1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2022 Intel Corporation 4 */ 5 6 #include "xe_guc_ct.h" 7 8 #include <linux/bitfield.h> 9 #include <linux/circ_buf.h> 10 #include <linux/delay.h> 11 #include <linux/fault-inject.h> 12 13 #include <kunit/static_stub.h> 14 15 #include <drm/drm_managed.h> 16 17 #include "abi/guc_actions_abi.h" 18 #include "abi/guc_actions_sriov_abi.h" 19 #include "abi/guc_klvs_abi.h" 20 #include "xe_bo.h" 21 #include "xe_devcoredump.h" 22 #include "xe_device.h" 23 #include "xe_gt.h" 24 #include "xe_gt_printk.h" 25 #include "xe_gt_sriov_pf_control.h" 26 #include "xe_gt_sriov_pf_monitor.h" 27 #include "xe_guc.h" 28 #include "xe_guc_log.h" 29 #include "xe_guc_pagefault.h" 30 #include "xe_guc_relay.h" 31 #include "xe_guc_submit.h" 32 #include "xe_guc_tlb_inval.h" 33 #include "xe_map.h" 34 #include "xe_pm.h" 35 #include "xe_sriov_vf.h" 36 #include "xe_trace_guc.h" 37 38 static void receive_g2h(struct xe_guc_ct *ct); 39 static void g2h_worker_func(struct work_struct *w); 40 static void safe_mode_worker_func(struct work_struct *w); 41 static void ct_exit_safe_mode(struct xe_guc_ct *ct); 42 static void guc_ct_change_state(struct xe_guc_ct *ct, 43 enum xe_guc_ct_state state); 44 45 static struct xe_guc *ct_to_guc(struct xe_guc_ct *ct) 46 { 47 return container_of(ct, struct xe_guc, ct); 48 } 49 50 static struct xe_gt *ct_to_gt(struct xe_guc_ct *ct) 51 { 52 return container_of(ct, struct xe_gt, uc.guc.ct); 53 } 54 55 static struct xe_device *ct_to_xe(struct xe_guc_ct *ct) 56 { 57 return gt_to_xe(ct_to_gt(ct)); 58 } 59 60 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG) 61 enum { 62 /* Internal states, not error conditions */ 63 CT_DEAD_STATE_REARM, /* 0x0001 */ 64 CT_DEAD_STATE_CAPTURE, /* 0x0002 */ 65 66 /* Error conditions */ 67 CT_DEAD_SETUP, /* 0x0004 */ 68 CT_DEAD_H2G_WRITE, /* 0x0008 */ 69 CT_DEAD_H2G_HAS_ROOM, /* 0x0010 */ 70 CT_DEAD_G2H_READ, /* 0x0020 */ 71 CT_DEAD_G2H_RECV, /* 0x0040 */ 72 CT_DEAD_G2H_RELEASE, /* 0x0080 */ 73 CT_DEAD_DEADLOCK, /* 0x0100 */ 74 CT_DEAD_PROCESS_FAILED, /* 0x0200 */ 75 CT_DEAD_FAST_G2H, /* 0x0400 */ 76 CT_DEAD_PARSE_G2H_RESPONSE, /* 0x0800 */ 77 CT_DEAD_PARSE_G2H_UNKNOWN, /* 0x1000 */ 78 CT_DEAD_PARSE_G2H_ORIGIN, /* 0x2000 */ 79 CT_DEAD_PARSE_G2H_TYPE, /* 0x4000 */ 80 CT_DEAD_CRASH, /* 0x8000 */ 81 }; 82 83 static void ct_dead_worker_func(struct work_struct *w); 84 static void ct_dead_capture(struct xe_guc_ct *ct, struct guc_ctb *ctb, u32 reason_code); 85 86 static void ct_dead_fini(struct xe_guc_ct *ct) 87 { 88 cancel_work_sync(&ct->dead.worker); 89 } 90 91 static void ct_dead_init(struct xe_guc_ct *ct) 92 { 93 spin_lock_init(&ct->dead.lock); 94 INIT_WORK(&ct->dead.worker, ct_dead_worker_func); 95 96 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_GUC) 97 stack_depot_init(); 98 #endif 99 } 100 101 static void fast_req_stack_save(struct xe_guc_ct *ct, unsigned int slot) 102 { 103 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_GUC) 104 unsigned long entries[SZ_32]; 105 unsigned int n; 106 107 n = stack_trace_save(entries, ARRAY_SIZE(entries), 1); 108 /* May be called under spinlock, so avoid sleeping */ 109 ct->fast_req[slot].stack = stack_depot_save(entries, n, GFP_NOWAIT); 110 #endif 111 } 112 113 static void fast_req_dump(struct xe_guc_ct *ct, u16 fence, unsigned int slot) 114 { 115 struct xe_gt *gt = ct_to_gt(ct); 116 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_GUC) 117 char *buf __cleanup(kfree) = kmalloc(SZ_4K, GFP_NOWAIT); 118 119 if (buf && stack_depot_snprint(ct->fast_req[slot].stack, buf, SZ_4K, 0)) 120 xe_gt_err(gt, "Fence 0x%x was used by action %#04x sent at:\n%s\n", 121 fence, ct->fast_req[slot].action, buf); 122 else 123 xe_gt_err(gt, "Fence 0x%x was used by action %#04x [failed to retrieve stack]\n", 124 fence, ct->fast_req[slot].action); 125 #else 126 xe_gt_err(gt, "Fence 0x%x was used by action %#04x\n", 127 fence, ct->fast_req[slot].action); 128 #endif 129 } 130 131 static void fast_req_report(struct xe_guc_ct *ct, u16 fence) 132 { 133 u16 fence_min = U16_MAX, fence_max = 0; 134 struct xe_gt *gt = ct_to_gt(ct); 135 unsigned int n; 136 137 lockdep_assert_held(&ct->lock); 138 139 for (n = 0; n < ARRAY_SIZE(ct->fast_req); n++) { 140 if (ct->fast_req[n].fence < fence_min) 141 fence_min = ct->fast_req[n].fence; 142 if (ct->fast_req[n].fence > fence_max) 143 fence_max = ct->fast_req[n].fence; 144 145 if (ct->fast_req[n].fence != fence) 146 continue; 147 148 return fast_req_dump(ct, fence, n); 149 } 150 151 xe_gt_warn(gt, "Fence 0x%x not found - tracking buffer wrapped? [range = 0x%x -> 0x%x, next = 0x%X]\n", 152 fence, fence_min, fence_max, ct->fence_seqno); 153 } 154 155 static void fast_req_track(struct xe_guc_ct *ct, u16 fence, u16 action) 156 { 157 unsigned int slot = fence % ARRAY_SIZE(ct->fast_req); 158 159 fast_req_stack_save(ct, slot); 160 ct->fast_req[slot].fence = fence; 161 ct->fast_req[slot].action = action; 162 } 163 164 #define CT_DEAD(ct, ctb, reason_code) ct_dead_capture((ct), (ctb), CT_DEAD_##reason_code) 165 166 #else 167 168 static void ct_dead_fini(struct xe_guc_ct *ct) { } 169 static void ct_dead_init(struct xe_guc_ct *ct) { } 170 171 static void fast_req_report(struct xe_guc_ct *ct, u16 fence) { } 172 static void fast_req_track(struct xe_guc_ct *ct, u16 fence, u16 action) { } 173 174 #define CT_DEAD(ct, ctb, reason) \ 175 do { \ 176 struct guc_ctb *_ctb = (ctb); \ 177 if (_ctb) \ 178 _ctb->info.broken = true; \ 179 } while (0) 180 181 #endif 182 183 /* Used when a CT send wants to block and / or receive data */ 184 struct g2h_fence { 185 u32 *response_buffer; 186 u32 seqno; 187 u32 response_data; 188 u16 response_len; 189 u16 error; 190 u16 hint; 191 u16 reason; 192 bool cancel; 193 bool retry; 194 bool fail; 195 bool done; 196 }; 197 198 static void g2h_fence_init(struct g2h_fence *g2h_fence, u32 *response_buffer) 199 { 200 memset(g2h_fence, 0, sizeof(*g2h_fence)); 201 g2h_fence->response_buffer = response_buffer; 202 g2h_fence->seqno = ~0x0; 203 } 204 205 static void g2h_fence_cancel(struct g2h_fence *g2h_fence) 206 { 207 g2h_fence->cancel = true; 208 g2h_fence->fail = true; 209 210 /* WRITE_ONCE pairs with READ_ONCEs in guc_ct_send_recv. */ 211 WRITE_ONCE(g2h_fence->done, true); 212 } 213 214 static bool g2h_fence_needs_alloc(struct g2h_fence *g2h_fence) 215 { 216 return g2h_fence->seqno == ~0x0; 217 } 218 219 /** 220 * DOC: GuC CTB Blob 221 * 222 * We allocate single blob to hold both CTB descriptors and buffers: 223 * 224 * +--------+-----------------------------------------------+------+ 225 * | offset | contents | size | 226 * +========+===============================================+======+ 227 * | 0x0000 | H2G CTB Descriptor (send) | | 228 * +--------+-----------------------------------------------+ 4K | 229 * | 0x0800 | G2H CTB Descriptor (g2h) | | 230 * +--------+-----------------------------------------------+------+ 231 * | 0x1000 | H2G CT Buffer (send) | n*4K | 232 * | | | | 233 * +--------+-----------------------------------------------+------+ 234 * | 0x1000 | G2H CT Buffer (g2h) | m*4K | 235 * | + n*4K | | | 236 * +--------+-----------------------------------------------+------+ 237 * 238 * Size of each ``CT Buffer`` must be multiple of 4K. 239 * We don't expect too many messages in flight at any time, unless we are 240 * using the GuC submission. In that case each request requires a minimum 241 * 2 dwords which gives us a maximum 256 queue'd requests. Hopefully this 242 * enough space to avoid backpressure on the driver. We increase the size 243 * of the receive buffer (relative to the send) to ensure a G2H response 244 * CTB has a landing spot. 245 * 246 * In addition to submissions, the G2H buffer needs to be able to hold 247 * enough space for recoverable page fault notifications. The number of 248 * page faults is interrupt driven and can be as much as the number of 249 * compute resources available. However, most of the actual work for these 250 * is in a separate page fault worker thread. Therefore we only need to 251 * make sure the queue has enough space to handle all of the submissions 252 * and responses and an extra buffer for incoming page faults. 253 */ 254 255 #define CTB_DESC_SIZE ALIGN(sizeof(struct guc_ct_buffer_desc), SZ_2K) 256 #define CTB_H2G_BUFFER_OFFSET (CTB_DESC_SIZE * 2) 257 #define CTB_H2G_BUFFER_SIZE (SZ_4K) 258 #define CTB_H2G_BUFFER_DWORDS (CTB_H2G_BUFFER_SIZE / sizeof(u32)) 259 #define CTB_G2H_BUFFER_SIZE (SZ_128K) 260 #define CTB_G2H_BUFFER_DWORDS (CTB_G2H_BUFFER_SIZE / sizeof(u32)) 261 #define G2H_ROOM_BUFFER_SIZE (CTB_G2H_BUFFER_SIZE / 2) 262 #define G2H_ROOM_BUFFER_DWORDS (CTB_G2H_BUFFER_DWORDS / 2) 263 264 /** 265 * xe_guc_ct_queue_proc_time_jiffies - Return maximum time to process a full 266 * CT command queue 267 * @ct: the &xe_guc_ct. Unused at this moment but will be used in the future. 