1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB 2 /* Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. */ 3 4 #include <rdma/ib_umem_odp.h> 5 #include "mlx5_ib.h" 6 #include "umr.h" 7 #include "wr.h" 8 9 /* 10 * We can't use an array for xlt_emergency_page because dma_map_single doesn't 11 * work on kernel modules memory 12 */ 13 void *xlt_emergency_page; 14 static DEFINE_MUTEX(xlt_emergency_page_mutex); 15 16 static __be64 get_umr_enable_mr_mask(void) 17 { 18 u64 result; 19 20 result = MLX5_MKEY_MASK_KEY | 21 MLX5_MKEY_MASK_FREE; 22 23 return cpu_to_be64(result); 24 } 25 26 static __be64 get_umr_disable_mr_mask(void) 27 { 28 u64 result; 29 30 result = MLX5_MKEY_MASK_FREE; 31 32 return cpu_to_be64(result); 33 } 34 35 static __be64 get_umr_update_translation_mask(struct mlx5_ib_dev *dev) 36 { 37 u64 result; 38 39 result = MLX5_MKEY_MASK_LEN | 40 MLX5_MKEY_MASK_PAGE_SIZE | 41 MLX5_MKEY_MASK_START_ADDR; 42 if (MLX5_CAP_GEN_2(dev->mdev, umr_log_entity_size_5)) 43 result |= MLX5_MKEY_MASK_PAGE_SIZE_5; 44 45 return cpu_to_be64(result); 46 } 47 48 static __be64 get_umr_update_access_mask(struct mlx5_ib_dev *dev) 49 { 50 u64 result; 51 52 result = MLX5_MKEY_MASK_LR | 53 MLX5_MKEY_MASK_LW | 54 MLX5_MKEY_MASK_RR | 55 MLX5_MKEY_MASK_RW; 56 57 if (MLX5_CAP_GEN(dev->mdev, atomic)) 58 result |= MLX5_MKEY_MASK_A; 59 60 if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write_umr)) 61 result |= MLX5_MKEY_MASK_RELAXED_ORDERING_WRITE; 62 63 if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read_umr)) 64 result |= MLX5_MKEY_MASK_RELAXED_ORDERING_READ; 65 66 return cpu_to_be64(result); 67 } 68 69 static __be64 get_umr_update_pd_mask(void) 70 { 71 u64 result; 72 73 result = MLX5_MKEY_MASK_PD; 74 75 return cpu_to_be64(result); 76 } 77 78 static int umr_check_mkey_mask(struct mlx5_ib_dev *dev, u64 mask) 79 { 80 if (mask & MLX5_MKEY_MASK_PAGE_SIZE && 81 MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled)) 82 return -EPERM; 83 84 if (mask & MLX5_MKEY_MASK_A && 85 MLX5_CAP_GEN(dev->mdev, umr_modify_atomic_disabled)) 86 return -EPERM; 87 88 if (mask & MLX5_MKEY_MASK_RELAXED_ORDERING_WRITE && 89 !MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write_umr)) 90 return -EPERM; 91 92 if (mask & MLX5_MKEY_MASK_RELAXED_ORDERING_READ && 93 !MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read_umr)) 94 return -EPERM; 95 96 return 0; 97 } 98 99 enum { 100 MAX_UMR_WR = 128, 101 }; 102 103 static int mlx5r_umr_qp_rst2rts(struct mlx5_ib_dev *dev, struct ib_qp *qp) 104 { 105 struct ib_qp_attr attr = {}; 106 int ret; 107 108 attr.qp_state = IB_QPS_INIT; 109 attr.port_num = 1; 110 ret = ib_modify_qp(qp, &attr, 111 IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT); 112 if (ret) { 113 mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n"); 114 return ret; 115 } 116 117 memset(&attr, 0, sizeof(attr)); 118 attr.qp_state = IB_QPS_RTR; 119 120 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 121 if (ret) { 122 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n"); 123 return ret; 124 } 125 126 memset(&attr, 0, sizeof(attr)); 127 attr.qp_state = IB_QPS_RTS; 128 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 129 if (ret) { 130 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n"); 131 return ret; 132 } 133 134 return 0; 135 } 136 137 int mlx5r_umr_resource_init(struct mlx5_ib_dev *dev) 138 { 139 struct ib_qp_init_attr init_attr = {}; 140 struct ib_cq *cq; 141 struct ib_qp *qp; 142 int ret = 0; 143 144 145 /* 146 * UMR qp is set once, never changed until device unload. 