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: %u, q_idx: %d, seq: %llu\n",
 420                 cs_cmpl->hw_sob->sob_id, cs_cmpl->sob_val, q_idx,
 421                 cs_cmpl->cs_seq);
 422
 423         /* we set an EB since we must make sure all oeprations are done
 424          * when sending the signal
 425          */
 426         hdev->asic_funcs->gen_signal_cb(hdev, job->patched_cb,
 427                                 cs_cmpl->hw_sob->sob_id, 0, true);
 428
 429         rc = hl_cs_signal_sob_wraparound_handler(hdev, q_idx, &hw_sob, 1,
 430                                                                 false);
 431
 432         return rc;
 433 }
 434
 435 void hl_hw_queue_encaps_sig_set_sob_info(struct hl_device *hdev,
 436                         struct hl_cs *cs, struct hl_cs_job *job,
 437                         struct hl_cs_compl *cs_cmpl)
 438 {
 439         struct hl_cs_encaps_sig_handle *handle = cs->encaps_sig_hdl;
 440
 441         cs_cmpl->hw_sob = handle->hw_sob;
 442
 443         /* Note that encaps_sig_wait_offset was validated earlier in the flow
 444          * for offset value which exceeds the max reserved signal count.
 445          * always decrement 1 of the offset since when the user
 446          * set offset 1 for example he mean to wait only for the first
 447          * signal only, which will be pre_sob_val, and if he set offset 2
 448          * then the value required is (pre_sob_val + 1) and so on...
 449          */
 450         cs_cmpl->sob_val = handle->pre_sob_val +
 451                         (job->encaps_sig_wait_offset - 1);
 452 }
 453
 454 static int init_wait_cs(struct hl_device *hdev, struct hl_cs *cs,
 455                 struct hl_cs_job *job, struct hl_cs_compl *cs_cmpl)
 456 {
 457         struct hl_gen_wait_properties wait_prop;
 458         struct hl_sync_stream_properties *prop;
 459         struct hl_cs_compl *signal_cs_cmpl;
 460         u32 q_idx;
 461
 462         q_idx = job->hw_queue_id;
 463         prop = &hdev->kernel_queues[q_idx].sync_stream_prop;
 464
 465         signal_cs_cmpl = container_of(cs->signal_fence,
 466                                         struct hl_cs_compl,
 467                                         base_fence);
 468
 469         if (cs->encaps_signals) {
 470                 /* use the encaps signal handle stored earlier in the flow
 471                  * and set the SOB information from the encaps
 472                  * signals handle
 473                  */
 474                 hl_hw_queue_encaps_sig_set_sob_info(hdev, cs, job, cs_cmpl);
 475
 476                 dev_dbg(hdev->dev, "Wait for encaps signals handle, qidx(%u), CS sequence(%llu), sob val: 0x%x, offset: %u\n",
 477                                 cs->encaps_sig_hdl->q_idx,
 478                                 cs->encaps_sig_hdl->cs_seq,
 479                                 cs_cmpl->sob_val,
 480                                 job->encaps_sig_wait_offset);
 481         } else {
 482                 /* Copy the SOB id and value of the signal CS */
 483                 cs_cmpl->hw_sob = signal_cs_cmpl->hw_sob;
 484                 cs_cmpl->sob_val = signal_cs_cmpl->sob_val;
 485         }
 486
 487         /* check again if the signal cs already completed.
 488          * if yes then don't send any wait cs since the hw_sob
 489          * could be in reset already. if signal is not completed
 490          * then get refcount to hw_sob to prevent resetting the sob
 491          * while wait cs is not submitted.
 492          * note that this check is protected by two locks,
 493          * hw queue lock and completion object lock,
 494          * and the same completion object lock also protects
 495          * the hw_sob reset handler function.
 496          * The hw_queue lock prevent out of sync of hw_sob
 497          * refcount value, changed by signal/wait flows.