268 * 269 * Observation is that a 4KiB buffer full of commands takes a little over a 270 * second to process. Use that to calculate maximum time to process a full CT 271 * command queue. 272 * 273 * Return: Maximum time to process a full CT queue in jiffies. 274 */ 275 long xe_guc_ct_queue_proc_time_jiffies(struct xe_guc_ct *ct) 276 { 277 BUILD_BUG_ON(!IS_ALIGNED(CTB_H2G_BUFFER_SIZE, SZ_4)); 278 return (CTB_H2G_BUFFER_SIZE / SZ_4K) * HZ; 279 } 280 281 static size_t guc_ct_size(void) 282 { 283 return CTB_H2G_BUFFER_OFFSET + CTB_H2G_BUFFER_SIZE + 284 CTB_G2H_BUFFER_SIZE; 285 } 286 287 static void guc_ct_fini(struct drm_device *drm, void *arg) 288 { 289 struct xe_guc_ct *ct = arg; 290 291 ct_dead_fini(ct); 292 ct_exit_safe_mode(ct); 293 destroy_workqueue(ct->g2h_wq); 294 xa_destroy(&ct->fence_lookup); 295 } 296 297 static void primelockdep(struct xe_guc_ct *ct) 298 { 299 if (!IS_ENABLED(CONFIG_LOCKDEP)) 300 return; 301 302 fs_reclaim_acquire(GFP_KERNEL); 303 might_lock(&ct->lock); 304 fs_reclaim_release(GFP_KERNEL); 305 } 306 307 int xe_guc_ct_init_noalloc(struct xe_guc_ct *ct) 308 { 309 struct xe_device *xe = ct_to_xe(ct); 310 struct xe_gt *gt = ct_to_gt(ct); 311 int err; 312 313 xe_gt_assert(gt, !(guc_ct_size() % PAGE_SIZE)); 314 315 err = drmm_mutex_init(&xe->drm, &ct->lock); 316 if (err) 317 return err; 318 319 primelockdep(ct); 320 321 ct->g2h_wq = alloc_ordered_workqueue("xe-g2h-wq", WQ_MEM_RECLAIM); 322 if (!ct->g2h_wq) 323 return -ENOMEM; 324 325 spin_lock_init(&ct->fast_lock); 326 xa_init(&ct->fence_lookup); 327 INIT_WORK(&ct->g2h_worker, g2h_worker_func); 328 INIT_DELAYED_WORK(&ct->safe_mode_worker, safe_mode_worker_func); 329 330 ct_dead_init(ct); 331 init_waitqueue_head(&ct->wq); 332 init_waitqueue_head(&ct->g2h_fence_wq); 333 334 err = drmm_add_action_or_reset(&xe->drm, guc_ct_fini, ct); 335 if (err) 336 return err; 337 338 xe_gt_assert(gt, ct->state == XE_GUC_CT_STATE_NOT_INITIALIZED); 339 ct->state = XE_GUC_CT_STATE_DISABLED; 340 return 0; 341 } 342 ALLOW_ERROR_INJECTION(xe_guc_ct_init_noalloc, ERRNO); /* See xe_pci_probe() */ 343 344 static void guc_action_disable_ct(void *arg) 345 { 346 struct xe_guc_ct *ct = arg; 347 348 guc_ct_change_state(ct, XE_GUC_CT_STATE_DISABLED); 349 } 350 351 int xe_guc_ct_init(struct xe_guc_ct *ct) 352 { 353 struct xe_device *xe = ct_to_xe(ct); 354 struct xe_gt *gt = ct_to_gt(ct); 355 struct xe_tile *tile = gt_to_tile(gt); 356 struct xe_bo *bo; 357 358 bo = xe_managed_bo_create_pin_map(xe, tile, guc_ct_size(), 359 XE_BO_FLAG_SYSTEM | 360 XE_BO_FLAG_GGTT | 361 XE_BO_FLAG_GGTT_INVALIDATE | 362 XE_BO_FLAG_PINNED_NORESTORE); 363 if (IS_ERR(bo)) 364 return PTR_ERR(bo); 365 366 ct->bo = bo; 367 368 return devm_add_action_or_reset(xe->drm.dev, guc_action_disable_ct, ct); 369 } 370 ALLOW_ERROR_INJECTION(xe_guc_ct_init, ERRNO); /* See xe_pci_probe() */ 371 372 /** 373 * xe_guc_ct_init_post_hwconfig - Reinitialize the GuC CTB in VRAM 374 * @ct: the &xe_guc_ct 375 * 376 * Allocate a new BO in VRAM and free the previous BO that was allocated 377 * in system memory (SMEM). Applicable only for DGFX products. 378 * 379 * Return: 0 on success, or a negative errno on failure. 380 */ 381 int xe_guc_ct_init_post_hwconfig(struct xe_guc_ct *ct) 382 { 383 struct xe_device *xe = ct_to_xe(ct); 384 struct xe_gt *gt = ct_to_gt(ct); 385 struct xe_tile *tile = gt_to_tile(gt); 386 int ret; 387 388 xe_assert(xe, !xe_guc_ct_enabled(ct)); 389 390 if (IS_DGFX(xe)) { 391 ret = xe_managed_bo_reinit_in_vram(xe, tile, &ct->bo); 392 if (ret) 393 return ret; 394 } 395 396 devm_remove_action(xe->drm.dev, guc_action_disable_ct, ct); 397 return devm_add_action_or_reset(xe->drm.dev, guc_action_disable_ct, ct); 398 } 399 400 #define desc_read(xe_, guc_ctb__, field_) \ 401 xe_map_rd_field(xe_, &guc_ctb__->desc, 0, \ 402 struct guc_ct_buffer_desc, field_) 403 404 #define desc_write(xe_, guc_ctb__, field_, val_) \ 405 xe_map_wr_field(xe_, &guc_ctb__->desc, 0, \ 406 struct guc_ct_buffer_desc, field_, val_) 407 408 static void guc_ct_ctb_h2g_init(struct xe_device *xe, struct guc_ctb *h2g, 409 struct iosys_map *map) 410 { 411 h2g->info.size = CTB_H2G_BUFFER_DWORDS; 412 h2g->info.resv_space = 0; 413 h2g->info.tail = 0; 414 h2g->info.head = 0; 415 h2g->info.space = CIRC_SPACE(h2g->info.tail, h2g->info.head, 416 h2g->info.size) - 417 h2g->info.resv_space; 418 h2g->info.broken = false; 419 420 h2g->desc = *map; 421 xe_map_memset(xe, &h2g->desc, 0, 0, sizeof(struct guc_ct_buffer_desc)); 422 423 h2g->cmds = IOSYS_MAP_INIT_OFFSET(map, CTB_H2G_BUFFER_OFFSET); 424 } 425 426 static void guc_ct_ctb_g2h_init(struct xe_device *xe, struct guc_ctb *g2h, 427 struct iosys_map *map) 428 { 429 g2h->info.size = CTB_G2H_BUFFER_DWORDS; 430 g2h->info.resv_space = G2H_ROOM_BUFFER_DWORDS; 431 g2h->info.head = 0; 432 g2h->info.tail = 0; 433 g2h->info.space = CIRC_SPACE(g2h->info.tail, g2h->info.head, 434 g2h->info.size) - 435 g2h->info.resv_space; 436 g2h->info.broken = false; 437 438 g2h->desc = IOSYS_MAP_INIT_OFFSET(map, CTB_DESC_SIZE); 439 xe_map_memset(xe, &g2h->desc, 0, 0, sizeof(struct guc_ct_buffer_desc)); 440 441 g2h->cmds = IOSYS_MAP_INIT_OFFSET(map, CTB_H2G_BUFFER_OFFSET + 442 CTB_H2G_BUFFER_SIZE); 443 } 444 445 static int guc_ct_ctb_h2g_register(struct xe_guc_ct *ct) 446 { 447 struct xe_guc *guc = ct_to_guc(ct); 448 u32 desc_addr, ctb_addr, size; 449 int err; 450 451 desc_addr = xe_bo_ggtt_addr(ct->bo); 452 ctb_addr = xe_bo_ggtt_addr(ct->bo) + CTB_H2G_BUFFER_OFFSET; 453 size = ct->ctbs.h2g.info.size * sizeof(u32); 454 455 err = xe_guc_self_cfg64(guc, 456 GUC_KLV_SELF_CFG_H2G_CTB_DESCRIPTOR_ADDR_KEY, 457 desc_addr); 458 if (err) 459 return err; 460 461 err = xe_guc_self_cfg64(guc, 462 GUC_KLV_SELF_CFG_H2G_CTB_ADDR_KEY, 463 ctb_addr); 464 if (err) 465 return err; 466 467 return xe_guc_self_cfg32(guc, 468 GUC_KLV_SELF_CFG_H2G_CTB_SIZE_KEY, 469 size); 470 } 471 472 static int guc_ct_ctb_g2h_register(struct xe_guc_ct *ct) 473 { 474 struct xe_guc *guc = ct_to_guc(ct); 475 u32 desc_addr, ctb_addr, size; 476 int err; 477 478 desc_addr = xe_bo_ggtt_addr(ct->bo) + CTB_DESC_SIZE; 479 ctb_addr = xe_bo_ggtt_addr(ct->bo) + CTB_H2G_BUFFER_OFFSET + 480 CTB_H2G_BUFFER_SIZE; 481 size = ct->ctbs.g2h.info.size * sizeof(u32); 482 483 err = xe_guc_self_cfg64(guc, 484 GUC_KLV_SELF_CFG_G2H_CTB_DESCRIPTOR_ADDR_KEY, 485 desc_addr); 486 if (err) 487 return err; 488 489 err = xe_guc_self_cfg64(guc, 490 GUC_KLV_SELF_CFG_G2H_CTB_ADDR_KEY, 491 ctb_addr); 492 if (err) 493 return err; 494 495 return xe_guc_self_cfg32(guc, 496 GUC_KLV_SELF_CFG_G2H_CTB_SIZE_KEY, 497 size); 498 } 499 500 static int guc_ct_control_toggle(struct xe_guc_ct *ct, bool enable) 501 { 502 u32 request[HOST2GUC_CONTROL_CTB_REQUEST_MSG_LEN] = { 503 FIELD_PREP(GUC_HXG_MSG_0_ORIGIN, GUC_HXG_ORIGIN_HOST) | 504 FIELD_PREP(GUC_HXG_MSG_0_TYPE, GUC_HXG_TYPE_REQUEST) | 505 FIELD_PREP(GUC_HXG_REQUEST_MSG_0_ACTION, 506 GUC_ACTION_HOST2GUC_CONTROL_CTB), 507 FIELD_PREP(HOST2GUC_CONTROL_CTB_REQUEST_MSG_1_CONTROL, 508 enable ? GUC_CTB_CONTROL_ENABLE : 509 GUC_CTB_CONTROL_DISABLE), 510 }; 511 int ret = xe_guc_mmio_send(ct_to_guc(ct), request, ARRAY_SIZE(request)); 512 513 return ret > 0 ? -EPROTO : ret; 514 } 515 516 static void guc_ct_change_state(struct xe_guc_ct *ct, 517 enum xe_guc_ct_state state) 518 { 519 struct xe_gt *gt = ct_to_gt(ct); 520 struct g2h_fence *g2h_fence; 521 unsigned long idx; 522 523 mutex_lock(&ct->lock); /* Serialise dequeue_one_g2h() */ 524 spin_lock_irq(&ct->fast_lock); /* Serialise CT fast-path */ 525 526 xe_gt_assert(ct_to_gt(ct), ct->g2h_outstanding == 0 || 527 state == XE_GUC_CT_STATE_STOPPED); 528 529 if (ct->g2h_outstanding) 530 xe_pm_runtime_put(ct_to_xe(ct)); 531 ct->g2h_outstanding = 0; 532 533 /* 534 * WRITE_ONCE pairs with READ_ONCEs in xe_guc_ct_initialized and 535 * xe_guc_ct_enabled. 536 */ 537 WRITE_ONCE(ct->state, state); 538 539 xe_gt_dbg(gt, "GuC CT communication channel %s\n", 540 state == XE_GUC_CT_STATE_STOPPED ? "stopped" : 541 str_enabled_disabled(state == XE_GUC_CT_STATE_ENABLED)); 542 543 spin_unlock_irq(&ct->fast_lock); 544 545 /* cancel all in-flight send-recv requests */ 546 xa_for_each(&ct->fence_lookup, idx, g2h_fence) 547 g2h_fence_cancel(g2h_fence); 548 549 /* make sure guc_ct_send_recv() will see g2h_fence changes */ 550 smp_mb(); 551 wake_up_all(&ct->g2h_fence_wq); 552 553 /* 554 * Lockdep doesn't like this under the fast lock and he destroy only 555 * needs to be serialized with the send path which ct lock provides. 