147 * Avoid taking the mutex if initialization is already done. 148 */ 149 if (dev->umrc.qp) 150 return 0; 151 152 mutex_lock(&dev->umrc.init_lock); 153 /* First user allocates the UMR resources. Skip if already allocated. */ 154 if (dev->umrc.qp) 155 goto unlock; 156 157 cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ); 158 if (IS_ERR(cq)) { 159 mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n"); 160 ret = PTR_ERR(cq); 161 goto unlock; 162 } 163 164 init_attr.send_cq = cq; 165 init_attr.recv_cq = cq; 166 init_attr.sq_sig_type = IB_SIGNAL_ALL_WR; 167 init_attr.cap.max_send_wr = MAX_UMR_WR; 168 init_attr.cap.max_send_sge = 1; 169 init_attr.qp_type = MLX5_IB_QPT_REG_UMR; 170 init_attr.port_num = 1; 171 qp = ib_create_qp(dev->umrc.pd, &init_attr); 172 if (IS_ERR(qp)) { 173 mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n"); 174 ret = PTR_ERR(qp); 175 goto destroy_cq; 176 } 177 178 ret = mlx5r_umr_qp_rst2rts(dev, qp); 179 if (ret) 180 goto destroy_qp; 181 182 dev->umrc.cq = cq; 183 184 sema_init(&dev->umrc.sem, MAX_UMR_WR); 185 mutex_init(&dev->umrc.lock); 186 dev->umrc.state = MLX5_UMR_STATE_ACTIVE; 187 dev->umrc.qp = qp; 188 189 mutex_unlock(&dev->umrc.init_lock); 190 return 0; 191 192 destroy_qp: 193 ib_destroy_qp(qp); 194 destroy_cq: 195 ib_free_cq(cq); 196 unlock: 197 mutex_unlock(&dev->umrc.init_lock); 198 return ret; 199 } 200 201 void mlx5r_umr_resource_cleanup(struct mlx5_ib_dev *dev) 202 { 203 if (dev->umrc.state == MLX5_UMR_STATE_UNINIT) 204 return; 205 mutex_destroy(&dev->umrc.lock); 206 /* After device init, UMR cp/qp are not unset during the lifetime. */ 207 ib_destroy_qp(dev->umrc.qp); 208 ib_free_cq(dev->umrc.cq); 209 } 210 211 int mlx5r_umr_init(struct mlx5_ib_dev *dev) 212 { 213 struct ib_pd *pd; 214 215 pd = ib_alloc_pd(&dev->ib_dev, 0); 216 if (IS_ERR(pd)) { 217 mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n"); 218 return PTR_ERR(pd); 219 } 220 dev->umrc.pd = pd; 221 222 mutex_init(&dev->umrc.init_lock); 223 224 return 0; 225 } 226 227 void mlx5r_umr_cleanup(struct mlx5_ib_dev *dev) 228 { 229 if (!dev->umrc.pd) 230 return; 231 232 mutex_destroy(&dev->umrc.init_lock); 233 ib_dealloc_pd(dev->umrc.pd); 234 } 235 236 237 static int mlx5r_umr_post_send(struct ib_qp *ibqp, u32 mkey, struct ib_cqe *cqe, 238 struct mlx5r_umr_wqe *wqe, bool with_data) 239 { 240 unsigned int wqe_size = 241 with_data ? sizeof(struct mlx5r_umr_wqe) : 242 sizeof(struct mlx5r_umr_wqe) - 243 sizeof(struct mlx5_wqe_data_seg); 244 struct mlx5_ib_dev *dev = to_mdev(ibqp->device); 245 struct mlx5_core_dev *mdev = dev->mdev; 246 struct mlx5_ib_qp *qp = to_mqp(ibqp); 247 struct mlx5_wqe_ctrl_seg *ctrl; 248 union { 249 struct ib_cqe *ib_cqe; 250 u64 wr_id; 251 } id; 252 void *cur_edge, *seg; 253 unsigned long flags; 254 unsigned int idx; 255 int size, err; 256 257 if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)) 258 return -EIO; 259 260 spin_lock_irqsave(&qp->sq.lock, flags); 261 262 err = mlx5r_begin_wqe(qp, &seg, &ctrl, &idx, &size, &cur_edge, 0, 263 cpu_to_be32(mkey), false, false); 264 if (WARN_ON(err)) 265 goto out; 266 267 qp->sq.wr_data[idx] = MLX5_IB_WR_UMR; 268 269 mlx5r_memcpy_send_wqe(&qp->sq, &cur_edge, &seg, &size, wqe, wqe_size); 270 271 id.ib_cqe = cqe; 272 mlx5r_finish_wqe(qp, ctrl, seg, size, cur_edge, idx, id.wr_id, 0, 273 MLX5_FENCE_MODE_INITIATOR_SMALL, MLX5_OPCODE_UMR); 274 275 mlx5r_ring_db(qp, 1, ctrl); 276 277 out: 278 spin_unlock_irqrestore(&qp->sq.lock, flags); 