 498          */
 499         spin_lock(&signal_cs_cmpl->lock);
 500
 501         if (completion_done(&cs->signal_fence->completion)) {
 502                 spin_unlock(&signal_cs_cmpl->lock);
 503                 return -EINVAL;
 504         }
 505
 506         kref_get(&cs_cmpl->hw_sob->kref);
 507
 508         spin_unlock(&signal_cs_cmpl->lock);
 509
 510         dev_dbg(hdev->dev,
 511                 "generate wait CB, sob_id: %d, sob_val: 0x%x, mon_id: %d, q_idx: %d, seq: %llu\n",
 512                 cs_cmpl->hw_sob->sob_id, cs_cmpl->sob_val,
 513                 prop->base_mon_id, q_idx, cs->sequence);
 514
 515         wait_prop.data = (void *) job->patched_cb;
 516         wait_prop.sob_base = cs_cmpl->hw_sob->sob_id;
 517         wait_prop.sob_mask = 0x1;
 518         wait_prop.sob_val = cs_cmpl->sob_val;
 519         wait_prop.mon_id = prop->base_mon_id;
 520         wait_prop.q_idx = q_idx;
 521         wait_prop.size = 0;
 522
 523         hdev->asic_funcs->gen_wait_cb(hdev, &wait_prop);
 524
 525         mb();
 526         hl_fence_put(cs->signal_fence);
 527         cs->signal_fence = NULL;
 528
 529         return 0;
 530 }
 531
 532 /*
 533  * init_signal_wait_cs - initialize a signal/wait CS
 534  * @cs: pointer to the signal/wait CS
 535  *
 536  * H/W queues spinlock should be taken before calling this function
 537  */
 538 static int init_signal_wait_cs(struct hl_cs *cs)
 539 {
 540         struct hl_ctx *ctx = cs->ctx;
 541         struct hl_device *hdev = ctx->hdev;
 542         struct hl_cs_job *job;
 543         struct hl_cs_compl *cs_cmpl =
 544                         container_of(cs->fence, struct hl_cs_compl, base_fence);
 545         int rc = 0;
 546
 547         /* There is only one job in a signal/wait CS */
 548         job = list_first_entry(&cs->job_list, struct hl_cs_job,
 549                                 cs_node);
 550
 551         if (cs->type & CS_TYPE_SIGNAL)
 552                 rc = init_signal_cs(hdev, job, cs_cmpl);
 553         else if (cs->type & CS_TYPE_WAIT)
 554                 rc = init_wait_cs(hdev, cs, job, cs_cmpl);
 555
 556         return rc;
 557 }
 558
 559 static int encaps_sig_first_staged_cs_handler
 560                         (struct hl_device *hdev, struct hl_cs *cs)
 561 {
 562         struct hl_cs_compl *cs_cmpl =
 563                         container_of(cs->fence,
 564                                         struct hl_cs_compl, base_fence);
 565         struct hl_cs_encaps_sig_handle *encaps_sig_hdl;
 566         struct hl_encaps_signals_mgr *mgr;
 567         int rc = 0;
 568
 569         mgr = &hdev->compute_ctx->sig_mgr;
 570
 571         spin_lock(&mgr->lock);
 572         encaps_sig_hdl = idr_find(&mgr->handles, cs->encaps_sig_hdl_id);
 573         if (encaps_sig_hdl) {
 574                 /*
 575                  * Set handler CS sequence,
 576                  * the CS which contains the encapsulated signals.
 577                  */
 578                 encaps_sig_hdl->cs_seq = cs->sequence;
 579                 /* store the handle and set encaps signal indication,
 580                  * to be used later in cs_do_release to put the last
 581                  * reference to encaps signals handlers.
