drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c

   1 /*
   2  * Copyright 2014 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  */
  23
  24 #include <linux/slab.h>
  25 #include <linux/mutex.h>
  26 #include "kfd_device_queue_manager.h"
  27 #include "kfd_kernel_queue.h"
  28 #include "kfd_priv.h"
  29
  30 static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
  31                                 unsigned int buffer_size_bytes)
  32 {
  33         unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t);
  34
  35         WARN((temp * sizeof(uint32_t)) > buffer_size_bytes,
  36              "Runlist IB overflow");
  37         *wptr = temp;
  38 }
  39
  40 static void pm_calc_rlib_size(struct packet_manager *pm,
  41                                 unsigned int *rlib_size,
  42                                 bool *over_subscription)
  43 {
  44         unsigned int process_count, queue_count, compute_queue_count;
  45         unsigned int map_queue_size;
  46         unsigned int max_proc_per_quantum = 1;
  47         struct kfd_dev *dev = pm->dqm->dev;
  48
  49         process_count = pm->dqm->processes_count;
  50         queue_count = pm->dqm->queue_count;
  51         compute_queue_count = queue_count - pm->dqm->sdma_queue_count;
  52
  53         /* check if there is over subscription
  54          * Note: the arbitration between the number of VMIDs and
  55          * hws_max_conc_proc has been done in
  56          * kgd2kfd_device_init().
  57          */
  58         *over_subscription = false;
  59
  60         if (dev->max_proc_per_quantum > 1)
  61                 max_proc_per_quantum = dev->max_proc_per_quantum;
  62
  63         if ((process_count > max_proc_per_quantum) ||
  64             compute_queue_count > get_queues_num(pm->dqm)) {
  65                 *over_subscription = true;
  66                 pr_debug("Over subscribed runlist\n");
  67         }
  68
  69         map_queue_size = pm->pmf->map_queues_size;
  70         /* calculate run list ib allocation size */
  71         *rlib_size = process_count * pm->pmf->map_process_size +
  72                      queue_count * map_queue_size;
  73
  74         /*
  75          * Increase the allocation size in case we need a chained run list
  76          * when over subscription
  77          */
  78         if (*over_subscription)
  79                 *rlib_size += pm->pmf->runlist_size;
  80
  81         pr_debug("runlist ib size %d\n", *rlib_size);
  82 }
  83
  84 static int pm_allocate_runlist_ib(struct packet_manager *pm,
  85                                 unsigned int **rl_buffer,
  86                                 uint64_t *rl_gpu_buffer,
  87                                 unsigned int *rl_buffer_size,
  88                                 bool *is_over_subscription)
  89 {
  90         int retval;
  91
  92         if (WARN_ON(pm->allocated))
  93                 return -EINVAL;
  94
  95         pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription);
  96
  97         mutex_lock(&pm->lock);
  98
  99         retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size,
 100                                         &pm->ib_buffer_obj);
 101
 102         if (retval) {
 103                 pr_err("Failed to allocate runlist IB\n");
 104                 goto out;
 105         }
 106
 107         *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
 108         *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
 109
 110         memset(*rl_buffer, 0, *rl_buffer_size);
 111         pm->allocated = true;
 112
 113 out:
 114         mutex_unlock(&pm->lock);
 115         return retval;
 116 }
 117
 118 static int pm_create_runlist_ib(struct packet_manager *pm,
 119                                 struct list_head *queues,
 120                                 uint64_t *rl_gpu_addr,
 121                                 size_t *rl_size_bytes)
 122 {
 123         unsigned int alloc_size_bytes;
 124         unsigned int *rl_buffer, rl_wptr, i;
 125         int retval, proccesses_mapped;
 126         struct device_process_node *cur;
 127         struct qcm_process_device *qpd;
 128         struct queue *q;
 129         struct kernel_queue *kq;
 130         bool is_over_subscription;
 131
 132         rl_wptr = retval = proccesses_mapped = 0;
 133
 134         retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
 135                                 &alloc_size_bytes, &is_over_subscription);
 136         if (retval)
 137                 return retval;
 138
 139         *rl_size_bytes = alloc_size_bytes;
 140         pm->ib_size_bytes = alloc_size_bytes;
 141
 142         pr_debug("Building runlist ib process count: %d queues count %d\n",
 143                 pm->dqm->processes_count, pm->dqm->queue_count);
 144
 145         /* build the run list ib packet */
 146         list_for_each_entry(cur, queues, list) {
 147                 qpd = cur->qpd;
 148                 /* build map process packet */
 149                 if (proccesses_mapped >= pm->dqm->processes_count) {
 150                         pr_debug("Not enough space left in runlist IB\n");
 151                         pm_release_ib(pm);
 152                         return -ENOMEM;
 153                 }
 154
 155                 retval = pm->pmf->map_process(pm, &rl_buffer[rl_wptr], qpd);