279 280 return err; 281 } 282 283 static int mlx5r_umr_recover(struct mlx5_ib_dev *dev, u32 mkey, 284 struct mlx5r_umr_context *umr_context, 285 struct mlx5r_umr_wqe *wqe, bool with_data) 286 { 287 struct umr_common *umrc = &dev->umrc; 288 struct ib_qp_attr attr; 289 int err; 290 291 mutex_lock(&umrc->lock); 292 /* Preventing any further WRs to be sent now */ 293 if (umrc->state != MLX5_UMR_STATE_RECOVER) { 294 mlx5_ib_warn(dev, "UMR recovery encountered an unexpected state=%d\n", 295 umrc->state); 296 umrc->state = MLX5_UMR_STATE_RECOVER; 297 } 298 mutex_unlock(&umrc->lock); 299 300 /* Sending a final/barrier WR (the failed one) and wait for its completion. 301 * This will ensure that all the previous WRs got a completion before 302 * we set the QP state to RESET. 303 */ 304 err = mlx5r_umr_post_send(umrc->qp, mkey, &umr_context->cqe, wqe, 305 with_data); 306 if (err) { 307 mlx5_ib_warn(dev, "UMR recovery post send failed, err %d\n", err); 308 goto err; 309 } 310 311 /* Since the QP is in an error state, it will only receive 312 * IB_WC_WR_FLUSH_ERR. However, as it serves only as a barrier 313 * we don't care about its status. 314 */ 315 wait_for_completion(&umr_context->done); 316 317 attr.qp_state = IB_QPS_RESET; 318 err = ib_modify_qp(umrc->qp, &attr, IB_QP_STATE); 319 if (err) { 320 mlx5_ib_warn(dev, "Couldn't modify UMR QP to RESET, err=%d\n", err); 321 goto err; 322 } 323 324 err = mlx5r_umr_qp_rst2rts(dev, umrc->qp); 325 if (err) { 326 mlx5_ib_warn(dev, "Couldn't modify UMR QP to RTS, err=%d\n", err); 327 goto err; 328 } 329 330 umrc->state = MLX5_UMR_STATE_ACTIVE; 331 return 0; 332 333 err: 334 umrc->state = MLX5_UMR_STATE_ERR; 335 return err; 336 } 337 338 static void mlx5r_umr_done(struct ib_cq *cq, struct ib_wc *wc) 339 { 340 struct mlx5_ib_umr_context *context = 341 container_of(wc->wr_cqe, struct mlx5_ib_umr_context, cqe); 342 343 context->status = wc->status; 344 complete(&context->done); 345 } 346 347 static inline void mlx5r_umr_init_context(struct mlx5r_umr_context *context) 348 { 349 context->cqe.done = mlx5r_umr_done; 350 init_completion(&context->done); 351 } 352 353 static int mlx5r_umr_post_send_wait(struct mlx5_ib_dev *dev, u32 mkey, 354 struct mlx5r_umr_wqe *wqe, bool with_data) 355 { 356 struct umr_common *umrc = &dev->umrc; 357 struct mlx5r_umr_context umr_context; 358 int err; 359 360 err = umr_check_mkey_mask(dev, be64_to_cpu(wqe->ctrl_seg.mkey_mask)); 361 if (WARN_ON(err)) 362 return err; 363 364 mlx5r_umr_init_context(&umr_context); 365 366 down(&umrc->sem); 367 while (true) { 368 mutex_lock(&umrc->lock); 369 if (umrc->state == MLX5_UMR_STATE_ERR) { 370 mutex_unlock(&umrc->lock); 371 err = -EFAULT; 372 break; 373 } 374 375 if (umrc->state == MLX5_UMR_STATE_RECOVER) { 376 mutex_unlock(&umrc->lock); 377 usleep_range(3000, 5000); 378 continue; 379 } 380 381 err = mlx5r_umr_post_send(umrc->qp, mkey, &umr_context.cqe, wqe, 382 with_data); 383 mutex_unlock(&umrc->lock); 384 if (err) { 385 mlx5_ib_warn(dev, "UMR post send failed, err %d\n", 386 err); 387 break; 388 } 389 390 wait_for_completion(&umr_context.done); 391 392 if (umr_context.status == IB_WC_SUCCESS) 393 break; 394 395 if (umr_context.status == IB_WC_WR_FLUSH_ERR) 396 continue; 397 398 WARN_ON_ONCE(1); 399 mlx5_ib_warn(dev, 400 "reg umr failed (%u). Trying to recover and resubmit the flushed WQEs, mkey = %u\n", 401 umr_context.status, mkey); 402 err = mlx5r_umr_recover(dev, mkey, &umr_context, wqe, with_data); 403 if (err) 404 mlx5_ib_warn(dev, "couldn't recover UMR, err %d\n", 405 err); 406 err = -EFAULT; 407 break; 408 } 409 up(&umrc->sem); 410 return err; 411 } 412 413 /** 414 * mlx5r_umr_revoke_mr - Fence all DMA on the MR 415 * @mr: The MR to fence 416 * 417 * Upon return the NIC will not be doing any DMA to the pages under the MR, 418 * and any DMA in progress will be completed. Failure of this function 419 * indicates the HW has failed catastrophically. 