 582                  */
 583                 cs_cmpl->encaps_signals = true;
 584                 cs_cmpl->encaps_sig_hdl = encaps_sig_hdl;
 585
 586                 /* set hw_sob pointer in completion object
 587                  * since it's used in cs_do_release flow to put
 588                  * refcount to sob
 589                  */
 590                 cs_cmpl->hw_sob = encaps_sig_hdl->hw_sob;
 591                 cs_cmpl->sob_val = encaps_sig_hdl->pre_sob_val +
 592                                                 encaps_sig_hdl->count;
 593
 594                 dev_dbg(hdev->dev, "CS seq (%llu) added to encaps signal handler id (%u), count(%u), qidx(%u), sob(%u), val(%u)\n",
 595                                 cs->sequence, encaps_sig_hdl->id,
 596                                 encaps_sig_hdl->count,
 597                                 encaps_sig_hdl->q_idx,
 598                                 cs_cmpl->hw_sob->sob_id,
 599                                 cs_cmpl->sob_val);
 600
 601         } else {
 602                 dev_err(hdev->dev, "encaps handle id(%u) wasn't found!\n",
 603                                 cs->encaps_sig_hdl_id);
 604                 rc = -EINVAL;
 605         }
 606
 607         spin_unlock(&mgr->lock);
 608
 609         return rc;
 610 }
 611
 612 /*
 613  * hl_hw_queue_schedule_cs - schedule a command submission
 614  * @cs: pointer to the CS
 615  */
 616 int hl_hw_queue_schedule_cs(struct hl_cs *cs)
 617 {
 618         enum hl_device_status status;
 619         struct hl_cs_counters_atomic *cntr;
 620         struct hl_ctx *ctx = cs->ctx;
 621         struct hl_device *hdev = ctx->hdev;
 622         struct hl_cs_job *job, *tmp;
 623         struct hl_hw_queue *q;
 624         int rc = 0, i, cq_cnt;
 625         bool first_entry;
 626         u32 max_queues;
 627
 628         cntr = &hdev->aggregated_cs_counters;
 629
 630         hdev->asic_funcs->hw_queues_lock(hdev);
 631
 632         if (!hl_device_operational(hdev, &status)) {
 633                 atomic64_inc(&cntr->device_in_reset_drop_cnt);
 634                 atomic64_inc(&ctx->cs_counters.device_in_reset_drop_cnt);
 635                 dev_err(hdev->dev,
 636                         "device is %s, CS rejected!\n", hdev->status[status]);
 637                 rc = -EPERM;
 638                 goto out;
 639         }
 640
 641         max_queues = hdev->asic_prop.max_queues;
 642
 643         q = &hdev->kernel_queues[0];
 644         for (i = 0, cq_cnt = 0 ; i < max_queues ; i++, q++) {
 645                 if (cs->jobs_in_queue_cnt[i]) {
 646                         switch (q->queue_type) {
 647                         case QUEUE_TYPE_EXT:
 648                                 rc = ext_queue_sanity_checks(hdev, q,
 649                                                 cs->jobs_in_queue_cnt[i],
 650                                                 cs_needs_completion(cs) ?
 651                                                                 true : false);
 652                                 break;
 653                         case QUEUE_TYPE_INT:
 654                                 rc = int_queue_sanity_checks(hdev, q,
 655                                                 cs->jobs_in_queue_cnt[i]);
 656                                 break;
 657                         case QUEUE_TYPE_HW:
 658                                 rc = hw_queue_sanity_checks(hdev, q,
 659                                                 cs->jobs_in_queue_cnt[i]);
 660                                 break;
 661                         default:
 662                                 dev_err(hdev->dev, "Queue type %d is invalid\n",
 663                                         q->queue_type);
 664                                 rc = -EINVAL;
 665                                 break;
 666                         }
 667
 668                         if (rc) {
 669                                 atomic64_inc(
 670                                         &ctx->cs_counters.queue_full_drop_cnt);
 671                                 atomic64_inc(&cntr->queue_full_drop_cnt);
 672                                 goto unroll_cq_resv;
 673                         }
 674
 675                         if (q->queue_type == QUEUE_TYPE_EXT)
 676                                 cq_cnt++;
 677                 }
 678         }
 679
 680         if ((cs->type == CS_TYPE_SIGNAL) || (cs->type == CS_TYPE_WAIT)) {
