| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * Copyright (C) 2024, Advanced Micro Devices, Inc. |
| 4 | */ |
| 5 | |
| 6 | #include <drm/amdxdna_accel.h> |
| 7 | #include <drm/drm_device.h> |
| 8 | #include <drm/drm_gem.h> |
| 9 | #include <drm/drm_gem_shmem_helper.h> |
| 10 | #include <drm/drm_print.h> |
| 11 | #include <drm/drm_syncobj.h> |
| 12 | #include <linux/hmm.h> |
| 13 | #include <linux/types.h> |
| 14 | #include <linux/xarray.h> |
| 15 | #include <trace/events/amdxdna.h> |
| 16 | |
| 17 | #include "aie2_msg_priv.h" |
| 18 | #include "aie2_pci.h" |
| 19 | #include "aie2_solver.h" |
| 20 | #include "amdxdna_ctx.h" |
| 21 | #include "amdxdna_gem.h" |
| 22 | #include "amdxdna_mailbox.h" |
| 23 | #include "amdxdna_pci_drv.h" |
| 24 | |
| 25 | static bool force_cmdlist; |
| 26 | module_param(force_cmdlist, bool, 0600); |
| 27 | MODULE_PARM_DESC(force_cmdlist, "Force use command list (Default false)" ); |
| 28 | |
| 29 | #define HWCTX_MAX_TIMEOUT 60000 /* milliseconds */ |
| 30 | |
| 31 | static void aie2_job_release(struct kref *ref) |
| 32 | { |
| 33 | struct amdxdna_sched_job *job; |
| 34 | |
| 35 | job = container_of(ref, struct amdxdna_sched_job, refcnt); |
| 36 | amdxdna_sched_job_cleanup(job); |
| 37 | atomic64_inc(v: &job->hwctx->job_free_cnt); |
| 38 | wake_up(&job->hwctx->priv->job_free_wq); |
| 39 | if (job->out_fence) |
| 40 | dma_fence_put(fence: job->out_fence); |
| 41 | kfree(objp: job); |
| 42 | } |
| 43 | |
| 44 | static void aie2_job_put(struct amdxdna_sched_job *job) |
| 45 | { |
| 46 | kref_put(kref: &job->refcnt, release: aie2_job_release); |
| 47 | } |
| 48 | |
| 49 | /* The bad_job is used in aie2_sched_job_timedout, otherwise, set it to NULL */ |
| 50 | static void aie2_hwctx_stop(struct amdxdna_dev *xdna, struct amdxdna_hwctx *hwctx, |
| 51 | struct drm_sched_job *bad_job) |
| 52 | { |
| 53 | drm_sched_stop(sched: &hwctx->priv->sched, bad: bad_job); |
| 54 | aie2_destroy_context(ndev: xdna->dev_handle, hwctx); |
| 55 | } |
| 56 | |
| 57 | static int aie2_hwctx_restart(struct amdxdna_dev *xdna, struct amdxdna_hwctx *hwctx) |
| 58 | { |
| 59 | struct amdxdna_gem_obj *heap = hwctx->priv->heap; |
| 60 | int ret; |
| 61 | |
| 62 | ret = aie2_create_context(ndev: xdna->dev_handle, hwctx); |
| 63 | if (ret) { |
| 64 | XDNA_ERR(xdna, "Create hwctx failed, ret %d" , ret); |
| 65 | goto out; |
| 66 | } |
| 67 | |
| 68 | ret = aie2_map_host_buf(ndev: xdna->dev_handle, context_id: hwctx->fw_ctx_id, |
| 69 | addr: heap->mem.userptr, size: heap->mem.size); |
| 70 | if (ret) { |
| 71 | XDNA_ERR(xdna, "Map host buf failed, ret %d" , ret); |
| 72 | goto out; |
| 73 | } |
| 74 | |
| 75 | if (hwctx->status != HWCTX_STAT_READY) { |
| 76 | XDNA_DBG(xdna, "hwctx is not ready, status %d" , hwctx->status); |
| 77 | goto out; |
| 78 | } |
| 79 | |
| 80 | ret = aie2_config_cu(hwctx); |
| 81 | if (ret) { |
| 82 | XDNA_ERR(xdna, "Config cu failed, ret %d" , ret); |
| 83 | goto out; |
| 84 | } |
| 85 | |
| 86 | out: |
| 87 | drm_sched_start(sched: &hwctx->priv->sched, errno: 0); |
| 88 | XDNA_DBG(xdna, "%s restarted, ret %d" , hwctx->name, ret); |
| 89 | return ret; |
| 90 | } |
| 91 | |
| 92 | void aie2_restart_ctx(struct amdxdna_client *client) |
| 93 | { |
| 94 | struct amdxdna_dev *xdna = client->xdna; |
| 95 | struct amdxdna_hwctx *hwctx; |
| 96 | unsigned long hwctx_id; |
| 97 | |
| 98 | drm_WARN_ON(&xdna->ddev, !mutex_is_locked(&xdna->dev_lock)); |
| 99 | mutex_lock(&client->hwctx_lock); |
| 100 | amdxdna_for_each_hwctx(client, hwctx_id, hwctx) { |
| 101 | if (hwctx->status != HWCTX_STAT_STOP) |
| 102 | continue; |
| 103 | |
| 104 | hwctx->status = hwctx->old_status; |
| 105 | XDNA_DBG(xdna, "Resetting %s" , hwctx->name); |
| 106 | aie2_hwctx_restart(xdna, hwctx); |
| 107 | } |
| 108 | mutex_unlock(lock: &client->hwctx_lock); |
| 109 | } |
| 110 | |
| 111 | static struct dma_fence *aie2_cmd_get_out_fence(struct amdxdna_hwctx *hwctx, u64 seq) |
| 112 | { |
| 113 | struct dma_fence *fence, *out_fence = NULL; |
| 114 | int ret; |
| 115 | |
| 116 | fence = drm_syncobj_fence_get(syncobj: hwctx->priv->syncobj); |
| 117 | if (!fence) |