556 */ 557 xa_destroy(&ct->fence_lookup); 558 559 mutex_unlock(&ct->lock); 560 } 561 562 static bool ct_needs_safe_mode(struct xe_guc_ct *ct) 563 { 564 return !pci_dev_msi_enabled(to_pci_dev(ct_to_xe(ct)->drm.dev)); 565 } 566 567 static bool ct_restart_safe_mode_worker(struct xe_guc_ct *ct) 568 { 569 if (!ct_needs_safe_mode(ct)) 570 return false; 571 572 queue_delayed_work(ct->g2h_wq, &ct->safe_mode_worker, HZ / 10); 573 return true; 574 } 575 576 static void safe_mode_worker_func(struct work_struct *w) 577 { 578 struct xe_guc_ct *ct = container_of(w, struct xe_guc_ct, safe_mode_worker.work); 579 580 receive_g2h(ct); 581 582 if (!ct_restart_safe_mode_worker(ct)) 583 xe_gt_dbg(ct_to_gt(ct), "GuC CT safe-mode canceled\n"); 584 } 585 586 static void ct_enter_safe_mode(struct xe_guc_ct *ct) 587 { 588 if (ct_restart_safe_mode_worker(ct)) 589 xe_gt_dbg(ct_to_gt(ct), "GuC CT safe-mode enabled\n"); 590 } 591 592 static void ct_exit_safe_mode(struct xe_guc_ct *ct) 593 { 594 if (cancel_delayed_work_sync(&ct->safe_mode_worker)) 595 xe_gt_dbg(ct_to_gt(ct), "GuC CT safe-mode disabled\n"); 596 } 597 598 static int __xe_guc_ct_start(struct xe_guc_ct *ct, bool needs_register) 599 { 600 struct xe_device *xe = ct_to_xe(ct); 601 struct xe_gt *gt = ct_to_gt(ct); 602 int err; 603 604 xe_gt_assert(gt, !xe_guc_ct_enabled(ct)); 605 606 if (needs_register) { 607 xe_map_memset(xe, &ct->bo->vmap, 0, 0, xe_bo_size(ct->bo)); 608 guc_ct_ctb_h2g_init(xe, &ct->ctbs.h2g, &ct->bo->vmap); 609 guc_ct_ctb_g2h_init(xe, &ct->ctbs.g2h, &ct->bo->vmap); 610 611 err = guc_ct_ctb_h2g_register(ct); 612 if (err) 613 goto err_out; 614 615 err = guc_ct_ctb_g2h_register(ct); 616 if (err) 617 goto err_out; 618 619 err = guc_ct_control_toggle(ct, true); 620 if (err) 621 goto err_out; 622 } else { 623 ct->ctbs.h2g.info.broken = false; 624 ct->ctbs.g2h.info.broken = false; 625 /* Skip everything in H2G buffer */ 626 xe_map_memset(xe, &ct->bo->vmap, CTB_H2G_BUFFER_OFFSET, 0, 627 CTB_H2G_BUFFER_SIZE); 628 } 629 630 guc_ct_change_state(ct, XE_GUC_CT_STATE_ENABLED); 631 632 smp_mb(); 633 wake_up_all(&ct->wq); 634 635 if (ct_needs_safe_mode(ct)) 636 ct_enter_safe_mode(ct); 637 638 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG) 639 /* 640 * The CT has now been reset so the dumper can be re-armed 641 * after any existing dead state has been dumped. 642 */ 643 spin_lock_irq(&ct->dead.lock); 644 if (ct->dead.reason) { 645 ct->dead.reason |= (1 << CT_DEAD_STATE_REARM); 646 queue_work(system_unbound_wq, &ct->dead.worker); 647 } 648 spin_unlock_irq(&ct->dead.lock); 649 #endif 650 651 return 0; 652 653 err_out: 654 xe_gt_err(gt, "Failed to enable GuC CT (%pe)\n", ERR_PTR(err)); 655 CT_DEAD(ct, NULL, SETUP); 656 657 return err; 658 } 659 660 /** 661 * xe_guc_ct_restart() - Restart GuC CT 662 * @ct: the &xe_guc_ct 663 * 664 * Restart GuC CT to an empty state without issuing a CT register MMIO command. 665 * 666 * Return: 0 on success, or a negative errno on failure. 667 */ 668 int xe_guc_ct_restart(struct xe_guc_ct *ct) 669 { 670 return __xe_guc_ct_start(ct, false); 671 } 672 673 /** 674 * xe_guc_ct_enable() - Enable GuC CT 675 * @ct: the &xe_guc_ct 676 * 677 * Enable GuC CT to an empty state and issue a CT register MMIO command. 678 * 679 * Return: 0 on success, or a negative errno on failure. 680 */ 681 int xe_guc_ct_enable(struct xe_guc_ct *ct) 682 { 683 return __xe_guc_ct_start(ct, true); 684 } 685 686 static void stop_g2h_handler(struct xe_guc_ct *ct) 687 { 688 cancel_work_sync(&ct->g2h_worker); 689 } 690 691 /** 692 * xe_guc_ct_disable - Set GuC to disabled state 693 * @ct: the &xe_guc_ct 694 * 695 * Set GuC CT to disabled state and stop g2h handler. No outstanding g2h expected 696 * in this transition. 697 */ 698 void xe_guc_ct_disable(struct xe_guc_ct *ct) 699 { 700 guc_ct_change_state(ct, XE_GUC_CT_STATE_DISABLED); 701 ct_exit_safe_mode(ct); 702 stop_g2h_handler(ct); 703 } 704 705 /** 706 * xe_guc_ct_flush_and_stop - Flush and stop all processing of G2H / H2G 707 * @ct: the &xe_guc_ct 708 */ 709 void xe_guc_ct_flush_and_stop(struct xe_guc_ct *ct) 710 { 711 receive_g2h(ct); 712 xe_guc_ct_stop(ct); 713 } 714 715 /** 716 * xe_guc_ct_stop - Set GuC to stopped state 717 * @ct: the &xe_guc_ct 718 * 719 * Set GuC CT to stopped state, stop g2h handler, and clear any outstanding g2h 720 */ 721 void xe_guc_ct_stop(struct xe_guc_ct *ct) 722 { 723 if (!xe_guc_ct_initialized(ct)) 724 return; 725 726 guc_ct_change_state(ct, XE_GUC_CT_STATE_STOPPED); 727 stop_g2h_handler(ct); 728 } 729 730 /** 731 * xe_guc_ct_runtime_suspend() - GuC CT runtime suspend 732 * @ct: the &xe_guc_ct 733 * 734 * Set GuC CT to disabled state. 735 */ 736 void xe_guc_ct_runtime_suspend(struct xe_guc_ct *ct) 737 { 738 struct guc_ctb *g2h = &ct->ctbs.g2h; 739 u32 credits = CIRC_SPACE(0, 0, CTB_G2H_BUFFER_DWORDS) - G2H_ROOM_BUFFER_DWORDS; 740 741 /* We should be back to guc_ct_ctb_g2h_init() values */ 742 xe_gt_assert(ct_to_gt(ct), g2h->info.space == credits); 743 744 /* 745 * Since we're already in runtime suspend path, we shouldn't have pending 746 * messages. But if there happen to be any, we'd probably want them to be 747 * thrown as errors for further investigation. 748 */ 749 xe_guc_ct_disable(ct); 750 } 751 752 /** 753 * xe_guc_ct_runtime_resume() - GuC CT runtime resume 754 * @ct: the &xe_guc_ct 755 * 756 * Restart GuC CT and set it to enabled state. 757 */ 758 void xe_guc_ct_runtime_resume(struct xe_guc_ct *ct) 759 { 760 xe_guc_ct_restart(ct); 761 } 762 763 static bool h2g_has_room(struct xe_guc_ct *ct, u32 cmd_len) 764 { 765 struct guc_ctb *h2g = &ct->ctbs.h2g; 766 767 lockdep_assert_held(&ct->lock); 768 769 if (cmd_len > h2g->info.space) { 770 h2g->info.head = desc_read(ct_to_xe(ct), h2g, head); 771 772 if (h2g->info.head > h2g->info.size) { 773 struct xe_device *xe = ct_to_xe(ct); 774 u32 desc_status = desc_read(xe, h2g, status); 775 776 desc_write(xe, h2g, status, desc_status | GUC_CTB_STATUS_OVERFLOW); 777 778 xe_gt_err(ct_to_gt(ct), "CT: invalid head offset %u >= %u)\n", 779 h2g->info.head, h2g->info.size); 780 CT_DEAD(ct, h2g, H2G_HAS_ROOM); 781 return false; 782 } 783 784 h2g->info.space = CIRC_SPACE(h2g->info.tail, h2g->info.head, 785 h2g->info.size) - 786 h2g->info.resv_space; 787 if (cmd_len > h2g->info.space) 788 return false; 789 } 790 791 return true; 792 } 793 794 static bool g2h_has_room(struct xe_guc_ct *ct, u32 g2h_len) 795 { 796 if (!g2h_len) 797 return true; 798 799 lockdep_assert_held(&ct->fast_lock); 800 801 return ct->ctbs.g2h.info.space > g2h_len; 802 } 803 804 static int has_room(struct xe_guc_ct *ct, u32 cmd_len, u32 g2h_len) 805 { 806 lockdep_assert_held(&ct->lock); 807 808 if (!g2h_has_room(ct, g2h_len) || !h2g_has_room(ct, cmd_len)) 809 return -EBUSY; 810 811 return 0; 812 } 813 814 static void h2g_reserve_space(struct xe_guc_ct *ct, u32 cmd_len) 815 { 816 lockdep_assert_held(&ct->lock); 817 ct->ctbs.h2g.info.space -= cmd_len; 818 } 819 820 static void __g2h_reserve_space(struct xe_guc_ct *ct, u32 g2h_len, u32 num_g2h) 821 { 822 xe_gt_assert(ct_to_gt(ct), g2h_len <= ct->ctbs.g2h.info.space); 823 xe_gt_assert(ct_to_gt(ct), (!g2h_len && !num_g2h) || 824 (g2h_len && num_g2h)); 825 826 if (g2h_len) { 827 lockdep_assert_held(&ct->fast_lock); 828 829 if (!ct->g2h_outstanding) 830 xe_pm_runtime_get_noresume(ct_to_xe(ct)); 831 832 ct->ctbs.g2h.info.space -= g2h_len; 833 ct->g2h_outstanding += num_g2h; 834 } 835 } 836 837 static void __g2h_release_space(struct xe_guc_ct *ct, u32 g2h_len) 838 { 839 bool bad = false; 840 841 lockdep_assert_held(&ct->fast_lock); 842 843 bad = ct->ctbs.g2h.info.space + g2h_len > 844 ct->ctbs.g2h.info.size - ct->ctbs.g2h.info.resv_space; 845 bad |= !ct->g2h_outstanding; 846 847 if (bad) { 848 xe_gt_err(ct_to_gt(ct), "Invalid G2H release: %d + %d vs %d - %d -> %d vs %d, outstanding = %d!\n", 849 ct->ctbs.g2h.info.space, g2h_len, 850 ct->ctbs.g2h.info.size, ct->ctbs.g2h.info.resv_space, 851 ct->ctbs.g2h.info.space + g2h_len, 852 ct->ctbs.g2h.info.size - ct->ctbs.g2h.info.resv_space, 853 ct->g2h_outstanding); 854 CT_DEAD(ct, &ct->ctbs.g2h, G2H_RELEASE); 855 return; 856 } 857 858 ct->ctbs.g2h.info.space += g2h_len; 859 if (!