 681                 rc = init_signal_wait_cs(cs);
 682                 if (rc)
 683                         goto unroll_cq_resv;
 684         } else if (cs->type == CS_TYPE_COLLECTIVE_WAIT) {
 685                 rc = hdev->asic_funcs->collective_wait_init_cs(cs);
 686                 if (rc)
 687                         goto unroll_cq_resv;
 688         }
 689
 690
 691         if (cs->encaps_signals && cs->staged_first) {
 692                 rc = encaps_sig_first_staged_cs_handler(hdev, cs);
 693                 if (rc)
 694                         goto unroll_cq_resv;
 695         }
 696
 697         spin_lock(&hdev->cs_mirror_lock);
 698
 699         /* Verify staged CS exists and add to the staged list */
 700         if (cs->staged_cs && !cs->staged_first) {
 701                 struct hl_cs *staged_cs;
 702
 703                 staged_cs = hl_staged_cs_find_first(hdev, cs->staged_sequence);
 704                 if (!staged_cs) {
 705                         dev_err(hdev->dev,
 706                                 "Cannot find staged submission sequence %llu",
 707                                 cs->staged_sequence);
 708                         rc = -EINVAL;
 709                         goto unlock_cs_mirror;
 710                 }
 711
 712                 if (is_staged_cs_last_exists(hdev, staged_cs)) {
 713                         dev_err(hdev->dev,
 714                                 "Staged submission sequence %llu already submitted",
 715                                 cs->staged_sequence);
 716                         rc = -EINVAL;
 717                         goto unlock_cs_mirror;
 718                 }
 719
 720                 list_add_tail(&cs->staged_cs_node, &staged_cs->staged_cs_node);
 721
 722                 /* update stream map of the first CS */
 723                 if (hdev->supports_wait_for_multi_cs)
 724                         staged_cs->fence->stream_master_qid_map |=
 725                                         cs->fence->stream_master_qid_map;
 726         }
 727
 728         list_add_tail(&cs->mirror_node, &hdev->cs_mirror_list);
 729
 730         /* Queue TDR if the CS is the first entry and if timeout is wanted */
 731         first_entry = list_first_entry(&hdev->cs_mirror_list,
 732                                         struct hl_cs, mirror_node) == cs;
 733         if ((hdev->timeout_jiffies != MAX_SCHEDULE_TIMEOUT) &&
 734                                 first_entry && cs_needs_timeout(cs)) {
 735                 cs->tdr_active = true;
 736                 schedule_delayed_work(&cs->work_tdr, cs->timeout_jiffies);
 737
 738         }
 739
 740         spin_unlock(&hdev->cs_mirror_lock);
 741
 742         list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
 743                 switch (job->queue_type) {
 744                 case QUEUE_TYPE_EXT:
 745                         ext_queue_schedule_job(job);
 746                         break;
 747                 case QUEUE_TYPE_INT:
 748                         int_queue_schedule_job(job);
 749                         break;
 750                 case QUEUE_TYPE_HW:
 751                         hw_queue_schedule_job(job);
 752                         break;
 753                 default:
 754                         break;
 755                 }
 756
 757         cs->submitted = true;
 758
 759         goto out;
 760
 761 unlock_cs_mirror:
 762         spin_unlock(&hdev->cs_mirror_lock);
 763 unroll_cq_resv:
 764         q = &hdev->kernel_queues[0];
 765         for (i = 0 ; (i < max_queues) && (cq_cnt > 0) ; i++, q++) {
 766                 if ((q->queue_type == QUEUE_TYPE_EXT) &&
 767                                                 (cs->jobs_in_queue_cnt[i])) {
 768                         atomic_t *free_slots =
 769                                 &hdev->completion_queue[i].free_slots_cnt;
 770                         atomic_add(cs->jobs_in_queue_cnt[i], free_slots);
 771                         cq_cnt--;
 772                 }
 773         }
 774
 775 out:
 776         hdev->asic_funcs->hw_queues_unlock(hdev);
 777
 778         return rc;
 779 }
 780
 781 /*
 782  * hl_hw_queue_inc_ci_kernel - increment ci for kernel's queue
 783  *
 784  * @hdev: pointer to hl_device structure
 785  * @hw_queue_id: which queue to increment its ci
 786  */
 787 void hl_hw_queue_inc_ci_kernel(struct hl_device *hdev, u32 hw_queue_id)
 788 {