| 118 | return NULL; |
| 119 | |
| 120 | ret = dma_fence_chain_find_seqno(pfence: &fence, seqno: seq); |
| 121 | if (ret) |
| 122 | goto out; |
| 123 | |
| 124 | out_fence = dma_fence_get(fence: dma_fence_chain_contained(fence)); |
| 125 | |
| 126 | out: |
| 127 | dma_fence_put(fence); |
| 128 | return out_fence; |
| 129 | } |
| 130 | |
| 131 | static void aie2_hwctx_wait_for_idle(struct amdxdna_hwctx *hwctx) |
| 132 | { |
| 133 | struct dma_fence *fence; |
| 134 | |
| 135 | fence = aie2_cmd_get_out_fence(hwctx, seq: hwctx->priv->seq - 1); |
| 136 | if (!fence) |
| 137 | return; |
| 138 | |
| 139 | /* Wait up to 2 seconds for fw to finish all pending requests */ |
| 140 | dma_fence_wait_timeout(fence, intr: false, timeout: msecs_to_jiffies(m: 2000)); |
| 141 | dma_fence_put(fence); |
| 142 | } |
| 143 | |
| 144 | void aie2_hwctx_suspend(struct amdxdna_hwctx *hwctx) |
| 145 | { |
| 146 | struct amdxdna_dev *xdna = hwctx->client->xdna; |
| 147 | |
| 148 | /* |
| 149 | * Command timeout is unlikely. But if it happens, it doesn't |
| 150 | * break the system. aie2_hwctx_stop() will destroy mailbox |
| 151 | * and abort all commands. |
| 152 | */ |
| 153 | drm_WARN_ON(&xdna->ddev, !mutex_is_locked(&xdna->dev_lock)); |
| 154 | aie2_hwctx_wait_for_idle(hwctx); |
| 155 | aie2_hwctx_stop(xdna, hwctx, NULL); |
| 156 | hwctx->old_status = hwctx->status; |
| 157 | hwctx->status = HWCTX_STAT_STOP; |
| 158 | } |
| 159 | |
| 160 | void aie2_hwctx_resume(struct amdxdna_hwctx *hwctx) |
| 161 | { |
| 162 | struct amdxdna_dev *xdna = hwctx->client->xdna; |
| 163 | |
| 164 | /* |
| 165 | * The resume path cannot guarantee that mailbox channel can be |
| 166 | * regenerated. If this happen, when submit message to this |
| 167 | * mailbox channel, error will return. |
| 168 | */ |
| 169 | drm_WARN_ON(&xdna->ddev, !mutex_is_locked(&xdna->dev_lock)); |
| 170 | hwctx->status = hwctx->old_status; |
| 171 | aie2_hwctx_restart(xdna, hwctx); |
| 172 | } |
| 173 | |
| 174 | static void |
| 175 | aie2_sched_notify(struct amdxdna_sched_job *job) |
| 176 | { |
| 177 | struct dma_fence *fence = job->fence; |
| 178 | |
| 179 | trace_xdna_job(sched_job: &job->base, name: job->hwctx->name, str: "signaled fence" , seq: job->seq); |
| 180 | job->hwctx->priv->completed++; |
| 181 | dma_fence_signal(fence); |
| 182 | |
| 183 | up(sem: &job->hwctx->priv->job_sem); |
| 184 | job->job_done = true; |
| 185 | dma_fence_put(fence); |
| 186 | mmput_async(job->mm); |
| 187 | aie2_job_put(job); |
| 188 | } |
| 189 | |
| 190 | static int |
| 191 | aie2_sched_resp_handler(void *handle, void __iomem *data, size_t size) |
| 192 | { |
| 193 | struct amdxdna_sched_job *job = handle; |
| 194 | struct amdxdna_gem_obj *cmd_abo; |
| 195 | u32 ret = 0; |
| 196 | u32 status; |
| 197 | |
| 198 | cmd_abo = job->cmd_bo; |
| 199 | |
| 200 | if (unlikely(!data)) |
| 201 | goto out; |
| 202 | |
| 203 | if (unlikely(size != sizeof(u32))) { |
| 204 | amdxdna_cmd_set_state(abo: cmd_abo, s: ERT_CMD_STATE_ABORT); |
| 205 | ret = -EINVAL; |
| 206 | goto out; |
| 207 | } |
| 208 | |
| 209 | status = readl(addr: data); |
| 210 | XDNA_DBG(job->hwctx->client->xdna, "Resp status 0x%x" , status); |
| 211 | if (status == AIE2_STATUS_SUCCESS) |
| 212 | amdxdna_cmd_set_state(abo: cmd_abo, s: ERT_CMD_STATE_COMPLETED); |
| 213 | else |
| 214 | amdxdna_cmd_set_state(abo: cmd_abo, s: ERT_CMD_STATE_ERROR); |
| 215 | |
| 216 | out: |
| 217 | aie2_sched_notify(job); |
| 218 | return ret; |
| 219 | } |
| 220 | |
| 221 | static int |
| 222 | aie2_sched_nocmd_resp_handler(void *handle, void __iomem *data, size_t size) |
| 223 | { |
| 224 | struct amdxdna_sched_job *job = handle; |
| 225 | u32 ret = 0; |
| 226 | u32 status; |
| 227 | |
| 228 | if (unlikely(!data)) |
| 229 | goto out; |
| 230 | |
| 231 | if (unlikely(size != sizeof(u32))) { |
| 232 | ret = -EINVAL; |
| 233 | goto out; |
| 234 | } |
| 235 | |
| 236 | status = readl(addr: data); |
| 237 | XDNA_DBG(job->hwctx->client->xdna, "Resp status 0x%x" , status); |
| 238 | |
| 239 | out: |
| 240 | aie2_sched_notify(job); |
| 241 | return ret; |
| 242 | } |
| 243 | |
