839                         &hdev->kernel_queues[q_idx].sync_stream_prop;
 840
 841         /*
 842          * In case we got here due to a stuck CS, the refcnt might be bigger
 843          * than 1 and therefore we reset it.
 844          */
 845         kref_init(&prop->hw_sob[prop->curr_sob_offset].kref);
 846         prop->curr_sob_offset = 0;
 847         prop->next_sob_val = 1;
 848 }
 849
 850 /*
 851  * queue_init - main initialization function for H/W queue object
 852  *
 853  * @hdev: pointer to hl_device device structure
 854  * @q: pointer to hl_hw_queue queue structure
 855  * @hw_queue_id: The id of the H/W queue
 856  *
 857  * Allocate dma-able memory for the queue and initialize fields
 858  * Returns 0 on success
 859  */
 860 static int queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
 861                         u32 hw_queue_id)
 862 {
 863         int rc;
 864
 865         q->hw_queue_id = hw_queue_id;
 866
 867         switch (q->queue_type) {
 868         case QUEUE_TYPE_EXT:
 869                 rc = ext_queue_init(hdev, q);
 870                 break;
 871         case QUEUE_TYPE_INT:
 872                 rc = int_queue_init(hdev, q);
 873                 break;
 874         case QUEUE_TYPE_CPU:
 875                 rc = cpu_queue_init(hdev, q);
 876                 break;
 877         case QUEUE_TYPE_HW:
 878                 rc = hw_queue_init(hdev, q);
 879                 break;
 880         case QUEUE_TYPE_NA:
 881                 q->valid = 0;
 882                 return 0;
 883         default:
 884                 dev_crit(hdev->dev, "wrong queue type %d during init\n",
 885                         q->queue_type);
 886                 rc = -EINVAL;
 887                 break;
 888         }
 889
 890         sync_stream_queue_init(hdev, q->hw_queue_id);
 891
 892         if (rc)
 893                 return rc;
 894
 895         q->valid = 1;
 896
 897         return 0;
 898 }
 899
 900 /*
 901  * hw_queue_fini - destroy queue
 902  *
 903  * @hdev: pointer to hl_device device structure
 904  * @q: pointer to hl_hw_queue queue structure
 905  *
 906  * Free the queue memory
 907  */
 908 static void queue_fini(struct hl_device *hdev, struct hl_hw_queue *q)
 909 {
 910         if (!q->valid)
 911                 return;
 912
 913         /*
 914          * If we arrived here, there are no jobs waiting on this queue
 915          * so we can safely remove it.
 916          * This is because this function can only called when:
 917          * 1. Either a context is deleted, which only can occur if all its
 918          *    jobs were finished
 919          * 2. A context wasn't able to be created due to failure or timeout,
 920          *    which means there are no jobs on the queue yet
 921          *
 922          * The only exception are the queues of the kernel context, but
 923          * if they are being destroyed, it means that the entire module is
 924          * being removed. If the module is removed, it means there is no open
 925          * user context. It also means that if a job was submitted by
 926          * the kernel driver (e.g. context creation), the job itself was
 927          * released by the kernel driver when a timeout occurred on its
 928          * Completion. Thus, we don't need to release it again.
 929          */
 930
 931         if (q->queue_type == QUEUE_TYPE_INT)
 932                 return;
 933
 934         kfree(q->shadow_queue);
 935
 936         if (q->queue_type == QUEUE_TYPE_CPU)
 937                 hdev->asic_funcs->cpu_accessible_dma_pool_free(hdev,
 938                                         HL_QUEUE_SIZE_IN_BYTES,
 939                                         q->kernel_address);
 940         else
 941                 hdev->asic_funcs->asic_dma_free_coherent(hdev,
 942                                         HL_QUEUE_SIZE_IN_BYTES,
 943                                         q->kernel_address,
 944                                         q->bus_address);
 945 }
 946
 947 int hl_hw_queues_create(struct hl_device *hdev)
 948 {
 949         struct asic_fixed_properties *asic = &hdev->asic_prop;
 950         struct hl_hw_queue *q;
 951         int i, rc, q_ready_cnt;
 952
 953         hdev->kernel_queues = kcalloc(asic->max_queues,
 954                                 sizeof(*hdev->kernel_queues), GFP_KERNEL);
 955
 956         if (!hdev->kernel_queues) {
 957                 dev_err(hdev->dev, "Not enough memory for H/W queues\n");
 958                 return -ENOMEM;
 959         }
 960
 961         /* Initialize the H/W queues */
 962         for (i = 0, q_ready_cnt = 0, q = hdev->kernel_queues;
 963                         i < asic->max_queues ; i++, q_ready_cnt++, q++) {
 964
 965                 q->queue_type = asic->hw_queues_props[i].type;
 966                 q->supports_sync_stream =
 967                                 asic->hw_queues_props[i].supports_sync_stream;
 968                 q->collective_mode = asic->hw_queues_props[i].collective_mode;
 969                 rc = queue_init(hdev, q, i);
 970                 if (rc) {
 971                         dev_err(hdev->dev,
 972                                 "failed to initialize queue %d\n", i);
 973                         goto release_queues;
 974                 }
 975         }
 976
 977         return 0;
 978
 979 release_queues:
 980         for (i = 0, q = hdev->kernel_queues ; i < q_ready_cnt ; i++, q++)
 981                 queue_fini(hdev, q);
 982
 983         kfree(hdev->kernel_queues);
 984
 985         return rc;
 986 }
 987
 988 void hl_hw_queues_destroy(struct hl_device *hdev)
 989 {
 990         struct hl_hw_queue *q;
 991         u32 max_queues = hdev->asic_prop.max_queues;
 992         int i;
 993
 994         for (i = 0, q = hdev->kernel_queues ; i < max_queues ; i++, q++)
 995                 queue_fini(hdev, q);
 996
 997         kfree(hdev->kernel_queues);
 998 }
 999
1000 void hl_hw_queue_reset(struct hl_device *hdev, bool hard_reset)
1001 {
1002         struct hl_hw_queue *q;
1003         u32 max_queues = hdev->asic_prop.max_queues;
1004         int i;
1005
1006         for (i = 0, q = hdev->kernel_queues ; i < max_queues ; i++, q++) {
1007                 if ((!q->valid) ||
1008                         ((!hard_reset) && (q->queue_type == QUEUE_TYPE_CPU)))
1009                         continue;
1010                 q->pi = 0;
1011                 atomic_set(&q->ci, 0);
1012
1013                 if (q->supports_sync_stream)
1014                         sync_stream_queue_reset(hdev, q->hw_queue_id);
1015         }
1016 }