 789         struct hl_hw_queue *q = &hdev->kernel_queues[hw_queue_id];
 790
 791         atomic_inc(&q->ci);
 792 }
 793
 794 static int ext_and_cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
 795                                         bool is_cpu_queue)
 796 {
 797         void *p;
 798         int rc;
 799
 800         if (is_cpu_queue)
 801                 p = hdev->asic_funcs->cpu_accessible_dma_pool_alloc(hdev,
 802                                                         HL_QUEUE_SIZE_IN_BYTES,
 803                                                         &q->bus_address);
 804         else
 805                 p = hdev->asic_funcs->asic_dma_alloc_coherent(hdev,
 806                                                 HL_QUEUE_SIZE_IN_BYTES,
 807                                                 &q->bus_address,
 808                                                 GFP_KERNEL | __GFP_ZERO);
 809         if (!p)
 810                 return -ENOMEM;
 811
 812         q->kernel_address = p;
 813
 814         q->shadow_queue = kmalloc_array(HL_QUEUE_LENGTH,
 815                                         sizeof(*q->shadow_queue),
 816                                         GFP_KERNEL);
 817         if (!q->shadow_queue) {
 818                 dev_err(hdev->dev,
 819                         "Failed to allocate shadow queue for H/W queue %d\n",
 820                         q->hw_queue_id);
 821                 rc = -ENOMEM;
 822                 goto free_queue;
 823         }
 824
 825         /* Make sure read/write pointers are initialized to start of queue */
 826         atomic_set(&q->ci, 0);
 827         q->pi = 0;
 828
 829         return 0;
 830
 831 free_queue:
 832         if (is_cpu_queue)
 833                 hdev->asic_funcs->cpu_accessible_dma_pool_free(hdev,
 834                                         HL_QUEUE_SIZE_IN_BYTES,
 835                                         q->kernel_address);
 836         else
 837                 hdev->asic_funcs->asic_dma_free_coherent(hdev,
 838                                         HL_QUEUE_SIZE_IN_BYTES,
 839                                         q->kernel_address,
 840                                         q->bus_address);
 841
 842         return rc;
 843 }
 844
 845 static int int_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 846 {
 847         void *p;
 848
 849         p = hdev->asic_funcs->get_int_queue_base(hdev, q->hw_queue_id,
 850                                         &q->bus_address, &q->int_queue_len);
 851         if (!p) {
 852                 dev_err(hdev->dev,
 853                         "Failed to get base address for internal queue %d\n",
 854                         q->hw_queue_id);
 855                 return -EFAULT;
 856         }
 857
 858         q->kernel_address = p;
 859         q->pi = 0;
 860         atomic_set(&q->ci, 0);
 861
 862         return 0;
 863 }
 864
 865 static int cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 866 {
 867         return ext_and_cpu_queue_init(hdev, q, true);
 868 }
 869
 870 static int ext_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 871 {
 872         return ext_and_cpu_queue_init(hdev, q, false);
 873 }
 874
 875 static int hw_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
 876 {
 877         void *p;
 878
 879         p = hdev->asic_funcs->asic_dma_alloc_coherent(hdev,
 880                                                 HL_QUEUE_SIZE_IN_BYTES,
 881                                                 &q->bus_address,
 882                                                 GFP_KERNEL | __GFP_ZERO);
 883         if (!p)
 884                 return -ENOMEM;
 885
 886         q->kernel_address = p;
 887
 888         /* Make sure read/write pointers are initialized to start of queue */
 889         atomic_set(&q->ci, 0);
 890         q->pi = 0;
 891
 892         return 0;
 893 }
 894
 895 static void sync_stream_queue_init(struct hl_device *hdev, u32 q_idx)
 896 {
 897         struct hl_sync_stream_properties *sync_stream_prop;
 898         struct asic_fixed_properties *prop = &hdev->asic_prop;
 899         struct hl_hw_sob *hw_sob;
 900         int sob, reserved_mon_idx, queue_idx;
 901
 902         sync_stream_prop = &hdev->kernel_queues[q_idx].sync_stream_prop;
 903
 904         /* We use 'collective_mon_idx' as a running index in order to reserve
 905          * monitors for collective master/slave queues.