| 244 | static int |
| 245 | aie2_sched_cmdlist_resp_handler(void *handle, void __iomem *data, size_t size) |
| 246 | { |
| 247 | struct amdxdna_sched_job *job = handle; |
| 248 | struct amdxdna_gem_obj *cmd_abo; |
| 249 | struct amdxdna_dev *xdna; |
| 250 | u32 fail_cmd_status; |
| 251 | u32 fail_cmd_idx; |
| 252 | u32 cmd_status; |
| 253 | u32 ret = 0; |
| 254 | |
| 255 | cmd_abo = job->cmd_bo; |
| 256 | if (unlikely(!data) || unlikely(size != sizeof(u32) * 3)) { |
| 257 | amdxdna_cmd_set_state(abo: cmd_abo, s: ERT_CMD_STATE_ABORT); |
| 258 | ret = -EINVAL; |
| 259 | goto out; |
| 260 | } |
| 261 | |
| 262 | cmd_status = readl(addr: data + offsetof(struct cmd_chain_resp, status)); |
| 263 | xdna = job->hwctx->client->xdna; |
| 264 | XDNA_DBG(xdna, "Status 0x%x" , cmd_status); |
| 265 | if (cmd_status == AIE2_STATUS_SUCCESS) { |
| 266 | amdxdna_cmd_set_state(abo: cmd_abo, s: ERT_CMD_STATE_COMPLETED); |
| 267 | goto out; |
| 268 | } |
| 269 | |
| 270 | /* Slow path to handle error, read from ringbuf on BAR */ |
| 271 | fail_cmd_idx = readl(addr: data + offsetof(struct cmd_chain_resp, fail_cmd_idx)); |
| 272 | fail_cmd_status = readl(addr: data + offsetof(struct cmd_chain_resp, fail_cmd_status)); |
| 273 | XDNA_DBG(xdna, "Failed cmd idx %d, status 0x%x" , |
| 274 | fail_cmd_idx, fail_cmd_status); |
| 275 | |
| 276 | if (fail_cmd_status == AIE2_STATUS_SUCCESS) { |
| 277 | amdxdna_cmd_set_state(abo: cmd_abo, s: ERT_CMD_STATE_ABORT); |
| 278 | ret = -EINVAL; |
| 279 | goto out; |
| 280 | } |
| 281 | amdxdna_cmd_set_state(abo: cmd_abo, s: fail_cmd_status); |
| 282 | |
| 283 | if (amdxdna_cmd_get_op(abo: cmd_abo) == ERT_CMD_CHAIN) { |
| 284 | struct amdxdna_cmd_chain *cc = amdxdna_cmd_get_payload(abo: cmd_abo, NULL); |
| 285 | |
| 286 | cc->error_index = fail_cmd_idx; |
| 287 | if (cc->error_index >= cc->command_count) |
| 288 | cc->error_index = 0; |
| 289 | } |
| 290 | out: |
| 291 | aie2_sched_notify(job); |
| 292 | return ret; |
| 293 | } |
| 294 | |
| 295 | static struct dma_fence * |
| 296 | aie2_sched_job_run(struct drm_sched_job *sched_job) |
| 297 | { |
| 298 | struct amdxdna_sched_job *job = drm_job_to_xdna_job(sched_job); |
| 299 | struct amdxdna_gem_obj *cmd_abo = job->cmd_bo; |
| 300 | struct amdxdna_hwctx *hwctx = job->hwctx; |
| 301 | struct dma_fence *fence; |
| 302 | int ret; |
| 303 | |
| 304 | if (!mmget_not_zero(mm: job->mm)) |
| 305 | return ERR_PTR(error: -ESRCH); |
| 306 | |
| 307 | kref_get(kref: &job->refcnt); |
| 308 | fence = dma_fence_get(fence: job->fence); |
| 309 | |
| 310 | if (unlikely(!cmd_abo)) { |
| 311 | ret = aie2_sync_bo(hwctx, job, notify_cb: aie2_sched_nocmd_resp_handler); |
| 312 | goto out; |
| 313 | } |
| 314 | |
| 315 | amdxdna_cmd_set_state(abo: cmd_abo, s: ERT_CMD_STATE_NEW); |
| 316 | |
| 317 | if (amdxdna_cmd_get_op(abo: cmd_abo) == ERT_CMD_CHAIN) |
| 318 | ret = aie2_cmdlist_multi_execbuf(hwctx, job, notify_cb: aie2_sched_cmdlist_resp_handler); |
| 319 | else if (force_cmdlist) |
| 320 | ret = aie2_cmdlist_single_execbuf(hwctx, job, notify_cb: aie2_sched_cmdlist_resp_handler); |
| 321 | else |
| 322 | ret = aie2_execbuf(hwctx, job, notify_cb: aie2_sched_resp_handler); |
| 323 | |
| 324 | out: |
| 325 | if (ret) { |
| 326 | dma_fence_put(fence: job->fence); |
| 327 | aie2_job_put(job); |
| 328 | mmput(job->mm); |
| 329 | fence = ERR_PTR(error: ret); |
| 330 | } |
| 331 | trace_xdna_job(sched_job, name: hwctx->name, str: "sent to device" , seq: job->seq); |
| 332 | |
| 333 | return fence; |
| 334 | } |
| 335 | |
| 336 | static void aie2_sched_job_free(struct drm_sched_job *sched_job) |
| 337 | { |
| 338 | struct amdxdna_sched_job *job = drm_job_to_xdna_job(sched_job); |
| 339 | struct amdxdna_hwctx *hwctx = job->hwctx; |
| 340 | |
| 341 | trace_xdna_job(sched_job, name: hwctx->name, str: "job free" , seq: job->seq); |
| 342 | if (!job->job_done) |
| 343 | up(sem: &hwctx->priv->job_sem); |
| 344 | |
| 345 | drm_sched_job_cleanup(job: sched_job); |
| 346 | aie2_job_put(job); |
| 347 | } |
| 348 | |
| 349 | static enum drm_gpu_sched_stat |
| 350 | aie2_sched_job_timedout(struct drm_sched_job *sched_job) |