--ct->g2h_outstanding) 860 xe_pm_runtime_put(ct_to_xe(ct)); 861 } 862 863 static void g2h_release_space(struct xe_guc_ct *ct, u32 g2h_len) 864 { 865 spin_lock_irq(&ct->fast_lock); 866 __g2h_release_space(ct, g2h_len); 867 spin_unlock_irq(&ct->fast_lock); 868 } 869 870 /* 871 * The CT protocol accepts a 16 bits fence. This field is fully owned by the 872 * driver, the GuC will just copy it to the reply message. Since we need to 873 * be able to distinguish between replies to REQUEST and FAST_REQUEST messages, 874 * we use one bit of the seqno as an indicator for that and a rolling counter 875 * for the remaining 15 bits. 876 */ 877 #define CT_SEQNO_MASK GENMASK(14, 0) 878 #define CT_SEQNO_UNTRACKED BIT(15) 879 static u16 next_ct_seqno(struct xe_guc_ct *ct, bool is_g2h_fence) 880 { 881 u32 seqno = ct->fence_seqno++ & CT_SEQNO_MASK; 882 883 if (!is_g2h_fence) 884 seqno |= CT_SEQNO_UNTRACKED; 885 886 return seqno; 887 } 888 889 #define MAKE_ACTION(type, __action) \ 890 ({ \ 891 FIELD_PREP(GUC_HXG_MSG_0_TYPE, type) | \ 892 FIELD_PREP(GUC_HXG_EVENT_MSG_0_ACTION | \ 893 GUC_HXG_EVENT_MSG_0_DATA0, __action); \ 894 }) 895 896 static bool vf_action_can_safely_fail(struct xe_device *xe, u32 action) 897 { 898 /* 899 * When resuming a VF, we can't reliably track whether context 900 * registration has completed in the GuC state machine. It is harmless 901 * to resend the request, as it will fail silently if GUC_HXG_TYPE_EVENT 902 * is used. Additionally, if there is an H2G protocol issue on a VF, 903 * subsequent H2G messages sent as GUC_HXG_TYPE_FAST_REQUEST will likely 904 * fail. 905 */ 906 return IS_SRIOV_VF(xe) && xe_sriov_vf_migration_supported(xe) && 907 (action == XE_GUC_ACTION_REGISTER_CONTEXT_MULTI_LRC || 908 action == XE_GUC_ACTION_REGISTER_CONTEXT); 909 } 910 911 #define H2G_CT_HEADERS (GUC_CTB_HDR_LEN + 1) /* one DW CTB header and one DW HxG header */ 912 913 static int h2g_write(struct xe_guc_ct *ct, const u32 *action, u32 len, 914 u32 ct_fence_value, bool want_response) 915 { 916 struct xe_device *xe = ct_to_xe(ct); 917 struct xe_gt *gt = ct_to_gt(ct); 918 struct guc_ctb *h2g = &ct->ctbs.h2g; 919 u32 cmd[H2G_CT_HEADERS]; 920 u32 tail = h2g->info.tail; 921 u32 full_len; 922 struct iosys_map map = IOSYS_MAP_INIT_OFFSET(&h2g->cmds, 923 tail * sizeof(u32)); 924 u32 desc_status; 925 926 full_len = len + GUC_CTB_HDR_LEN; 927 928 lockdep_assert_held(&ct->lock); 929 xe_gt_assert(gt, full_len <= GUC_CTB_MSG_MAX_LEN); 930 931 desc_status = desc_read(xe, h2g, status); 932 if (desc_status) { 933 xe_gt_err(gt, "CT write: non-zero status: %u\n", desc_status); 934 goto corrupted; 935 } 936 937 if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) { 938 u32 desc_tail = desc_read(xe, h2g, tail); 939 u32 desc_head = desc_read(xe, h2g, head); 940 941 if (tail != desc_tail) { 942 desc_write(xe, h2g, status, desc_status | GUC_CTB_STATUS_MISMATCH); 943 xe_gt_err(gt, "CT write: tail was modified %u != %u\n", desc_tail, tail); 944 goto corrupted; 945 } 946 947 if (tail > h2g->info.size) { 948 desc_write(xe, h2g, status, desc_status | GUC_CTB_STATUS_OVERFLOW); 949 xe_gt_err(gt, "CT write: tail out of range: %u vs %u\n", 950 tail, h2g->info.size); 951 goto corrupted; 952 } 953 954 if (desc_head >= h2g->info.size) { 955 desc_write(xe, h2g, status, desc_status | GUC_CTB_STATUS_OVERFLOW); 956 xe_gt_err(gt, "CT write: invalid head offset %u >= %u)\n", 957 desc_head, h2g->info.size); 958 goto corrupted; 959 } 960 } 961 962 /* Command will wrap, zero fill (NOPs), return and check credits again */ 963 if (tail + full_len > h2g->info.size) { 964 xe_map_memset(xe, &map, 0, 0, 965 (h2g->info.size - tail) * sizeof(u32)); 966 h2g_reserve_space(ct, (h2g->info.size - tail)); 967 h2g->info.tail = 0; 968 desc_write(xe, h2g, tail, h2g->info.tail); 969 970 return -EAGAIN; 971 } 972 973 /* 974 * dw0: CT header (including fence) 975 * dw1: HXG header (including action code) 976 * dw2+: action data 977 */ 978 cmd[0] = FIELD_PREP(GUC_CTB_MSG_0_FORMAT, GUC_CTB_FORMAT_HXG) | 979 FIELD_PREP(GUC_CTB_MSG_0_NUM_DWORDS, len) | 980 FIELD_PREP(GUC_CTB_MSG_0_FENCE, ct_fence_value); 981 if (want_response) { 982 cmd[1] = MAKE_ACTION(GUC_HXG_TYPE_REQUEST, action[0]); 983 } else if (vf_action_can_safely_fail(xe, action[0])) { 984 cmd[1] = MAKE_ACTION(GUC_HXG_TYPE_EVENT, action[0]); 985 } else { 986 fast_req_track(ct, ct_fence_value, 987 FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, action[0])); 988 989 cmd[1] = MAKE_ACTION(GUC_HXG_TYPE_FAST_REQUEST, action[0]); 990 } 991 992 /* H2G header in cmd[1] replaces action[0] so: */ 993 --len; 994 ++action; 995 996 /* Write H2G ensuring visible before descriptor update */ 997 xe_map_memcpy_to(xe, &map, 0, cmd, H2G_CT_HEADERS * sizeof(u32)); 998 xe_map_memcpy_to(xe, &map, H2G_CT_HEADERS * sizeof(u32), action, len * sizeof(u32)); 999 xe_device_wmb(xe); 1000 1001 /* Update local copies */ 1002 h2g->info.tail = (tail + full_len) % h2g->info.size; 1003 h2g_reserve_space(ct, full_len); 1004 1005 /* Update descriptor */ 1006 desc_write(xe, h2g, tail, h2g->info.tail); 1007 1008 trace_xe_guc_ctb_h2g(xe, gt->info.id, *(action - 1), full_len, 1009 desc_read(xe, h2g, head), h2g->info.tail); 1010 1011 return 0; 1012 1013 corrupted: 1014 CT_DEAD(ct, &ct->ctbs.h2g, H2G_WRITE); 1015 return -EPIPE; 1016 } 1017 1018 static int __guc_ct_send_locked(struct xe_guc_ct *ct, const u32 *action, 1019 u32 len, u32 g2h_len, u32 num_g2h, 1020 struct g2h_fence *g2h_fence) 1021 { 1022 struct xe_gt *gt = ct_to_gt(ct); 1023 u16 seqno; 1024 int ret; 1025 1026 xe_gt_assert(gt, xe_guc_ct_initialized(ct)); 1027 xe_gt_assert(gt, !g2h_len || !g2h_fence); 1028 xe_gt_assert(gt, !num_g2h || !g2h_fence); 1029 xe_gt_assert(gt, !g2h_len || num_g2h); 1030 xe_gt_assert(gt, g2h_len || !num_g2h); 1031 lockdep_assert_held(&ct->lock); 1032 1033 if (unlikely(ct->ctbs.h2g.info.broken)) { 1034 ret = -EPIPE; 1035 goto out; 1036 } 1037 1038 if (ct->state == XE_GUC_CT_STATE_DISABLED) { 1039 ret = -ENODEV; 1040 goto out; 1041 } 1042 1043 if (ct->state == XE_GUC_CT_STATE_STOPPED || xe_gt_recovery_pending(gt)) { 1044 ret = -ECANCELED; 1045 goto out; 1046 } 1047 1048 xe_gt_assert(gt, xe_guc_ct_enabled(ct)); 1049 1050 if (g2h_fence) { 1051 g2h_len = GUC_CTB_HXG_MSG_MAX_LEN; 1052 num_g2h = 1; 1053 1054 if (g2h_fence_needs_alloc(g2h_fence)) { 1055 g2h_fence->seqno = next_ct_seqno(ct, true); 1056 ret = xa_err(xa_store(&ct->fence_lookup, 1057 g2h_fence->seqno, g2h_fence, 1058 GFP_ATOMIC)); 1059 if (ret) 1060 goto out; 1061 } 1062 1063 seqno = g2h_fence->seqno; 1064 } else { 1065 seqno = next_ct_seqno(ct, false); 1066 } 1067 1068 if (g2h_len) 1069 spin_lock_irq(&ct->fast_lock); 1070 retry: 1071 ret = has_room(ct, len + GUC_CTB_HDR_LEN, g2h_len); 1072 if (unlikely(ret)) 1073 goto out_unlock; 1074 1075 ret = h2g_write(ct, action, len, seqno, !!g2h_fence); 1076 if (unlikely(ret)) { 1077 if (ret == -EAGAIN) 1078 goto retry; 1079 goto out_unlock; 1080 } 1081 1082 __g2h_reserve_space(ct, g2h_len, num_g2h); 1083 xe_guc_notify(ct_to_guc(ct)); 1084 out_unlock: 1085 if (g2h_len) 1086 spin_unlock_irq(&ct->fast_lock); 1087 out: 1088 return ret; 1089 } 1090 1091 static void kick_reset(struct xe_guc_ct *ct) 1092 { 1093 xe_gt_reset_async(ct_to_gt(ct)); 1094 } 1095 1096 static int dequeue_one_g2h(struct xe_guc_ct *ct); 1097 1098 /* 1099 * wait before retry of sending h2g message 1100 * Return: true if ready for retry, false if the wait timeouted 1101 */ 1102 static bool guc_ct_send_wait_for_retry(struct xe_guc_ct *ct, u32 len, 1103 u32 g2h_len, struct g2h_fence *g2h_fence, 1104 unsigned int *sleep_period_ms) 1105 { 1106 struct xe_device *xe = ct_to_xe(ct); 1107 1108 /* 1109 * We wait to try to restore credits for about 1 second before bailing. 1110 * In the case of H2G credits we have no choice but just to wait for the 1111 * GuC to consume H2Gs in the channel so we use a wait / sleep loop. In 1112 * the case of G2H we process any G2H in the channel, hopefully freeing 1113 * credits as we consume the G2H messages. 1114 */ 1115 if (!h2g_has_room(ct, len + GUC_CTB_HDR_LEN)) { 1116 struct guc_ctb *h2g = &ct->ctbs.h2g; 1117 1118 if (*sleep_period_ms == 1024) 1119 return false; 1120 1121 trace_xe_guc_ct_h2g_flow_control(xe, h2g->info.head, h2g->info.tail, 1122 h2g->info.size, 1123 h2g->info.space, 1124 len + GUC_CTB_HDR_LEN); 1125 msleep(*sleep_period_ms); 1126 *sleep_period_ms <<= 1; 1127 } else { 1128 struct xe_device *xe = ct_to_xe(ct); 1129 struct guc_ctb *g2h = &ct->ctbs.g2h; 1130 int ret; 1131 1132 trace_xe_guc_ct_g2h_flow_control(xe, g2h->info.head, 1133 desc_read(xe, g2h, tail), 1134 g2h->info.size, 1135 g2h->info.space, 1136 g2h_fence ? 