 156                 if (retval)
 157                         return retval;
 158
 159                 proccesses_mapped++;
 160                 inc_wptr(&rl_wptr, pm->pmf->map_process_size,
 161                                 alloc_size_bytes);
 162
 163                 list_for_each_entry(kq, &qpd->priv_queue_list, list) {
 164                         if (!kq->queue->properties.is_active)
 165                                 continue;
 166
 167                         pr_debug("static_queue, mapping kernel q %d, is debug status %d\n",
 168                                 kq->queue->queue, qpd->is_debug);
 169
 170                         retval = pm->pmf->map_queues(pm,
 171                                                 &rl_buffer[rl_wptr],
 172                                                 kq->queue,
 173                                                 qpd->is_debug);
 174                         if (retval)
 175                                 return retval;
 176
 177                         inc_wptr(&rl_wptr,
 178                                 pm->pmf->map_queues_size,
 179                                 alloc_size_bytes);
 180                 }
 181
 182                 list_for_each_entry(q, &qpd->queues_list, list) {
 183                         if (!q->properties.is_active)
 184                                 continue;
 185
 186                         pr_debug("static_queue, mapping user queue %d, is debug status %d\n",
 187                                 q->queue, qpd->is_debug);
 188
 189                         retval = pm->pmf->map_queues(pm,
 190                                                 &rl_buffer[rl_wptr],
 191                                                 q,
 192                                                 qpd->is_debug);
 193
 194                         if (retval)
 195                                 return retval;
 196
 197                         inc_wptr(&rl_wptr,
 198                                 pm->pmf->map_queues_size,
 199                                 alloc_size_bytes);
 200                 }
 201         }
 202
 203         pr_debug("Finished map process and queues to runlist\n");
 204
 205         if (is_over_subscription)
 206                 retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr],
 207                                         *rl_gpu_addr,
 208                                         alloc_size_bytes / sizeof(uint32_t),
 209                                         true);
 210
 211         for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
 212                 pr_debug("0x%2X ", rl_buffer[i]);
 213         pr_debug("\n");
 214
 215         return retval;
 216 }
 217
 218 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
 219 {
 220         switch (dqm->dev->device_info->asic_family) {
 221         case CHIP_KAVERI:
 222         case CHIP_HAWAII:
 223                 /* PM4 packet structures on CIK are the same as on VI */
 224         case CHIP_CARRIZO:
 225         case CHIP_TONGA:
 226         case CHIP_FIJI:
 227         case CHIP_POLARIS10:
 228         case CHIP_POLARIS11:
 229                 pm->pmf = &kfd_vi_pm_funcs;
 230                 break;
 231         case CHIP_VEGA10:
 232         case CHIP_RAVEN:
 233                 pm->pmf = &kfd_v9_pm_funcs;
 234                 break;
 235         default:
 236                 WARN(1, "Unexpected ASIC family %u",
 237                      dqm->dev->device_info->asic_family);
 238                 return -EINVAL;
 239         }
 240
 241         pm->dqm = dqm;
 242         mutex_init(&pm->lock);
 243         pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
 244         if (!pm->priv_queue) {
 245                 mutex_destroy(&pm->lock);
 246                 return -ENOMEM;
 247         }
 248         pm->allocated = false;
 249
 250         return 0;
 251 }
 252
 253 void pm_uninit(struct packet_manager *pm)
 254 {
 255         mutex_destroy(&pm->lock);
 256         kernel_queue_uninit(pm->priv_queue);
 257 }
 258
 259 int pm_send_set_resources(struct packet_manager *pm,
 260                                 struct scheduling_resources *res)
 261 {
 262         uint32_t *buffer, size;
 263         int retval = 0;
 264
 265         size = pm->pmf->set_resources_size;
 266         mutex_lock(&pm->lock);
 267         pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
 268                                         size / sizeof(uint32_t),
 269                                         (unsigned int **)&buffer);
 270         if (!buffer) {
 271                 pr_err("Failed to allocate buffer on kernel queue\n");
 272                 retval = -ENOMEM;
 273                 goto out;
 274         }
 275
 276         retval = pm->pmf->set_resources(pm, buffer, res);
 277         if (!retval)
 278                 pm->priv_queue->ops.submit_packet(pm->priv_queue);
 279         else
 280                 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
 281
 282 out:
 283         mutex_unlock(&pm->lock);
 284
 285         return retval;
 286 }
 287
 288 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
 289 {
 290         uint64_t rl_gpu_ib_addr;
 291         uint32_t *rl_buffer;
 292         size_t rl_ib_size, packet_size_dwords;
 293         int retval;
 294