| 351 | { |
| 352 | struct amdxdna_sched_job *job = drm_job_to_xdna_job(sched_job); |
| 353 | struct amdxdna_hwctx *hwctx = job->hwctx; |
| 354 | struct amdxdna_dev *xdna; |
| 355 | |
| 356 | xdna = hwctx->client->xdna; |
| 357 | trace_xdna_job(sched_job, name: hwctx->name, str: "job timedout" , seq: job->seq); |
| 358 | mutex_lock(&xdna->dev_lock); |
| 359 | aie2_hwctx_stop(xdna, hwctx, bad_job: sched_job); |
| 360 | |
| 361 | aie2_hwctx_restart(xdna, hwctx); |
| 362 | mutex_unlock(lock: &xdna->dev_lock); |
| 363 | |
| 364 | return DRM_GPU_SCHED_STAT_NOMINAL; |
| 365 | } |
| 366 | |
| 367 | static const struct drm_sched_backend_ops sched_ops = { |
| 368 | .run_job = aie2_sched_job_run, |
| 369 | .free_job = aie2_sched_job_free, |
| 370 | .timedout_job = aie2_sched_job_timedout, |
| 371 | }; |
| 372 | |
| 373 | static int aie2_hwctx_col_list(struct amdxdna_hwctx *hwctx) |
| 374 | { |
| 375 | struct amdxdna_dev *xdna = hwctx->client->xdna; |
| 376 | struct amdxdna_dev_hdl *ndev; |
| 377 | int start, end, first, last; |
| 378 | u32 width = 1, entries = 0; |
| 379 | int i; |
| 380 | |
| 381 | if (!hwctx->num_tiles) { |
| 382 | XDNA_ERR(xdna, "Number of tiles is zero" ); |
| 383 | return -EINVAL; |
| 384 | } |
| 385 | |
| 386 | ndev = xdna->dev_handle; |
| 387 | if (unlikely(!ndev->metadata.core.row_count)) { |
| 388 | XDNA_WARN(xdna, "Core tile row count is zero" ); |
| 389 | return -EINVAL; |
| 390 | } |
| 391 | |
| 392 | hwctx->num_col = hwctx->num_tiles / ndev->metadata.core.row_count; |
| 393 | if (!hwctx->num_col || hwctx->num_col > ndev->total_col) { |
| 394 | XDNA_ERR(xdna, "Invalid num_col %d" , hwctx->num_col); |
| 395 | return -EINVAL; |
| 396 | } |
| 397 | |
| 398 | if (ndev->priv->col_align == COL_ALIGN_NATURE) |
| 399 | width = hwctx->num_col; |
| 400 | |
| 401 | /* |
| 402 | * In range [start, end], find out columns that is multiple of width. |
| 403 | * 'first' is the first column, |
| 404 | * 'last' is the last column, |
| 405 | * 'entries' is the total number of columns. |
| 406 | */ |
| 407 | start = xdna->dev_info->first_col; |
| 408 | end = ndev->total_col - hwctx->num_col; |
| 409 | if (start > 0 && end == 0) { |
| 410 | XDNA_DBG(xdna, "Force start from col 0" ); |
| 411 | start = 0; |
| 412 | } |
| 413 | first = start + (width - start % width) % width; |
| 414 | last = end - end % width; |
| 415 | if (last >= first) |
| 416 | entries = (last - first) / width + 1; |
| 417 | XDNA_DBG(xdna, "start %d end %d first %d last %d" , |
| 418 | start, end, first, last); |
| 419 | |
| 420 | if (unlikely(!entries)) { |
| 421 | XDNA_ERR(xdna, "Start %d end %d width %d" , |
| 422 | start, end, width); |
| 423 | return -EINVAL; |
| 424 | } |
| 425 | |
| 426 | hwctx->col_list = kmalloc_array(entries, sizeof(*hwctx->col_list), GFP_KERNEL); |
| 427 | if (!hwctx->col_list) |
| 428 | return -ENOMEM; |
| 429 | |
| 430 | hwctx->col_list_len = entries; |
| 431 | hwctx->col_list[0] = first; |
| 432 | for (i = 1; i < entries; i++) |
| 433 | hwctx->col_list[i] = hwctx->col_list[i - 1] + width; |
| 434 | |
| 435 | print_hex_dump_debug("col_list: " , DUMP_PREFIX_OFFSET, 16, 4, hwctx->col_list, |
| 436 | entries * sizeof(*hwctx->col_list), false); |
| 437 | return 0; |
| 438 | } |
| 439 | |
| 440 | static int aie2_alloc_resource(struct amdxdna_hwctx *hwctx) |
| 441 | { |
| 442 | struct amdxdna_dev *xdna = hwctx->client->xdna; |
| 443 | struct alloc_requests *xrs_req; |
| 444 | int ret; |
| 445 | |
| 446 | xrs_req = kzalloc(sizeof(*xrs_req), GFP_KERNEL); |
| 447 | if (!xrs_req) |
| 448 | return -ENOMEM; |
| 449 | |
| 450 | xrs_req->cdo.start_cols = hwctx->col_list; |
| 451 | xrs_req->cdo.cols_len = hwctx->col_list_len; |
| 452 | xrs_req->cdo.ncols = hwctx->num_col; |
| 453 | xrs_req->cdo.qos_cap.opc = hwctx->max_opc; |
| 454 | |
| 455 | xrs_req->rqos.gops = hwctx->qos.gops; |
| 456 | xrs_req->rqos.fps = hwctx->qos.fps; |
| 457 | xrs_req->rqos.dma_bw = hwctx->qos.dma_bandwidth; |
| 458 | xrs_req->rqos.latency = hwctx->qos.latency; |
| 459 | xrs_req->rqos.exec_time = hwctx->qos.frame_exec_time; |
| 460 | xrs_req->rqos.priority = hwctx->qos.priority; |
| 461 | |