1137 GUC_CTB_HXG_MSG_MAX_LEN : 1138 g2h_len); 1139 1140 #define g2h_avail(ct) \ 1141 (desc_read(ct_to_xe(ct), (&ct->ctbs.g2h), tail) != ct->ctbs.g2h.info.head) 1142 if (!wait_event_timeout(ct->wq, !ct->g2h_outstanding || 1143 g2h_avail(ct), HZ)) 1144 return false; 1145 #undef g2h_avail 1146 1147 ret = dequeue_one_g2h(ct); 1148 if (ret < 0) { 1149 if (ret != -ECANCELED) 1150 xe_gt_err(ct_to_gt(ct), "CTB receive failed (%pe)", 1151 ERR_PTR(ret)); 1152 return false; 1153 } 1154 } 1155 return true; 1156 } 1157 1158 static int guc_ct_send_locked(struct xe_guc_ct *ct, const u32 *action, u32 len, 1159 u32 g2h_len, u32 num_g2h, 1160 struct g2h_fence *g2h_fence) 1161 { 1162 struct xe_gt *gt = ct_to_gt(ct); 1163 unsigned int sleep_period_ms = 1; 1164 int ret; 1165 1166 xe_gt_assert(gt, !g2h_len || !g2h_fence); 1167 lockdep_assert_held(&ct->lock); 1168 xe_device_assert_mem_access(ct_to_xe(ct)); 1169 1170 try_again: 1171 ret = __guc_ct_send_locked(ct, action, len, g2h_len, num_g2h, 1172 g2h_fence); 1173 1174 if (unlikely(ret == -EBUSY)) { 1175 if (!guc_ct_send_wait_for_retry(ct, len, g2h_len, g2h_fence, 1176 &sleep_period_ms)) 1177 goto broken; 1178 goto try_again; 1179 } 1180 1181 return ret; 1182 1183 broken: 1184 xe_gt_err(gt, "No forward process on H2G, reset required\n"); 1185 CT_DEAD(ct, &ct->ctbs.h2g, DEADLOCK); 1186 1187 return -EDEADLK; 1188 } 1189 1190 static int guc_ct_send(struct xe_guc_ct *ct, const u32 *action, u32 len, 1191 u32 g2h_len, u32 num_g2h, struct g2h_fence *g2h_fence) 1192 { 1193 int ret; 1194 1195 xe_gt_assert(ct_to_gt(ct), !g2h_len || !g2h_fence); 1196 1197 mutex_lock(&ct->lock); 1198 ret = guc_ct_send_locked(ct, action, len, g2h_len, num_g2h, g2h_fence); 1199 mutex_unlock(&ct->lock); 1200 1201 return ret; 1202 } 1203 1204 int xe_guc_ct_send(struct xe_guc_ct *ct, const u32 *action, u32 len, 1205 u32 g2h_len, u32 num_g2h) 1206 { 1207 int ret; 1208 1209 ret = guc_ct_send(ct, action, len, g2h_len, num_g2h, NULL); 1210 if (ret == -EDEADLK) 1211 kick_reset(ct); 1212 1213 return ret; 1214 } 1215 1216 int xe_guc_ct_send_locked(struct xe_guc_ct *ct, const u32 *action, u32 len, 1217 u32 g2h_len, u32 num_g2h) 1218 { 1219 int ret; 1220 1221 ret = guc_ct_send_locked(ct, action, len, g2h_len, num_g2h, NULL); 1222 if (ret == -EDEADLK) 1223 kick_reset(ct); 1224 1225 return ret; 1226 } 1227 1228 int xe_guc_ct_send_g2h_handler(struct xe_guc_ct *ct, const u32 *action, u32 len) 1229 { 1230 int ret; 1231 1232 lockdep_assert_held(&ct->lock); 1233 1234 ret = guc_ct_send_locked(ct, action, len, 0, 0, NULL); 1235 if (ret == -EDEADLK) 1236 kick_reset(ct); 1237 1238 return ret; 1239 } 1240 1241 /* 1242 * Check if a GT reset is in progress or will occur and if GT reset brought the 1243 * CT back up. Randomly picking 5 seconds for an upper limit to do a GT a reset. 1244 */ 1245 static bool retry_failure(struct xe_guc_ct *ct, int ret) 1246 { 1247 if (!(ret == -EDEADLK || ret == -EPIPE || ret == -ENODEV)) 1248 return false; 1249 1250 #define ct_alive(ct) \ 1251 (xe_guc_ct_enabled(ct) && !ct->ctbs.h2g.info.broken && \ 1252 !ct->ctbs.g2h.info.broken) 1253 if (!wait_event_interruptible_timeout(ct->wq, ct_alive(ct), HZ * 5)) 1254 return false; 1255 #undef ct_alive 1256 1257 return true; 1258 } 1259 1260 #define GUC_SEND_RETRY_LIMIT 50 1261 #define GUC_SEND_RETRY_MSLEEP 5 1262 1263 static int guc_ct_send_recv(struct xe_guc_ct *ct, const u32 *action, u32 len, 1264 u32 *response_buffer, bool no_fail) 1265 { 1266 struct xe_gt *gt = ct_to_gt(ct); 1267 struct g2h_fence g2h_fence; 1268 unsigned int retries = 0; 1269 int ret = 0; 1270 1271 /* 1272 * We use a fence to implement blocking sends / receiving response data. 1273 * The seqno of the fence is sent in the H2G, returned in the G2H, and 1274 * an xarray is used as storage media with the seqno being to key. 1275 * Fields in the fence hold success, failure, retry status and the 1276 * response data. Safe to allocate on the stack as the xarray is the 1277 * only reference and it cannot be present after this function exits. 1278 */ 1279 retry: 1280 g2h_fence_init(&g2h_fence, response_buffer); 1281 retry_same_fence: 1282 ret = guc_ct_send(ct, action, len, 0, 0, &g2h_fence); 1283 if (unlikely(ret == -ENOMEM)) { 1284 /* Retry allocation /w GFP_KERNEL */ 1285 ret = xa_err(xa_store(&ct->fence_lookup, g2h_fence.seqno, 1286 &g2h_fence, GFP_KERNEL)); 1287 if (ret) 1288 return ret; 1289 1290 goto retry_same_fence; 1291 } else if (unlikely(ret)) { 1292 if (ret == -EDEADLK) 1293 kick_reset(ct); 1294 1295 if (no_fail && retry_failure(ct, ret)) 1296 goto retry_same_fence; 1297 1298 if (!g2h_fence_needs_alloc(&g2h_fence)) 1299 xa_erase(&ct->fence_lookup, g2h_fence.seqno); 1300 1301 return ret; 1302 } 1303 1304 /* READ_ONCEs pairs with WRITE_ONCEs in parse_g2h_response 1305 * and g2h_fence_cancel. 1306 */ 1307 ret = wait_event_timeout(ct->g2h_fence_wq, READ_ONCE(g2h_fence.done), HZ); 1308 if (!ret) { 1309 LNL_FLUSH_WORK(&ct->g2h_worker); 1310 if (READ_ONCE(g2h_fence.done)) { 1311 xe_gt_warn(gt, "G2H fence %u, action %04x, done\n", 1312 g2h_fence.seqno, action[0]); 1313 ret = 1; 1314 } 1315 } 1316 1317 /* 1318 * Ensure we serialize with completion side to prevent UAF with fence going out of scope on 1319 * the stack, since we have no clue if it will fire after the timeout before we can erase 1320 * from the xa. Also we have some dependent loads and stores below for which we need the 1321 * correct ordering, and we lack the needed barriers. 1322 */ 1323 mutex_lock(&ct->lock); 1324 if (!ret) { 1325 xe_gt_err(gt, "Timed out wait for G2H, fence %u, action %04x, done %s", 1326 g2h_fence.seqno, action[0], str_yes_no(g2h_fence.done)); 1327 xa_erase(&ct->fence_lookup, g2h_fence.seqno); 1328 mutex_unlock(&ct->lock); 1329 return -ETIME; 1330 } 1331 1332 if (g2h_fence.retry) { 1333 xe_gt_dbg(gt, "H2G action %#x retrying: reason %#x\n", 1334 action[0], g2h_fence.reason); 1335 mutex_unlock(&ct->lock); 1336 if (++retries > GUC_SEND_RETRY_LIMIT) { 1337 xe_gt_err(gt, "H2G action %#x reached retry limit=%u, aborting\n", 1338 action[0], GUC_SEND_RETRY_LIMIT); 1339 return -ELOOP; 1340 } 1341 msleep(GUC_SEND_RETRY_MSLEEP * retries); 1342 goto retry; 1343 } 1344 if (g2h_fence.fail) { 1345 if (g2h_fence.cancel) { 1346 xe_gt_dbg(gt, "H2G request %#x canceled!\n", action[0]); 1347 ret = -ECANCELED; 1348 goto unlock; 1349 } 1350 xe_gt_err(gt, "H2G request %#x failed: error %#x hint %#x\n", 1351 action[0], g2h_fence.error, g2h_fence.hint); 1352 ret = -EIO; 1353 } 1354 1355 if (ret > 0) 1356 ret = response_buffer ? g2h_fence.response_len : g2h_fence.response_data; 1357 1358 unlock: 1359 mutex_unlock(&ct->lock); 1360 1361 return ret; 1362 } 1363 1364 /** 1365 * xe_guc_ct_send_recv - Send and receive HXG to the GuC 1366 * @ct: the &xe_guc_ct 1367 * @action: the dword array with `HXG Request`_ message (can't be NULL) 1368 * @len: length of the `HXG Request`_ message (in dwords, can't be 0) 1369 * @response_buffer: placeholder for the `HXG Response`_ message (can be NULL) 1370 * 1371 * Send a `HXG Request`_ message to the GuC over CT communication channel and 1372 * blocks until GuC replies with a `HXG Response`_ message. 1373 * 1374 * For non-blocking communication with GuC use xe_guc_ct_send(). 1375 * 1376 * Note: The size of &response_buffer must be at least GUC_CTB_MAX_DWORDS_. 1377 * 1378 * Return: response length (in dwords) if &response_buffer was not NULL, or 1379 * DATA0 from `HXG Response`_ if &response_buffer was NULL, or 1380 * a negative error code on failure. 1381 */ 1382 int xe_guc_ct_send_recv(struct xe_guc_ct *ct, const u32 *action, u32 len, 1383 u32 *response_buffer) 1384 { 1385 KUNIT_STATIC_STUB_REDIRECT(xe_guc_ct_send_recv, ct, action, len, response_buffer); 1386 return guc_ct_send_recv(ct, action, len, response_buffer, false); 1387 } 1388 ALLOW_ERROR_INJECTION(xe_guc_ct_send_recv, ERRNO); 1389 1390 int xe_guc_ct_send_recv_no_fail(struct xe_guc_ct *ct, const u32 *action, 1391 u32 len, u32 *response_buffer) 1392 { 1393 return guc_ct_send_recv(ct, action, len, response_buffer, true); 1394 } 1395 1396 static u32 *msg_to_hxg(u32 *msg) 1397 { 1398 return msg + GUC_CTB_MSG_MIN_LEN; 1399 } 1400 1401 static u32 msg_len_to_hxg_len(u32 len) 1402 { 1403 return len - GUC_CTB_MSG_MIN_LEN; 1404 } 1405 1406 static int parse_g2h_event(struct xe_guc_ct *ct, u32 *msg, u32 len) 1407 { 1408 u32 *hxg = msg_to_hxg(msg); 1409 u32 action = FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, hxg[0]); 1410 1411 lockdep_assert_held(&ct->lock); 1412 1413 switch (action) { 1414 case XE_GUC_ACTION_NOTIFY_MULTI_QUEUE_CONTEXT_CGP_SYNC_DONE: 1415 case XE_GUC_ACTION_SCHED_CONTEXT_MODE_DONE: 1416 case XE_GUC_ACTION_DEREGISTER_CONTEXT_DONE: 1417 case XE_GUC_ACTION_SCHED_ENGINE_MODE_DONE: 1418 case XE_GUC_ACTION_TLB_INVALIDATION_DONE: 1419 case XE_GUC_ACTION_PAGE_RECLAMATION_DONE: 1420 g2h_release_space(ct, len); 1421 } 1422 1423 return 0; 1424 } 1425 1426 static int guc_crash_process_msg(struct xe_guc_ct *ct, u32 action) 1427 { 1428 struct xe_gt *gt = ct_to_gt(ct); 1429 1430 if (action == XE_GUC_ACTION_NOTIFY_CRASH_DUMP_POSTED) 1431 xe_gt_err(gt, "GuC Crash dump notification\n"); 1432 else if (action == XE_GUC_ACTION_NOTIFY_EXCEPTION) 1433 xe_gt_err(gt, "GuC Exception notification\n"); 1434 else 1435 xe_gt_err(gt, "Unknown GuC crash notification: 0x%04X\n", action); 1436 1437 CT_DEAD(ct, NULL, CRASH); 1438 1439 kick_reset(ct); 1440 1441 return 0; 1442 } 1443 1444 static int parse_g2h_response(struct xe_guc_ct *ct, u32 *msg, u32 len) 1445 { 1446 struct xe_gt *gt = ct_to_gt(ct); 1447 u32 *hxg = msg_to_hxg(msg); 1448 u32 hxg_len = msg_len_to_hxg_len(len); 1449 u32 fence = FIELD_GET(GUC_CTB_MSG_0_FENCE, msg[0]); 1450 u32 type = FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]); 1451 struct g2h_fence *g2h_fence; 1452 1453 lockdep_assert_held(&ct->lock); 1454 1455 /* 1456 * Fences for FAST_REQUEST messages are not tracked in ct->fence_lookup. 