 906          * collective master queue gets 2 reserved monitors
 907          * collective slave queue gets 1 reserved monitor
 908          */
 909         if (hdev->kernel_queues[q_idx].collective_mode ==
 910                         HL_COLLECTIVE_MASTER) {
 911                 reserved_mon_idx = hdev->collective_mon_idx;
 912
 913                 /* reserve the first monitor for collective master queue */
 914                 sync_stream_prop->collective_mstr_mon_id[0] =
 915                         prop->collective_first_mon + reserved_mon_idx;
 916
 917                 /* reserve the second monitor for collective master queue */
 918                 sync_stream_prop->collective_mstr_mon_id[1] =
 919                         prop->collective_first_mon + reserved_mon_idx + 1;
 920
 921                 hdev->collective_mon_idx += HL_COLLECTIVE_RSVD_MSTR_MONS;
 922         } else if (hdev->kernel_queues[q_idx].collective_mode ==
 923                         HL_COLLECTIVE_SLAVE) {
 924                 reserved_mon_idx = hdev->collective_mon_idx++;
 925
 926                 /* reserve a monitor for collective slave queue */
 927                 sync_stream_prop->collective_slave_mon_id =
 928                         prop->collective_first_mon + reserved_mon_idx;
 929         }
 930
 931         if (!hdev->kernel_queues[q_idx].supports_sync_stream)
 932                 return;
 933
 934         queue_idx = hdev->sync_stream_queue_idx++;
 935
 936         sync_stream_prop->base_sob_id = prop->sync_stream_first_sob +
 937                         (queue_idx * HL_RSVD_SOBS);
 938         sync_stream_prop->base_mon_id = prop->sync_stream_first_mon +
 939                         (queue_idx * HL_RSVD_MONS);
 940         sync_stream_prop->next_sob_val = 1;
 941         sync_stream_prop->curr_sob_offset = 0;
 942
 943         for (sob = 0 ; sob < HL_RSVD_SOBS ; sob++) {
 944                 hw_sob = &sync_stream_prop->hw_sob[sob];
 945                 hw_sob->hdev = hdev;
 946                 hw_sob->sob_id = sync_stream_prop->base_sob_id + sob;
 947                 hw_sob->sob_addr =
 948                         hdev->asic_funcs->get_sob_addr(hdev, hw_sob->sob_id);
 949                 hw_sob->q_idx = q_idx;
 950                 kref_init(&hw_sob->kref);
 951         }
 952 }
 953
 954 static void sync_stream_queue_reset(struct hl_device *hdev, u32 q_idx)
 955 {
 956         struct hl_sync_stream_properties *prop =
 957                         &hdev->kernel_queues[q_idx].sync_stream_prop;
 958
 959         /*
 960          * In case we got here due to a stuck CS, the refcnt might be bigger
 961          * than 1 and therefore we reset it.
 962          */
 963         kref_init(&prop->hw_sob[prop->curr_sob_offset].kref);
 964         prop->curr_sob_offset = 0;
 965         prop->next_sob_val = 1;
 966 }
 967
 968 /*
 969  * queue_init - main initialization function for H/W queue object
 970  *
 971  * @hdev: pointer to hl_device device structure
 972  * @q: pointer to hl_hw_queue queue structure
 973  * @hw_queue_id: The id of the H/W queue
 974  *
 975  * Allocate dma-able memory for the queue and initialize fields
 976  * Returns 0 on success
 977  */
 978 static int queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
 979                         u32 hw_queue_id)
 980 {
 981         int rc;
 982
 983         q->hw_queue_id = hw_queue_id;
 984
 985         switch (q->queue_type) {
 986         case QUEUE_TYPE_EXT:
 987                 rc = ext_queue_init(hdev, q);
 988                 break;
 989         case QUEUE_TYPE_INT:
 990                 rc = int_queue_init(hdev, q);
 991                 break;
 992         case QUEUE_TYPE_CPU:
 993                 rc = cpu_queue_init(hdev, q);
 994                 break;
 995         case QUEUE_TYPE_HW:
 996                 rc = hw_queue_init(hdev, q);
 997                 break;
 998         case QUEUE_TYPE_NA:
 999                 q->valid = 0;
1000                 return 0;
1001         default:
1002                 dev_crit(hdev->dev, "wrong queue type %d during init\n",
1003                         q->queue_type);
1004                 rc = -EINVAL;
1005                 break;
1006         }
1007
1008         sync_stream_queue_init(hdev, q->hw_queue_id);
1009
1010         if (rc)
1011                 return rc;
1012
1013         q->valid = 1;
1014
1015         return 0;
1016 }
1017
1018 /*
1019  * hw_queue_fini - destroy queue
1020  *
1021  * @hdev: pointer to hl_device device structure
1022  * @q: pointer to hl_hw_queue queue structure
1023  *
1024  * Free the queue memory
1025  */
1026 static void queue_fini(struct hl_device *hdev, struct hl_hw_queue *q)
1027 {
1028         if (!q->valid)
1029                 return;
1030
1031         /*
1032          * If we arrived here, there are no jobs waiting on this queue
1033          * so we can safely remove it.