| 462 | xrs_req->rid = (uintptr_t)hwctx; |
| 463 | |
| 464 | ret = xrs_allocate_resource(hdl: xdna->xrs_hdl, req: xrs_req, cb_arg: hwctx); |
| 465 | if (ret) |
| 466 | XDNA_ERR(xdna, "Allocate AIE resource failed, ret %d" , ret); |
| 467 | |
| 468 | kfree(objp: xrs_req); |
| 469 | return ret; |
| 470 | } |
| 471 | |
| 472 | static void aie2_release_resource(struct amdxdna_hwctx *hwctx) |
| 473 | { |
| 474 | struct amdxdna_dev *xdna = hwctx->client->xdna; |
| 475 | int ret; |
| 476 | |
| 477 | ret = xrs_release_resource(hdl: xdna->xrs_hdl, rid: (uintptr_t)hwctx); |
| 478 | if (ret) |
| 479 | XDNA_ERR(xdna, "Release AIE resource failed, ret %d" , ret); |
| 480 | } |
| 481 | |
| 482 | static int aie2_ctx_syncobj_create(struct amdxdna_hwctx *hwctx) |
| 483 | { |
| 484 | struct amdxdna_dev *xdna = hwctx->client->xdna; |
| 485 | struct drm_file *filp = hwctx->client->filp; |
| 486 | struct drm_syncobj *syncobj; |
| 487 | u32 hdl; |
| 488 | int ret; |
| 489 | |
| 490 | hwctx->syncobj_hdl = AMDXDNA_INVALID_FENCE_HANDLE; |
| 491 | |
| 492 | ret = drm_syncobj_create(out_syncobj: &syncobj, flags: 0, NULL); |
| 493 | if (ret) { |
| 494 | XDNA_ERR(xdna, "Create ctx syncobj failed, ret %d" , ret); |
| 495 | return ret; |
| 496 | } |
| 497 | ret = drm_syncobj_get_handle(file_private: filp, syncobj, handle: &hdl); |
| 498 | if (ret) { |
| 499 | drm_syncobj_put(obj: syncobj); |
| 500 | XDNA_ERR(xdna, "Create ctx syncobj handle failed, ret %d" , ret); |
| 501 | return ret; |
| 502 | } |
| 503 | hwctx->priv->syncobj = syncobj; |
| 504 | hwctx->syncobj_hdl = hdl; |
| 505 | |
| 506 | return 0; |
| 507 | } |
| 508 | |
| 509 | static void aie2_ctx_syncobj_destroy(struct amdxdna_hwctx *hwctx) |
| 510 | { |
| 511 | /* |
| 512 | * The syncobj_hdl is owned by user space and will be cleaned up |
| 513 | * separately. |
| 514 | */ |
| 515 | drm_syncobj_put(obj: hwctx->priv->syncobj); |
| 516 | } |
| 517 | |
| 518 | int aie2_hwctx_init(struct amdxdna_hwctx *hwctx) |
| 519 | { |
| 520 | struct amdxdna_client *client = hwctx->client; |
| 521 | struct amdxdna_dev *xdna = client->xdna; |
| 522 | const struct drm_sched_init_args args = { |
| 523 | .ops = &sched_ops, |
| 524 | .num_rqs = DRM_SCHED_PRIORITY_COUNT, |
| 525 | .credit_limit = HWCTX_MAX_CMDS, |
| 526 | .timeout = msecs_to_jiffies(HWCTX_MAX_TIMEOUT), |
| 527 | .name = hwctx->name, |
| 528 | .dev = xdna->ddev.dev, |
| 529 | }; |
| 530 | struct drm_gpu_scheduler *sched; |
| 531 | struct amdxdna_hwctx_priv *priv; |
| 532 | struct amdxdna_gem_obj *heap; |
| 533 | struct amdxdna_dev_hdl *ndev; |
| 534 | int i, ret; |
| 535 | |
| 536 | priv = kzalloc(sizeof(*hwctx->priv), GFP_KERNEL); |
| 537 | if (!priv) |
| 538 | return -ENOMEM; |
| 539 | hwctx->priv = priv; |
| 540 | |
| 541 | mutex_lock(&client->mm_lock); |
| 542 | heap = client->dev_heap; |
| 543 | if (!heap) { |
| 544 | XDNA_ERR(xdna, "The client dev heap object not exist" ); |
| 545 | mutex_unlock(lock: &client->mm_lock); |
| 546 | ret = -ENOENT; |
| 547 | goto free_priv; |
| 548 | } |
| 549 | drm_gem_object_get(to_gobj(heap)); |
| 550 | mutex_unlock(lock: &client->mm_lock); |
| 551 | priv->heap = heap; |
| 552 | sema_init(sem: &priv->job_sem, HWCTX_MAX_CMDS); |
| 553 | |
| 554 | ret = amdxdna_gem_pin(abo: heap); |
| 555 | if (ret) { |
| 556 | XDNA_ERR(xdna, "Dev heap pin failed, ret %d" , ret); |
| 557 | goto put_heap; |
| 558 | } |
| 559 | |
| 560 | for (i = 0; i < ARRAY_SIZE(priv->cmd_buf); i++) { |
| 561 | struct amdxdna_gem_obj *abo; |
| 562 | struct amdxdna_drm_create_bo args = { |
| 563 | .flags = 0, |
| 564 | .type = AMDXDNA_BO_DEV, |
| 565 | .vaddr = 0, |
| 566 | .size = MAX_CHAIN_CMDBUF_SIZE, |
| 567 | }; |
| 568 | |
| 569 | abo = amdxdna_drm_alloc_dev_bo(dev: &xdna->ddev, args: &args, filp: client->filp, use_vmap: true); |
| 570 | if (IS_ERR(ptr: abo)) { |
| 571 | ret = PTR_ERR(ptr: abo); |
| 572 | goto free_cmd_bufs; |
| 573 | } |
| 574 | |
| 575 | XDNA_DBG(xdna, "Command buf %d addr 0x%llx size 0x%lx" , |
| 576 | i, abo->mem.dev_addr, abo->mem.size); |
| 577 | priv->cmd_buf[i] = abo; |
| 578 | } |
| 579 | |
| 580 | sched = &priv->sched; |