1457 * Those messages should never fail, so if we do get an error back it 1458 * means we're likely doing an illegal operation and the GuC is 1459 * rejecting it. We have no way to inform the code that submitted the 1460 * H2G that the message was rejected, so we need to escalate the 1461 * failure to trigger a reset. 1462 */ 1463 if (fence & CT_SEQNO_UNTRACKED) { 1464 if (type == GUC_HXG_TYPE_RESPONSE_FAILURE) 1465 xe_gt_err(gt, "FAST_REQ H2G fence 0x%x failed! e=0x%x, h=%u\n", 1466 fence, 1467 FIELD_GET(GUC_HXG_FAILURE_MSG_0_ERROR, hxg[0]), 1468 FIELD_GET(GUC_HXG_FAILURE_MSG_0_HINT, hxg[0])); 1469 else 1470 xe_gt_err(gt, "unexpected response %u for FAST_REQ H2G fence 0x%x!\n", 1471 type, fence); 1472 1473 fast_req_report(ct, fence); 1474 1475 /* FIXME: W/A race in the GuC, will get in firmware soon */ 1476 if (xe_gt_recovery_pending(gt)) 1477 return 0; 1478 1479 CT_DEAD(ct, NULL, PARSE_G2H_RESPONSE); 1480 1481 return -EPROTO; 1482 } 1483 1484 g2h_fence = xa_erase(&ct->fence_lookup, fence); 1485 if (unlikely(!g2h_fence)) { 1486 /* Don't tear down channel, as send could've timed out */ 1487 /* CT_DEAD(ct, NULL, PARSE_G2H_UNKNOWN); */ 1488 xe_gt_warn(gt, "G2H fence (%u) not found!\n", fence); 1489 g2h_release_space(ct, GUC_CTB_HXG_MSG_MAX_LEN); 1490 return 0; 1491 } 1492 1493 xe_gt_assert(gt, fence == g2h_fence->seqno); 1494 1495 if (type == GUC_HXG_TYPE_RESPONSE_FAILURE) { 1496 g2h_fence->fail = true; 1497 g2h_fence->error = FIELD_GET(GUC_HXG_FAILURE_MSG_0_ERROR, hxg[0]); 1498 g2h_fence->hint = FIELD_GET(GUC_HXG_FAILURE_MSG_0_HINT, hxg[0]); 1499 } else if (type == GUC_HXG_TYPE_NO_RESPONSE_RETRY) { 1500 g2h_fence->retry = true; 1501 g2h_fence->reason = FIELD_GET(GUC_HXG_RETRY_MSG_0_REASON, hxg[0]); 1502 } else if (g2h_fence->response_buffer) { 1503 g2h_fence->response_len = hxg_len; 1504 memcpy(g2h_fence->response_buffer, hxg, hxg_len * sizeof(u32)); 1505 } else { 1506 g2h_fence->response_data = FIELD_GET(GUC_HXG_RESPONSE_MSG_0_DATA0, hxg[0]); 1507 } 1508 1509 g2h_release_space(ct, GUC_CTB_HXG_MSG_MAX_LEN); 1510 1511 /* WRITE_ONCE pairs with READ_ONCEs in guc_ct_send_recv. */ 1512 WRITE_ONCE(g2h_fence->done, true); 1513 smp_mb(); 1514 1515 wake_up_all(&ct->g2h_fence_wq); 1516 1517 return 0; 1518 } 1519 1520 static int parse_g2h_msg(struct xe_guc_ct *ct, u32 *msg, u32 len) 1521 { 1522 struct xe_gt *gt = ct_to_gt(ct); 1523 u32 *hxg = msg_to_hxg(msg); 1524 u32 origin, type; 1525 int ret; 1526 1527 lockdep_assert_held(&ct->lock); 1528 1529 origin = FIELD_GET(GUC_HXG_MSG_0_ORIGIN, hxg[0]); 1530 if (unlikely(origin != GUC_HXG_ORIGIN_GUC)) { 1531 xe_gt_err(gt, "G2H channel broken on read, origin=%u, reset required\n", 1532 origin); 1533 CT_DEAD(ct, &ct->ctbs.g2h, PARSE_G2H_ORIGIN); 1534 1535 return -EPROTO; 1536 } 1537 1538 type = FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]); 1539 switch (type) { 1540 case GUC_HXG_TYPE_EVENT: 1541 ret = parse_g2h_event(ct, msg, len); 1542 break; 1543 case GUC_HXG_TYPE_RESPONSE_SUCCESS: 1544 case GUC_HXG_TYPE_RESPONSE_FAILURE: 1545 case GUC_HXG_TYPE_NO_RESPONSE_RETRY: 1546 ret = parse_g2h_response(ct, msg, len); 1547 break; 1548 default: 1549 xe_gt_err(gt, "G2H channel broken on read, type=%u, reset required\n", 1550 type); 1551 CT_DEAD(ct, &ct->ctbs.g2h, PARSE_G2H_TYPE); 1552 1553 ret = -EOPNOTSUPP; 1554 } 1555 1556 return ret; 1557 } 1558 1559 static int process_g2h_msg(struct xe_guc_ct *ct, u32 *msg, u32 len) 1560 { 1561 struct xe_guc *guc = ct_to_guc(ct); 1562 struct xe_gt *gt = ct_to_gt(ct); 1563 u32 hxg_len = msg_len_to_hxg_len(len); 1564 u32 *hxg = msg_to_hxg(msg); 1565 u32 action, adj_len; 1566 u32 *payload; 1567 int ret = 0; 1568 1569 if (FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]) != GUC_HXG_TYPE_EVENT) 1570 return 0; 1571 1572 action = FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, hxg[0]); 1573 payload = hxg + GUC_HXG_EVENT_MSG_MIN_LEN; 1574 adj_len = hxg_len - GUC_HXG_EVENT_MSG_MIN_LEN; 1575 1576 switch (action) { 1577 case XE_GUC_ACTION_SCHED_CONTEXT_MODE_DONE: 1578 ret = xe_guc_sched_done_handler(guc, payload, adj_len); 1579 break; 1580 case XE_GUC_ACTION_DEREGISTER_CONTEXT_DONE: 1581 ret = xe_guc_deregister_done_handler(guc, payload, adj_len); 1582 break; 1583 case XE_GUC_ACTION_CONTEXT_RESET_NOTIFICATION: 1584 ret = xe_guc_exec_queue_reset_handler(guc, payload, adj_len); 1585 break; 1586 case XE_GUC_ACTION_ENGINE_FAILURE_NOTIFICATION: 1587 ret = xe_guc_exec_queue_reset_failure_handler(guc, payload, 1588 adj_len); 1589 break; 1590 case XE_GUC_ACTION_SCHED_ENGINE_MODE_DONE: 1591 /* Selftest only at the moment */ 1592 break; 1593 case XE_GUC_ACTION_STATE_CAPTURE_NOTIFICATION: 1594 ret = xe_guc_error_capture_handler(guc, payload, adj_len); 1595 break; 1596 case XE_GUC_ACTION_NOTIFY_FLUSH_LOG_BUFFER_TO_FILE: 1597 /* FIXME: Handle this */ 1598 break; 1599 case XE_GUC_ACTION_NOTIFY_MEMORY_CAT_ERROR: 1600 ret = xe_guc_exec_queue_memory_cat_error_handler(guc, payload, 1601 adj_len); 1602 break; 1603 case XE_GUC_ACTION_REPORT_PAGE_FAULT_REQ_DESC: 1604 ret = xe_guc_pagefault_handler(guc, payload, adj_len); 1605 break; 1606 case XE_GUC_ACTION_TLB_INVALIDATION_DONE: 1607 case XE_GUC_ACTION_PAGE_RECLAMATION_DONE: 1608 /* 1609 * Page reclamation is an extension of TLB invalidation. Both 1610 * operations share the same seqno and fence. When either 1611 * action completes, we need to signal the corresponding 1612 * fence. Since the handling logic (lookup fence by seqno, 1613 * fence signalling) is identical, we use the same handler 1614 * for both G2H events. 1615 */ 1616 ret = xe_guc_tlb_inval_done_handler(guc, payload, adj_len); 1617 break; 1618 case XE_GUC_ACTION_GUC2PF_RELAY_FROM_VF: 1619 ret = xe_guc_relay_process_guc2pf(&guc->relay, hxg, hxg_len); 1620 break; 1621 case XE_GUC_ACTION_GUC2VF_RELAY_FROM_PF: 1622 ret = xe_guc_relay_process_guc2vf(&guc->relay, hxg, hxg_len); 1623 break; 1624 case GUC_ACTION_GUC2PF_VF_STATE_NOTIFY: 1625 ret = xe_gt_sriov_pf_control_process_guc2pf(gt, hxg, hxg_len); 1626 break; 1627 case GUC_ACTION_GUC2PF_ADVERSE_EVENT: 1628 ret = xe_gt_sriov_pf_monitor_process_guc2pf(gt, hxg, hxg_len); 1629 break; 1630 case XE_GUC_ACTION_NOTIFY_CRASH_DUMP_POSTED: 1631 case XE_GUC_ACTION_NOTIFY_EXCEPTION: 1632 ret = guc_crash_process_msg(ct, action); 1633 break; 1634 #if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST) 1635 case XE_GUC_ACTION_TEST_G2G_RECV: 1636 ret = xe_guc_g2g_test_notification(guc, payload, adj_len); 1637 break; 1638 #endif 1639 case XE_GUC_ACTION_NOTIFY_MULTI_QUEUE_CONTEXT_CGP_SYNC_DONE: 1640 ret = xe_guc_exec_queue_cgp_sync_done_handler(guc, payload, adj_len); 1641 break; 1642 case XE_GUC_ACTION_NOTIFY_MULTI_QUEUE_CGP_CONTEXT_ERROR: 1643 ret = xe_guc_exec_queue_cgp_context_error_handler(guc, payload, 1644 adj_len); 1645 break; 1646 default: 1647 xe_gt_err(gt, "unexpected G2H action 0x%04x\n", action); 1648 } 1649 1650 if (ret) { 1651 xe_gt_err(gt, "G2H action %#04x failed (%pe) len %u msg %*ph\n", 1652 action, ERR_PTR(ret), hxg_len, (int)sizeof(u32) * hxg_len, hxg); 1653 CT_DEAD(ct, NULL, PROCESS_FAILED); 1654 } 1655 1656 return 0; 1657 } 1658 1659 static int g2h_read(struct xe_guc_ct *ct, u32 *msg, bool fast_path) 1660 { 1661 struct xe_device *xe = ct_to_xe(ct); 1662 struct xe_gt *gt = ct_to_gt(ct); 1663 struct guc_ctb *g2h = &ct->ctbs.g2h; 1664 u32 tail, head, len, desc_status; 1665 s32 avail; 1666 u32 action; 1667 u32 *hxg; 1668 1669 xe_gt_assert(gt, xe_guc_ct_initialized(ct)); 1670 lockdep_assert_held(&ct->fast_lock); 1671 1672 if (ct->state == XE_GUC_CT_STATE_DISABLED) 1673 return -ENODEV; 1674 1675 if (ct->state == XE_GUC_CT_STATE_STOPPED) 1676 return -ECANCELED; 1677 1678 if (g2h->info.broken) 1679 return -EPIPE; 1680 1681 xe_gt_assert(gt, xe_guc_ct_enabled(ct)); 1682 1683 desc_status = desc_read(xe, g2h, status); 1684 if (desc_status) { 1685 if (desc_status & GUC_CTB_STATUS_DISABLED) { 1686 /* 1687 * Potentially valid if a CLIENT_RESET request resulted in 1688 * contexts/engines being reset. But should never happen as 1689 * no contexts should be active when CLIENT_RESET is sent. 1690 */ 1691 xe_gt_err(gt, "CT read: unexpected G2H after GuC has stopped!\n"); 1692 desc_status &= ~GUC_CTB_STATUS_DISABLED; 1693 } 1694 1695 if (desc_status) { 1696 xe_gt_err(gt, "CT read: non-zero status: %u\n", desc_status); 1697 goto corrupted; 1698 } 1699 } 1700 1701 if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) { 1702 u32 desc_tail = desc_read(xe, g2h, tail); 1703 /* 1704 u32 desc_head = desc_read(xe, g2h, head); 1705 1706 * info.head and desc_head are updated back-to-back at the end of 1707 * this function and nowhere else. Hence, they cannot be different 1708 * unless two g2h_read calls are running concurrently. Which is not 1709 * possible because it is guarded by ct->fast_lock. And yet, some 1710 * discrete platforms are regularly hitting this error :(. 