1034          * This is because this function can only called when:
1035          * 1. Either a context is deleted, which only can occur if all its
1036          *    jobs were finished
1037          * 2. A context wasn't able to be created due to failure or timeout,
1038          *    which means there are no jobs on the queue yet
1039          *
1040          * The only exception are the queues of the kernel context, but
1041          * if they are being destroyed, it means that the entire module is
1042          * being removed. If the module is removed, it means there is no open
1043          * user context. It also means that if a job was submitted by
1044          * the kernel driver (e.g. context creation), the job itself was
1045          * released by the kernel driver when a timeout occurred on its
1046          * Completion. Thus, we don't need to release it again.
1047          */
1048
1049         if (q->queue_type == QUEUE_TYPE_INT)
1050                 return;
1051
1052         kfree(q->shadow_queue);
1053
1054         if (q->queue_type == QUEUE_TYPE_CPU)
1055                 hdev->asic_funcs->cpu_accessible_dma_pool_free(hdev,
1056                                         HL_QUEUE_SIZE_IN_BYTES,
1057                                         q->kernel_address);
1058         else
1059                 hdev->asic_funcs->asic_dma_free_coherent(hdev,
1060                                         HL_QUEUE_SIZE_IN_BYTES,
1061                                         q->kernel_address,
1062                                         q->bus_address);
1063 }
1064
1065 int hl_hw_queues_create(struct hl_device *hdev)
1066 {
1067         struct asic_fixed_properties *asic = &hdev->asic_prop;
1068         struct hl_hw_queue *q;
1069         int i, rc, q_ready_cnt;
1070
1071         hdev->kernel_queues = kcalloc(asic->max_queues,
1072                                 sizeof(*hdev->kernel_queues), GFP_KERNEL);
1073
1074         if (!hdev->kernel_queues) {
1075                 dev_err(hdev->dev, "Not enough memory for H/W queues\n");
1076                 return -ENOMEM;
1077         }
1078
1079         /* Initialize the H/W queues */
1080         for (i = 0, q_ready_cnt = 0, q = hdev->kernel_queues;
1081                         i < asic->max_queues ; i++, q_ready_cnt++, q++) {
1082
1083                 q->queue_type = asic->hw_queues_props[i].type;
1084                 q->supports_sync_stream =
1085                                 asic->hw_queues_props[i].supports_sync_stream;
1086                 q->collective_mode = asic->hw_queues_props[i].collective_mode;
1087                 rc = queue_init(hdev, q, i);
1088                 if (rc) {
1089                         dev_err(hdev->dev,
1090                                 "failed to initialize queue %d\n", i);
1091                         goto release_queues;
1092                 }
1093         }
1094
1095         return 0;
1096
1097 release_queues:
1098         for (i = 0, q = hdev->kernel_queues ; i < q_ready_cnt ; i++, q++)
1099                 queue_fini(hdev, q);
1100
1101         kfree(hdev->kernel_queues);
1102
1103         return rc;
1104 }
1105
1106 void hl_hw_queues_destroy(struct hl_device *hdev)
1107 {
1108         struct hl_hw_queue *q;
1109         u32 max_queues = hdev->asic_prop.max_queues;
1110         int i;
1111
1112         for (i = 0, q = hdev->kernel_queues ; i < max_queues ; i++, q++)
1113                 queue_fini(hdev, q);
1114
1115         kfree(hdev->kernel_queues);
1116 }
1117
1118 void hl_hw_queue_reset(struct hl_device *hdev, bool hard_reset)
1119 {
1120         struct hl_hw_queue *q;
1121         u32 max_queues = hdev->asic_prop.max_queues;
1122         int i;
1123
1124         for (i = 0, q = hdev->kernel_queues ; i < max_queues ; i++, q++) {
1125                 if ((!q->valid) ||
1126                         ((!hard_reset) && (q->queue_type == QUEUE_TYPE_CPU)))
1127                         continue;
1128                 q->pi = 0;
1129                 atomic_set(&q->ci, 0);
1130
1131                 if (q->supports_sync_stream)
1132                         sync_stream_queue_reset(hdev, q->hw_queue_id);
1133         }
1134 }