| 581 | mutex_init(&priv->io_lock); |
| 582 | |
| 583 | fs_reclaim_acquire(GFP_KERNEL); |
| 584 | might_lock(&priv->io_lock); |
| 585 | fs_reclaim_release(GFP_KERNEL); |
| 586 | |
| 587 | ret = drm_sched_init(sched, args: &args); |
| 588 | if (ret) { |
| 589 | XDNA_ERR(xdna, "Failed to init DRM scheduler. ret %d" , ret); |
| 590 | goto free_cmd_bufs; |
| 591 | } |
| 592 | |
| 593 | ret = drm_sched_entity_init(entity: &priv->entity, priority: DRM_SCHED_PRIORITY_NORMAL, |
| 594 | sched_list: &sched, num_sched_list: 1, NULL); |
| 595 | if (ret) { |
| 596 | XDNA_ERR(xdna, "Failed to initial sched entiry. ret %d" , ret); |
| 597 | goto free_sched; |
| 598 | } |
| 599 | |
| 600 | ret = aie2_hwctx_col_list(hwctx); |
| 601 | if (ret) { |
| 602 | XDNA_ERR(xdna, "Create col list failed, ret %d" , ret); |
| 603 | goto free_entity; |
| 604 | } |
| 605 | |
| 606 | ret = aie2_alloc_resource(hwctx); |
| 607 | if (ret) { |
| 608 | XDNA_ERR(xdna, "Alloc hw resource failed, ret %d" , ret); |
| 609 | goto free_col_list; |
| 610 | } |
| 611 | |
| 612 | ret = aie2_map_host_buf(ndev: xdna->dev_handle, context_id: hwctx->fw_ctx_id, |
| 613 | addr: heap->mem.userptr, size: heap->mem.size); |
| 614 | if (ret) { |
| 615 | XDNA_ERR(xdna, "Map host buffer failed, ret %d" , ret); |
| 616 | goto release_resource; |
| 617 | } |
| 618 | |
| 619 | ret = aie2_ctx_syncobj_create(hwctx); |
| 620 | if (ret) { |
| 621 | XDNA_ERR(xdna, "Create syncobj failed, ret %d" , ret); |
| 622 | goto release_resource; |
| 623 | } |
| 624 | |
| 625 | hwctx->status = HWCTX_STAT_INIT; |
| 626 | ndev = xdna->dev_handle; |
| 627 | ndev->hwctx_num++; |
| 628 | init_waitqueue_head(&priv->job_free_wq); |
| 629 | |
| 630 | XDNA_DBG(xdna, "hwctx %s init completed" , hwctx->name); |
| 631 | |
| 632 | return 0; |
| 633 | |
| 634 | release_resource: |
| 635 | aie2_release_resource(hwctx); |
| 636 | free_col_list: |
| 637 | kfree(objp: hwctx->col_list); |
| 638 | free_entity: |
| 639 | drm_sched_entity_destroy(entity: &priv->entity); |
| 640 | free_sched: |
| 641 | drm_sched_fini(sched: &priv->sched); |
| 642 | free_cmd_bufs: |
| 643 | for (i = 0; i < ARRAY_SIZE(priv->cmd_buf); i++) { |
| 644 | if (!priv->cmd_buf[i]) |
| 645 | continue; |
| 646 | drm_gem_object_put(to_gobj(priv->cmd_buf[i])); |
| 647 | } |
| 648 | amdxdna_gem_unpin(abo: heap); |
| 649 | put_heap: |
| 650 | drm_gem_object_put(to_gobj(heap)); |
| 651 | free_priv: |
| 652 | kfree(objp: priv); |
| 653 | return ret; |
| 654 | } |
| 655 | |
| 656 | void aie2_hwctx_fini(struct amdxdna_hwctx *hwctx) |
| 657 | { |
| 658 | struct amdxdna_dev_hdl *ndev; |
| 659 | struct amdxdna_dev *xdna; |
| 660 | int idx; |
| 661 | |
| 662 | xdna = hwctx->client->xdna; |
| 663 | ndev = xdna->dev_handle; |
| 664 | ndev->hwctx_num--; |
| 665 | |
| 666 | XDNA_DBG(xdna, "%s sequence number %lld" , hwctx->name, hwctx->priv->seq); |
| 667 | drm_sched_entity_destroy(entity: &hwctx->priv->entity); |
| 668 | |
| 669 | aie2_hwctx_wait_for_idle(hwctx); |
| 670 | |
| 671 | /* Request fw to destroy hwctx and cancel the rest pending requests */ |
| 672 | aie2_release_resource(hwctx); |
| 673 | |
| 674 | /* Wait for all submitted jobs to be completed or canceled */ |
| 675 | wait_event(hwctx->priv->job_free_wq, |
| 676 | atomic64_read(&hwctx->job_submit_cnt) == |
| 677 | atomic64_read(&hwctx->job_free_cnt)); |
| 678 | |
| 679 | drm_sched_fini(sched: &hwctx->priv->sched); |
| 680 | aie2_ctx_syncobj_destroy(hwctx); |
| 681 | |
| 682 | for (idx = 0; idx < ARRAY_SIZE(hwctx->priv->cmd_buf); idx++) |
| 683 | drm_gem_object_put(to_gobj(hwctx->priv->cmd_buf[idx])); |
| 684 | amdxdna_gem_unpin(abo: hwctx->priv->heap); |
| 685 | drm_gem_object_put(to_gobj(hwctx->priv->heap)); |
| 686 | |
| 687 | mutex_destroy(lock: &hwctx->priv->io_lock); |
| 688 | kfree(objp: hwctx->col_list); |
| 689 | kfree(objp: hwctx->priv); |
| 690 | kfree(objp: hwctx->cus); |
| 691 | } |
| 692 | |
| 693 | static int aie2_hwctx_cu_config(struct amdxdna_hwctx *hwctx, void *buf, u32 size) |
| 694 | { |
| 695 | struct amdxdna_hwctx_param_config_cu *config = buf; |
| 696 | struct amdxdna_dev *xdna = hwctx->client->xdna; |