1711 * 1712 * desc_head rolling backwards shouldn't cause any noticeable 1713 * problems - just a delay in GuC being allowed to proceed past that 1714 * point in the queue. So for now, just disable the error until it 1715 * can be root caused. 1716 * 1717 if (g2h->info.head != desc_head) { 1718 desc_write(xe, g2h, status, desc_status | GUC_CTB_STATUS_MISMATCH); 1719 xe_gt_err(gt, "CT read: head was modified %u != %u\n", 1720 desc_head, g2h->info.head); 1721 goto corrupted; 1722 } 1723 */ 1724 1725 if (g2h->info.head > g2h->info.size) { 1726 desc_write(xe, g2h, status, desc_status | GUC_CTB_STATUS_OVERFLOW); 1727 xe_gt_err(gt, "CT read: head out of range: %u vs %u\n", 1728 g2h->info.head, g2h->info.size); 1729 goto corrupted; 1730 } 1731 1732 if (desc_tail >= g2h->info.size) { 1733 desc_write(xe, g2h, status, desc_status | GUC_CTB_STATUS_OVERFLOW); 1734 xe_gt_err(gt, "CT read: invalid tail offset %u >= %u)\n", 1735 desc_tail, g2h->info.size); 1736 goto corrupted; 1737 } 1738 } 1739 1740 /* Calculate DW available to read */ 1741 tail = desc_read(xe, g2h, tail); 1742 avail = tail - g2h->info.head; 1743 if (unlikely(avail == 0)) 1744 return 0; 1745 1746 if (avail < 0) 1747 avail += g2h->info.size; 1748 1749 /* Read header */ 1750 xe_map_memcpy_from(xe, msg, &g2h->cmds, sizeof(u32) * g2h->info.head, 1751 sizeof(u32)); 1752 len = FIELD_GET(GUC_CTB_MSG_0_NUM_DWORDS, msg[0]) + GUC_CTB_MSG_MIN_LEN; 1753 if (len > avail) { 1754 xe_gt_err(gt, "G2H channel broken on read, avail=%d, len=%d, reset required\n", 1755 avail, len); 1756 goto corrupted; 1757 } 1758 1759 head = (g2h->info.head + 1) % g2h->info.size; 1760 avail = len - 1; 1761 1762 /* Read G2H message */ 1763 if (avail + head > g2h->info.size) { 1764 u32 avail_til_wrap = g2h->info.size - head; 1765 1766 xe_map_memcpy_from(xe, msg + 1, 1767 &g2h->cmds, sizeof(u32) * head, 1768 avail_til_wrap * sizeof(u32)); 1769 xe_map_memcpy_from(xe, msg + 1 + avail_til_wrap, 1770 &g2h->cmds, 0, 1771 (avail - avail_til_wrap) * sizeof(u32)); 1772 } else { 1773 xe_map_memcpy_from(xe, msg + 1, 1774 &g2h->cmds, sizeof(u32) * head, 1775 avail * sizeof(u32)); 1776 } 1777 1778 hxg = msg_to_hxg(msg); 1779 action = FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, hxg[0]); 1780 1781 if (fast_path) { 1782 if (FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]) != GUC_HXG_TYPE_EVENT) 1783 return 0; 1784 1785 switch (action) { 1786 case XE_GUC_ACTION_REPORT_PAGE_FAULT_REQ_DESC: 1787 case XE_GUC_ACTION_TLB_INVALIDATION_DONE: 1788 case XE_GUC_ACTION_PAGE_RECLAMATION_DONE: 1789 break; /* Process these in fast-path */ 1790 default: 1791 return 0; 1792 } 1793 } 1794 1795 /* Update local / descriptor header */ 1796 g2h->info.head = (head + avail) % g2h->info.size; 1797 desc_write(xe, g2h, head, g2h->info.head); 1798 1799 trace_xe_guc_ctb_g2h(xe, ct_to_gt(ct)->info.id, 1800 action, len, g2h->info.head, tail); 1801 1802 return len; 1803 1804 corrupted: 1805 CT_DEAD(ct, &ct->ctbs.g2h, G2H_READ); 1806 return -EPROTO; 1807 } 1808 1809 static void g2h_fast_path(struct xe_guc_ct *ct, u32 *msg, u32 len) 1810 { 1811 struct xe_gt *gt = ct_to_gt(ct); 1812 struct xe_guc *guc = ct_to_guc(ct); 1813 u32 hxg_len = msg_len_to_hxg_len(len); 1814 u32 *hxg = msg_to_hxg(msg); 1815 u32 action = FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, hxg[0]); 1816 u32 *payload = hxg + GUC_HXG_MSG_MIN_LEN; 1817 u32 adj_len = hxg_len - GUC_HXG_MSG_MIN_LEN; 1818 int ret = 0; 1819 1820 switch (action) { 1821 case XE_GUC_ACTION_REPORT_PAGE_FAULT_REQ_DESC: 1822 ret = xe_guc_pagefault_handler(guc, payload, adj_len); 1823 break; 1824 case XE_GUC_ACTION_TLB_INVALIDATION_DONE: 1825 case XE_GUC_ACTION_PAGE_RECLAMATION_DONE: 1826 /* 1827 * Seqno and fence handling of page reclamation and TLB 1828 * invalidation is identical, so we can use the same handler 1829 * for both actions. 1830 */ 1831 __g2h_release_space(ct, len); 1832 ret = xe_guc_tlb_inval_done_handler(guc, payload, adj_len); 1833 break; 1834 default: 1835 xe_gt_warn(gt, "NOT_POSSIBLE"); 1836 } 1837 1838 if (ret) { 1839 xe_gt_err(gt, "G2H action 0x%04x failed (%pe)\n", 1840 action, ERR_PTR(ret)); 1841 CT_DEAD(ct, NULL, FAST_G2H); 1842 } 1843 } 1844 1845 /** 1846 * xe_guc_ct_fast_path - process critical G2H in the IRQ handler 1847 * @ct: GuC CT object 1848 * 1849 * Anything related to page faults is critical for performance, process these 1850 * critical G2H in the IRQ. This is safe as these handlers either just wake up 1851 * waiters or queue another worker. 1852 */ 1853 void xe_guc_ct_fast_path(struct xe_guc_ct *ct) 1854 { 1855 struct xe_device *xe = ct_to_xe(ct); 1856 bool ongoing; 1857 int len; 1858 1859 ongoing = xe_pm_runtime_get_if_active(ct_to_xe(ct)); 1860 if (!ongoing && xe_pm_read_callback_task(ct_to_xe(ct)) == NULL) 1861 return; 1862 1863 spin_lock(&ct->fast_lock); 1864 do { 1865 len = g2h_read(ct, ct->fast_msg, true); 1866 if (len > 0) 1867 g2h_fast_path(ct, ct->fast_msg, len); 1868 } while (len > 0); 1869 spin_unlock(&ct->fast_lock); 1870 1871 if (ongoing) 1872 xe_pm_runtime_put(xe); 1873 } 1874 1875 /* Returns less than zero on error, 0 on done, 1 on more available */ 1876 static int dequeue_one_g2h(struct xe_guc_ct *ct) 1877 { 1878 int len; 1879 int ret; 1880 1881 lockdep_assert_held(&ct->lock); 1882 1883 spin_lock_irq(&ct->fast_lock); 1884 len = g2h_read(ct, ct->msg, false); 1885 spin_unlock_irq(&ct->fast_lock); 1886 if (len <= 0) 1887 return len; 1888 1889 ret = parse_g2h_msg(ct, ct->msg, len); 1890 if (unlikely(ret < 0)) 1891 return ret; 1892 1893 ret = process_g2h_msg(ct, ct->msg, len); 1894 if (unlikely(ret < 0)) 1895 return ret; 1896 1897 return 1; 1898 } 1899 1900 static void receive_g2h(struct xe_guc_ct *ct) 1901 { 1902 bool ongoing; 1903 int ret; 1904 1905 /* 1906 * Normal users must always hold mem_access.ref around CT calls. However 1907 * during the runtime pm callbacks we rely on CT to talk to the GuC, but 1908 * at this stage we can't rely on mem_access.ref and even the 1909 * callback_task will be different than current. For such cases we just 1910 * need to ensure we always process the responses from any blocking 1911 * ct_send requests or where we otherwise expect some response when 1912 * initiated from those callbacks (which will need to wait for the below 1913 * dequeue_one_g2h()). The dequeue_one_g2h() will gracefully fail if 1914 * the device has suspended to the point that the CT communication has 1915 * been disabled. 1916 * 1917 * If we are inside the runtime pm callback, we can be the only task 1918 * still issuing CT requests (since that requires having the 1919 * mem_access.ref). It seems like it might in theory be possible to 1920 * receive unsolicited events from the GuC just as we are 1921 * suspending-resuming, but those will currently anyway be lost when 1922 * eventually exiting from suspend, hence no need to wake up the device 1923 * here. If we ever need something stronger than get_if_ongoing() then 1924 * we need to be careful with blocking the pm callbacks from getting CT 1925 * responses, if the worker here is blocked on those callbacks 1926 * completing, creating a deadlock. 