 295         retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
 296                                         &rl_ib_size);
 297         if (retval)
 298                 goto fail_create_runlist_ib;
 299
 300         pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
 301
 302         packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t);
 303         mutex_lock(&pm->lock);
 304
 305         retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
 306                                         packet_size_dwords, &rl_buffer);
 307         if (retval)
 308                 goto fail_acquire_packet_buffer;
 309
 310         retval = pm->pmf->runlist(pm, rl_buffer, rl_gpu_ib_addr,
 311                                         rl_ib_size / sizeof(uint32_t), false);
 312         if (retval)
 313                 goto fail_create_runlist;
 314
 315         pm->priv_queue->ops.submit_packet(pm->priv_queue);
 316
 317         mutex_unlock(&pm->lock);
 318
 319         return retval;
 320
 321 fail_create_runlist:
 322         pm->priv_queue->ops.rollback_packet(pm->priv_queue);
 323 fail_acquire_packet_buffer:
 324         mutex_unlock(&pm->lock);
 325 fail_create_runlist_ib:
 326         pm_release_ib(pm);
 327         return retval;
 328 }
 329
 330 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
 331                         uint32_t fence_value)
 332 {
 333         uint32_t *buffer, size;
 334         int retval = 0;
 335
 336         if (WARN_ON(!fence_address))
 337                 return -EFAULT;
 338
 339         size = pm->pmf->query_status_size;
 340         mutex_lock(&pm->lock);
 341         pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
 342                         size / sizeof(uint32_t), (unsigned int **)&buffer);
 343         if (!buffer) {
 344                 pr_err("Failed to allocate buffer on kernel queue\n");
 345                 retval = -ENOMEM;
 346                 goto out;
 347         }
 348
 349         retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value);
 350         if (!retval)
 351                 pm->priv_queue->ops.submit_packet(pm->priv_queue);
 352         else
 353                 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
 354
 355 out:
 356         mutex_unlock(&pm->lock);
 357         return retval;
 358 }
 359
 360 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
 361                         enum kfd_unmap_queues_filter filter,
 362                         uint32_t filter_param, bool reset,
 363                         unsigned int sdma_engine)
 364 {
 365         uint32_t *buffer, size;
 366         int retval = 0;
 367
 368         size = pm->pmf->unmap_queues_size;
 369         mutex_lock(&pm->lock);
 370         pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
 371                         size / sizeof(uint32_t), (unsigned int **)&buffer);
 372         if (!buffer) {
 373                 pr_err("Failed to allocate buffer on kernel queue\n");
 374                 retval = -ENOMEM;
 375                 goto out;
 376         }
 377
 378         retval = pm->pmf->unmap_queues(pm, buffer, type, filter, filter_param,
 379                                        reset, sdma_engine);
 380         if (!retval)
 381                 pm->priv_queue->ops.submit_packet(pm->priv_queue);
 382         else
 383                 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
 384
 385 out:
 386         mutex_unlock(&pm->lock);
 387         return retval;
 388 }
 389
 390 void pm_release_ib(struct packet_manager *pm)
 391 {
 392         mutex_lock(&pm->lock);
 393         if (pm->allocated) {
 394                 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
 395                 pm->allocated = false;
 396         }
 397         mutex_unlock(&pm->lock);
 398 }
 399
 400 #if defined(CONFIG_DEBUG_FS)
 401
 402 int pm_debugfs_runlist(struct seq_file *m, void *data)
 403 {
 404         struct packet_manager *pm = data;
 405
 406         mutex_lock(&pm->lock);
 407
 408         if (!pm->allocated) {
 409                 seq_puts(m, "  No active runlist\n");
 410                 goto out;
 411         }
 412
 413         seq_hex_dump(m, "  ", DUMP_PREFIX_OFFSET, 32, 4,
 414                      pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false);
 415
 416 out:
 417         mutex_unlock(&pm->lock);
 418         return 0;
 419 }
 420
 421 int pm_debugfs_hang_hws(struct packet_manager *pm)
 422 {
 423         uint32_t *buffer, size;
 424         int r = 0;
 425
 426         size = pm->pmf->query_status_size;
 427         mutex_lock(&pm->lock);
 428         pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
 429                         size / sizeof(uint32_t), (unsigned int **)&buffer);
 430         if (!buffer) {
 431                 pr_err("Failed to allocate buffer on kernel queue\n");
 432                 r = -ENOMEM;
 433                 goto out;
 434         }
 435         memset(buffer, 0x55, size);
 436         pm->priv_queue->ops.submit_packet(pm->priv_queue);
 437
 438         pr_info("Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.",
 439                 buffer[0], buffer[1], buffer[2], buffer[3],
 440                 buffer[4], buffer[5], buffer[6]);
 441 out:
 442         mutex_unlock(&pm->lock);
 443         return r;
 444 }
 445
 446
 447 #endif