| 697 | u32 total_size; |
| 698 | int ret; |
| 699 | |
| 700 | XDNA_DBG(xdna, "Config %d CU to %s" , config->num_cus, hwctx->name); |
| 701 | if (XDNA_MBZ_DBG(xdna, config->pad, sizeof(config->pad))) |
| 702 | return -EINVAL; |
| 703 | |
| 704 | if (hwctx->status != HWCTX_STAT_INIT) { |
| 705 | XDNA_ERR(xdna, "Not support re-config CU" ); |
| 706 | return -EINVAL; |
| 707 | } |
| 708 | |
| 709 | if (!config->num_cus) { |
| 710 | XDNA_ERR(xdna, "Number of CU is zero" ); |
| 711 | return -EINVAL; |
| 712 | } |
| 713 | |
| 714 | total_size = struct_size(config, cu_configs, config->num_cus); |
| 715 | if (total_size > size) { |
| 716 | XDNA_ERR(xdna, "CU config larger than size" ); |
| 717 | return -EINVAL; |
| 718 | } |
| 719 | |
| 720 | hwctx->cus = kmemdup(config, total_size, GFP_KERNEL); |
| 721 | if (!hwctx->cus) |
| 722 | return -ENOMEM; |
| 723 | |
| 724 | ret = aie2_config_cu(hwctx); |
| 725 | if (ret) { |
| 726 | XDNA_ERR(xdna, "Config CU to firmware failed, ret %d" , ret); |
| 727 | goto free_cus; |
| 728 | } |
| 729 | |
| 730 | wmb(); /* To avoid locking in command submit when check status */ |
| 731 | hwctx->status = HWCTX_STAT_READY; |
| 732 | |
| 733 | return 0; |
| 734 | |
| 735 | free_cus: |
| 736 | kfree(objp: hwctx->cus); |
| 737 | hwctx->cus = NULL; |
| 738 | return ret; |
| 739 | } |
| 740 | |
| 741 | int aie2_hwctx_config(struct amdxdna_hwctx *hwctx, u32 type, u64 value, void *buf, u32 size) |
| 742 | { |
| 743 | struct amdxdna_dev *xdna = hwctx->client->xdna; |
| 744 | |
| 745 | drm_WARN_ON(&xdna->ddev, !mutex_is_locked(&xdna->dev_lock)); |
| 746 | switch (type) { |
| 747 | case DRM_AMDXDNA_HWCTX_CONFIG_CU: |
| 748 | return aie2_hwctx_cu_config(hwctx, buf, size); |
| 749 | case DRM_AMDXDNA_HWCTX_ASSIGN_DBG_BUF: |
| 750 | case DRM_AMDXDNA_HWCTX_REMOVE_DBG_BUF: |
| 751 | return -EOPNOTSUPP; |
| 752 | default: |
| 753 | XDNA_DBG(xdna, "Not supported type %d" , type); |
| 754 | return -EOPNOTSUPP; |
| 755 | } |
| 756 | } |
| 757 | |
| 758 | static int aie2_populate_range(struct amdxdna_gem_obj *abo) |
| 759 | { |
| 760 | struct amdxdna_dev *xdna = to_xdna_dev(to_gobj(abo)->dev); |
| 761 | struct amdxdna_umap *mapp; |
| 762 | unsigned long timeout; |
| 763 | struct mm_struct *mm; |
| 764 | bool found; |
| 765 | int ret; |
| 766 | |
| 767 | timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT); |
| 768 | again: |
| 769 | found = false; |
| 770 | down_write(sem: &xdna->notifier_lock); |
| 771 | list_for_each_entry(mapp, &abo->mem.umap_list, node) { |
| 772 | if (mapp->invalid) { |
| 773 | found = true; |
| 774 | break; |
| 775 | } |
| 776 | } |
| 777 | |
| 778 | if (!found) { |
| 779 | abo->mem.map_invalid = false; |
| 780 | up_write(sem: &xdna->notifier_lock); |
| 781 | return 0; |
| 782 | } |
| 783 | kref_get(kref: &mapp->refcnt); |
| 784 | up_write(sem: &xdna->notifier_lock); |
| 785 | |
| 786 | XDNA_DBG(xdna, "populate memory range %lx %lx" , |
| 787 | mapp->vma->vm_start, mapp->vma->vm_end); |
| 788 | mm = mapp->notifier.mm; |
| 789 | if (!mmget_not_zero(mm)) { |
| 790 | amdxdna_umap_put(mapp); |
| 791 | return -EFAULT; |
| 792 | } |
| 793 | |
| 794 | mapp->range.notifier_seq = mmu_interval_read_begin(interval_sub: &mapp->notifier); |
| 795 | mmap_read_lock(mm); |
| 796 | ret = hmm_range_fault(range: &mapp->range); |
| 797 | mmap_read_unlock(mm); |
| 798 | if (ret) { |
| 799 | if (time_after(jiffies, timeout)) { |
| 800 | ret = -ETIME; |
| 801 | goto put_mm; |
| 802 | } |
| 803 | |
| 804 | if (ret == -EBUSY) { |
| 805 | amdxdna_umap_put(mapp); |
| 806 | goto again; |
| 807 | } |
| 808 | |
| 809 | goto put_mm; |
| 810 | } |
| 811 | |
| 812 | down_write(sem: &xdna->notifier_lock); |
| 813 | if (mmu_interval_read_retry(interval_sub: &mapp->notifier, seq: mapp->range.notifier_seq)) { |
| 814 | up_write(sem: &xdna->notifier_lock); |
| 815 | amdxdna_umap_put(mapp); |
| 816 | goto again; |
| 817 | } |
| 818 | mapp->invalid = false; |
| 819 | up_write(sem: &xdna->notifier_lock); |
| 820 | amdxdna_umap_put(mapp); |
| 821 | goto again; |
| 822 | |
| 823 | put_mm: |
| 824 | amdxdna_umap_put(mapp); |
| 825 | mmput(mm); |
| 826 | return ret; |