1927 */ 1928 ongoing = xe_pm_runtime_get_if_active(ct_to_xe(ct)); 1929 if (!ongoing && xe_pm_read_callback_task(ct_to_xe(ct)) == NULL) 1930 return; 1931 1932 do { 1933 mutex_lock(&ct->lock); 1934 ret = dequeue_one_g2h(ct); 1935 mutex_unlock(&ct->lock); 1936 1937 if (unlikely(ret == -EPROTO || ret == -EOPNOTSUPP)) { 1938 xe_gt_err(ct_to_gt(ct), "CT dequeue failed: %d", ret); 1939 CT_DEAD(ct, NULL, G2H_RECV); 1940 kick_reset(ct); 1941 } 1942 } while (ret == 1); 1943 1944 if (ongoing) 1945 xe_pm_runtime_put(ct_to_xe(ct)); 1946 } 1947 1948 static void g2h_worker_func(struct work_struct *w) 1949 { 1950 struct xe_guc_ct *ct = container_of(w, struct xe_guc_ct, g2h_worker); 1951 1952 receive_g2h(ct); 1953 } 1954 1955 static struct xe_guc_ct_snapshot *guc_ct_snapshot_alloc(struct xe_guc_ct *ct, bool atomic, 1956 bool want_ctb) 1957 { 1958 struct xe_guc_ct_snapshot *snapshot; 1959 1960 snapshot = kzalloc_obj(*snapshot, atomic ? GFP_ATOMIC : GFP_KERNEL); 1961 if (!snapshot) 1962 return NULL; 1963 1964 if (ct->bo && want_ctb) { 1965 snapshot->ctb_size = xe_bo_size(ct->bo); 1966 snapshot->ctb = kmalloc(snapshot->ctb_size, atomic ? GFP_ATOMIC : GFP_KERNEL); 1967 } 1968 1969 return snapshot; 1970 } 1971 1972 static void guc_ctb_snapshot_capture(struct xe_device *xe, struct guc_ctb *ctb, 1973 struct guc_ctb_snapshot *snapshot) 1974 { 1975 xe_map_memcpy_from(xe, &snapshot->desc, &ctb->desc, 0, 1976 sizeof(struct guc_ct_buffer_desc)); 1977 memcpy(&snapshot->info, &ctb->info, sizeof(struct guc_ctb_info)); 1978 } 1979 1980 static void guc_ctb_snapshot_print(struct guc_ctb_snapshot *snapshot, 1981 struct drm_printer *p) 1982 { 1983 drm_printf(p, "\tsize: %d\n", snapshot->info.size); 1984 drm_printf(p, "\tresv_space: %d\n", snapshot->info.resv_space); 1985 drm_printf(p, "\thead: %d\n", snapshot->info.head); 1986 drm_printf(p, "\ttail: %d\n", snapshot->info.tail); 1987 drm_printf(p, "\tspace: %d\n", snapshot->info.space); 1988 drm_printf(p, "\tbroken: %d\n", snapshot->info.broken); 1989 drm_printf(p, "\thead (memory): %d\n", snapshot->desc.head); 1990 drm_printf(p, "\ttail (memory): %d\n", snapshot->desc.tail); 1991 drm_printf(p, "\tstatus (memory): 0x%x\n", snapshot->desc.status); 1992 } 1993 1994 static struct xe_guc_ct_snapshot *guc_ct_snapshot_capture(struct xe_guc_ct *ct, bool atomic, 1995 bool want_ctb) 1996 { 1997 struct xe_device *xe = ct_to_xe(ct); 1998 struct xe_guc_ct_snapshot *snapshot; 1999 2000 snapshot = guc_ct_snapshot_alloc(ct, atomic, want_ctb); 2001 if (!snapshot) { 2002 xe_gt_err(ct_to_gt(ct), "Skipping CTB snapshot entirely.\n"); 2003 return NULL; 2004 } 2005 2006 if (xe_guc_ct_enabled(ct) || ct->state == XE_GUC_CT_STATE_STOPPED) { 2007 snapshot->ct_enabled = true; 2008 snapshot->g2h_outstanding = READ_ONCE(ct->g2h_outstanding); 2009 guc_ctb_snapshot_capture(xe, &ct->ctbs.h2g, &snapshot->h2g); 2010 guc_ctb_snapshot_capture(xe, &ct->ctbs.g2h, &snapshot->g2h); 2011 } 2012 2013 if (ct->bo && snapshot->ctb) 2014 xe_map_memcpy_from(xe, snapshot->ctb, &ct->bo->vmap, 0, snapshot->ctb_size); 2015 2016 return snapshot; 2017 } 2018 2019 /** 2020 * xe_guc_ct_snapshot_capture - Take a quick snapshot of the CT state. 2021 * @ct: GuC CT object. 2022 * 2023 * This can be printed out in a later stage like during dev_coredump 2024 * analysis. This is safe to be called during atomic context. 2025 * 2026 * Returns: a GuC CT snapshot object that must be freed by the caller 2027 * by using `xe_guc_ct_snapshot_free`. 2028 */ 2029 struct xe_guc_ct_snapshot *xe_guc_ct_snapshot_capture(struct xe_guc_ct *ct) 2030 { 2031 return guc_ct_snapshot_capture(ct, true, true); 2032 } 2033 2034 /** 2035 * xe_guc_ct_snapshot_print - Print out a given GuC CT snapshot. 2036 * @snapshot: GuC CT snapshot object. 2037 * @p: drm_printer where it will be printed out. 2038 * 2039 * This function prints out a given GuC CT snapshot object. 2040 */ 2041 void xe_guc_ct_snapshot_print(struct xe_guc_ct_snapshot *snapshot, 2042 struct drm_printer *p) 2043 { 2044 if (!snapshot) 2045 return; 2046 2047 if (snapshot->ct_enabled) { 2048 drm_puts(p, "H2G CTB (all sizes in DW):\n"); 2049 guc_ctb_snapshot_print(&snapshot->h2g, p); 2050 2051 drm_puts(p, "G2H CTB (all sizes in DW):\n"); 2052 guc_ctb_snapshot_print(&snapshot->g2h, p); 2053 drm_printf(p, "\tg2h outstanding: %d\n", 2054 snapshot->g2h_outstanding); 2055 2056 if (snapshot->ctb) { 2057 drm_printf(p, "[CTB].length: 0x%zx\n", snapshot->ctb_size); 2058 xe_print_blob_ascii85(p, "[CTB].data", '\n', 2059 snapshot->ctb, 0, snapshot->ctb_size); 2060 } 2061 } else { 2062 drm_puts(p, "CT disabled\n"); 2063 } 2064 } 2065 2066 /** 2067 * xe_guc_ct_snapshot_free - Free all allocated objects for a given snapshot. 2068 * @snapshot: GuC CT snapshot object. 2069 * 2070 * This function free all the memory that needed to be allocated at capture 2071 * time. 2072 */ 2073 void xe_guc_ct_snapshot_free(struct xe_guc_ct_snapshot *snapshot) 2074 { 2075 if (!snapshot) 2076 return; 2077 2078 kfree(snapshot->ctb); 2079 kfree(snapshot); 2080 } 2081 2082 /** 2083 * xe_guc_ct_print - GuC CT Print. 2084 * @ct: GuC CT. 2085 * @p: drm_printer where it will be printed out. 2086 * @want_ctb: Should the full CTB content be dumped (vs just the headers) 2087 * 2088 * This function will quickly capture a snapshot of the CT state 2089 * and immediately print it out. 2090 */ 2091 void xe_guc_ct_print(struct xe_guc_ct *ct, struct drm_printer *p, bool want_ctb) 2092 { 2093 struct xe_guc_ct_snapshot *snapshot; 2094 2095 snapshot = guc_ct_snapshot_capture(ct, false, want_ctb); 2096 xe_guc_ct_snapshot_print(snapshot, p); 2097 xe_guc_ct_snapshot_free(snapshot); 2098 } 2099 2100 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG) 2101 2102 #ifdef CONFIG_FUNCTION_ERROR_INJECTION 2103 /* 2104 * This is a helper function which assists the driver in identifying if a fault 2105 * injection test is currently active, allowing it to reduce unnecessary debug 2106 * output. Typically, the function returns zero, but the fault injection 2107 * framework can alter this to return an error. Since faults are injected 2108 * through this function, it's important to ensure the compiler doesn't optimize 2109 * it into an inline function. To avoid such optimization, the 'noinline' 2110 * attribute is applied. Compiler optimizes the static function defined in the 2111 * header file as an inline function. 2112 */ 2113 noinline int xe_is_injection_active(void) { return 0; } 2114 ALLOW_ERROR_INJECTION(xe_is_injection_active, ERRNO); 2115 #else 2116 int xe_is_injection_active(void) { return 0; } 2117 #endif 2118 2119 static void ct_dead_capture(struct xe_guc_ct *ct, struct guc_ctb *ctb, u32 reason_code) 2120 { 2121 struct xe_guc_log_snapshot *snapshot_log; 2122 struct xe_guc_ct_snapshot *snapshot_ct; 2123 struct xe_guc *guc = ct_to_guc(ct); 2124 unsigned long flags; 2125 bool have_capture; 2126 2127 if (ctb) 2128 ctb->info.broken = true; 2129 /* 2130 * Huge dump is getting generated when injecting error for guc CT/MMIO 2131 * functions. So, let us suppress the dump when fault is injected. 2132 */ 2133 if (xe_is_injection_active()) 2134 return; 2135 2136 /* Ignore further errors after the first dump until a reset */ 2137 if (ct->dead.reported) 2138 return; 2139 2140 spin_lock_irqsave(&ct->dead.lock, flags); 2141 2142 /* And only capture one dump at a time */ 2143 have_capture = ct->dead.reason & (1 << CT_DEAD_STATE_CAPTURE); 2144 ct->dead.reason |= (1 << reason_code) | 2145 (1 << CT_DEAD_STATE_CAPTURE); 2146 2147 spin_unlock_irqrestore(&ct->dead.lock, flags); 2148 2149 if (have_capture) 2150 return; 2151 2152 snapshot_log = xe_guc_log_snapshot_capture(&guc->log, true); 2153 snapshot_ct = xe_guc_ct_snapshot_capture((ct)); 2154 2155 spin_lock_irqsave(&ct->dead.lock, flags); 2156 2157 if (ct->dead.snapshot_log || ct->dead.snapshot_ct) { 2158 xe_gt_err(ct_to_gt(ct), "Got unexpected dead CT capture!\n"); 2159 xe_guc_log_snapshot_free(snapshot_log); 2160 xe_guc_ct_snapshot_free(snapshot_ct); 2161 } else { 2162 ct->dead.snapshot_log = snapshot_log; 2163 ct->dead.snapshot_ct = snapshot_ct; 2164 } 2165 2166 spin_unlock_irqrestore(&ct->dead.lock, flags); 2167 2168 queue_work(system_unbound_wq, &(ct)->dead.worker); 2169 } 2170 2171 static void ct_dead_print(struct xe_dead_ct *dead) 2172 { 2173 struct xe_guc_ct *ct = container_of(dead, struct xe_guc_ct, dead); 2174 struct xe_device *xe = ct_to_xe(ct); 2175 struct xe_gt *gt = ct_to_gt(ct); 2176 static int g_count; 2177 struct drm_printer ip = xe_gt_info_printer(gt); 2178 struct drm_printer lp = drm_line_printer(&ip, "Capture", ++g_count); 2179 2180 if (!dead->reason) { 2181 xe_gt_err(gt, "CTB is dead for no reason!?\n"); 2182 return; 2183 } 2184 2185 /* Can't generate a genuine core dump at this point, so just do the good bits */ 2186 drm_puts(&lp, "**** Xe Device Coredump ****\n"); 2187 drm_printf(&lp, "Reason: CTB is dead - 0x%X\n", dead->reason); 2188 xe_device_snapshot_print(xe, &lp); 2189 2190 drm_printf(&lp, "**** GT #%d ****\n", gt->info.id); 2191 drm_printf(&lp, "\tTile: %d\n", gt->tile->id); 2192 2193 drm_puts(&lp, "**** GuC Log ****\n"); 2194 xe_guc_log_snapshot_print(dead->snapshot_log, &lp); 2195 2196 drm_puts(&lp, "**** GuC CT ****\n"); 2197 xe_guc_ct_snapshot_print(dead->snapshot_ct, &lp); 2198 2199 drm_puts(&lp, "Done.\n"); 2200 } 2201 2202 static void ct_dead_worker_func(struct work_struct *w) 2203 { 2204 struct xe_guc_ct *ct = container_of(w, struct xe_guc_ct, dead.worker); 2205 2206 if (!ct->dead.reported) { 2207 ct->dead.reported = true; 2208 ct_dead_print(&ct->dead); 2209 } 2210 2211 spin_lock_irq(&ct->dead.lock); 2212 2213 xe_guc_log_snapshot_free(ct->dead.snapshot_log); 2214 ct->dead.snapshot_log = NULL; 2215 xe_guc_ct_snapshot_free(ct->dead.snapshot_ct); 2216 ct->dead.snapshot_ct = NULL; 2217 2218 if (ct->dead.reason & (1 << CT_DEAD_STATE_REARM)) { 2219 /* A reset has occurred so re-arm the error reporting */ 2220 ct->dead.reason = 0; 2221 ct->dead.reported = false; 2222 } 2223 2224 spin_unlock_irq(&ct->dead.lock); 2225 } 2226 #endif 2227