| 827 | } |
| 828 | |
| 829 | int aie2_cmd_submit(struct amdxdna_hwctx *hwctx, struct amdxdna_sched_job *job, u64 *seq) |
| 830 | { |
| 831 | struct amdxdna_dev *xdna = hwctx->client->xdna; |
| 832 | struct ww_acquire_ctx acquire_ctx; |
| 833 | struct dma_fence_chain *chain; |
| 834 | struct amdxdna_gem_obj *abo; |
| 835 | unsigned long timeout = 0; |
| 836 | int ret, i; |
| 837 | |
| 838 | ret = down_interruptible(sem: &hwctx->priv->job_sem); |
| 839 | if (ret) { |
| 840 | XDNA_ERR(xdna, "Grab job sem failed, ret %d" , ret); |
| 841 | return ret; |
| 842 | } |
| 843 | |
| 844 | chain = dma_fence_chain_alloc(); |
| 845 | if (!chain) { |
| 846 | XDNA_ERR(xdna, "Alloc fence chain failed" ); |
| 847 | ret = -ENOMEM; |
| 848 | goto up_sem; |
| 849 | } |
| 850 | |
| 851 | ret = drm_sched_job_init(job: &job->base, entity: &hwctx->priv->entity, credits: 1, owner: hwctx); |
| 852 | if (ret) { |
| 853 | XDNA_ERR(xdna, "DRM job init failed, ret %d" , ret); |
| 854 | goto free_chain; |
| 855 | } |
| 856 | |
| 857 | retry: |
| 858 | ret = drm_gem_lock_reservations(objs: job->bos, count: job->bo_cnt, acquire_ctx: &acquire_ctx); |
| 859 | if (ret) { |
| 860 | XDNA_WARN(xdna, "Failed to lock BOs, ret %d" , ret); |
| 861 | goto cleanup_job; |
| 862 | } |
| 863 | |
| 864 | for (i = 0; i < job->bo_cnt; i++) { |
| 865 | ret = dma_resv_reserve_fences(obj: job->bos[i]->resv, num_fences: 1); |
| 866 | if (ret) { |
| 867 | XDNA_WARN(xdna, "Failed to reserve fences %d" , ret); |
| 868 | drm_gem_unlock_reservations(objs: job->bos, count: job->bo_cnt, acquire_ctx: &acquire_ctx); |
| 869 | goto cleanup_job; |
| 870 | } |
| 871 | } |
| 872 | |
| 873 | down_read(sem: &xdna->notifier_lock); |
| 874 | for (i = 0; i < job->bo_cnt; i++) { |
| 875 | abo = to_xdna_obj(gobj: job->bos[i]); |
| 876 | if (abo->mem.map_invalid) { |
| 877 | up_read(sem: &xdna->notifier_lock); |
| 878 | drm_gem_unlock_reservations(objs: job->bos, count: job->bo_cnt, acquire_ctx: &acquire_ctx); |
| 879 | if (!timeout) { |
| 880 | timeout = jiffies + |
| 881 | msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT); |
| 882 | } else if (time_after(jiffies, timeout)) { |
| 883 | ret = -ETIME; |
| 884 | goto cleanup_job; |
| 885 | } |
| 886 | |
| 887 | ret = aie2_populate_range(abo); |
| 888 | if (ret) |
| 889 | goto cleanup_job; |
| 890 | goto retry; |
| 891 | } |
| 892 | } |
| 893 | |
| 894 | mutex_lock(&hwctx->priv->io_lock); |
| 895 | drm_sched_job_arm(job: &job->base); |
| 896 | job->out_fence = dma_fence_get(fence: &job->base.s_fence->finished); |
| 897 | for (i = 0; i < job->bo_cnt; i++) |
| 898 | dma_resv_add_fence(obj: job->bos[i]->resv, fence: job->out_fence, usage: DMA_RESV_USAGE_WRITE); |
| 899 | job->seq = hwctx->priv->seq++; |
| 900 | kref_get(kref: &job->refcnt); |
| 901 | drm_sched_entity_push_job(sched_job: &job->base); |
| 902 | |
| 903 | *seq = job->seq; |
| 904 | drm_syncobj_add_point(syncobj: hwctx->priv->syncobj, chain, fence: job->out_fence, point: *seq); |
| 905 | mutex_unlock(lock: &hwctx->priv->io_lock); |
| 906 | |
| 907 | up_read(sem: &xdna->notifier_lock); |
| 908 | drm_gem_unlock_reservations(objs: job->bos, count: job->bo_cnt, acquire_ctx: &acquire_ctx); |
| 909 | |
| 910 | aie2_job_put(job); |
| 911 | atomic64_inc(v: &hwctx->job_submit_cnt); |
| 912 | |
| 913 | return 0; |
| 914 | |
| 915 | cleanup_job: |
| 916 | drm_sched_job_cleanup(job: &job->base); |
| 917 | free_chain: |
| 918 | dma_fence_chain_free(chain); |
| 919 | up_sem: |
| 920 | up(sem: &hwctx->priv->job_sem); |
| 921 | job->job_done = true; |
| 922 | return ret; |
| 923 | } |
| 924 | |
| 925 | void aie2_hmm_invalidate(struct amdxdna_gem_obj *abo, |
| 926 | unsigned long cur_seq) |
| 927 | { |
| 928 | struct amdxdna_dev *xdna = to_xdna_dev(to_gobj(abo)->dev); |
| 929 | struct drm_gem_object *gobj = to_gobj(abo); |
| 930 | long ret; |
| 931 | |
| 932 | ret = dma_resv_wait_timeout(obj: gobj->resv, usage: DMA_RESV_USAGE_BOOKKEEP, |
| 933 | intr: true, MAX_SCHEDULE_TIMEOUT); |
| 934 | if (!ret || ret == -ERESTARTSYS) |
| 935 | XDNA_ERR(xdna, "Failed to wait for bo, ret %ld" , ret); |
| 936 | } |
| 937 | |