420 */ 421 int mlx5r_umr_revoke_mr(struct mlx5_ib_mr *mr) 422 { 423 struct mlx5_ib_dev *dev = mr_to_mdev(mr); 424 struct mlx5r_umr_wqe wqe = {}; 425 426 if (dev->mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) 427 return 0; 428 429 wqe.ctrl_seg.mkey_mask |= get_umr_update_pd_mask(); 430 wqe.ctrl_seg.mkey_mask |= get_umr_disable_mr_mask(); 431 wqe.ctrl_seg.flags |= MLX5_UMR_INLINE; 432 433 MLX5_SET(mkc, &wqe.mkey_seg, free, 1); 434 MLX5_SET(mkc, &wqe.mkey_seg, pd, to_mpd(dev->umrc.pd)->pdn); 435 MLX5_SET(mkc, &wqe.mkey_seg, qpn, 0xffffff); 436 MLX5_SET(mkc, &wqe.mkey_seg, mkey_7_0, 437 mlx5_mkey_variant(mr->mmkey.key)); 438 439 return mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, false); 440 } 441 442 static void mlx5r_umr_set_access_flags(struct mlx5_ib_dev *dev, 443 struct mlx5_mkey_seg *seg, 444 unsigned int access_flags) 445 { 446 bool ro_read = (access_flags & IB_ACCESS_RELAXED_ORDERING) && 447 (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read) || 448 pcie_relaxed_ordering_enabled(dev->mdev->pdev)); 449 450 MLX5_SET(mkc, seg, a, !!(access_flags & IB_ACCESS_REMOTE_ATOMIC)); 451 MLX5_SET(mkc, seg, rw, !!(access_flags & IB_ACCESS_REMOTE_WRITE)); 452 MLX5_SET(mkc, seg, rr, !!(access_flags & IB_ACCESS_REMOTE_READ)); 453 MLX5_SET(mkc, seg, lw, !!(access_flags & IB_ACCESS_LOCAL_WRITE)); 454 MLX5_SET(mkc, seg, lr, 1); 455 MLX5_SET(mkc, seg, relaxed_ordering_write, 456 !!(access_flags & IB_ACCESS_RELAXED_ORDERING)); 457 MLX5_SET(mkc, seg, relaxed_ordering_read, ro_read); 458 } 459 460 int mlx5r_umr_rereg_pd_access(struct mlx5_ib_mr *mr, struct ib_pd *pd, 461 int access_flags) 462 { 463 struct mlx5_ib_dev *dev = mr_to_mdev(mr); 464 struct mlx5r_umr_wqe wqe = {}; 465 int err; 466 467 wqe.ctrl_seg.mkey_mask = get_umr_update_access_mask(dev); 468 wqe.ctrl_seg.mkey_mask |= get_umr_update_pd_mask(); 469 wqe.ctrl_seg.flags = MLX5_UMR_CHECK_FREE; 470 wqe.ctrl_seg.flags |= MLX5_UMR_INLINE; 471 472 mlx5r_umr_set_access_flags(dev, &wqe.mkey_seg, access_flags); 473 MLX5_SET(mkc, &wqe.mkey_seg, pd, to_mpd(pd)->pdn); 474 MLX5_SET(mkc, &wqe.mkey_seg, qpn, 0xffffff); 475 MLX5_SET(mkc, &wqe.mkey_seg, mkey_7_0, 476 mlx5_mkey_variant(mr->mmkey.key)); 477 478 err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, false); 479 if (err) 480 return err; 481 482 mr->access_flags = access_flags; 483 return 0; 484 } 485 486 #define MLX5_MAX_UMR_CHUNK \ 487 ((1 << (MLX5_MAX_UMR_SHIFT + 4)) - MLX5_UMR_FLEX_ALIGNMENT) 488 #define MLX5_SPARE_UMR_CHUNK 0x10000 489 490 /* 491 * Allocate a temporary buffer to hold the per-page information to transfer to 492 * HW. For efficiency this should be as large as it can be, but buffer 493 * allocation failure is not allowed, so try smaller sizes. 494 */ 495 static void *mlx5r_umr_alloc_xlt(size_t *nents, size_t ent_size, gfp_t gfp_mask) 496 { 497 const size_t xlt_chunk_align = MLX5_UMR_FLEX_ALIGNMENT / ent_size; 498 size_t size; 499 void *res = NULL; 500 501 static_assert(PAGE_SIZE % MLX5_UMR_FLEX_ALIGNMENT == 0); 502 503 /* 504 * MLX5_IB_UPD_XLT_ATOMIC doesn't signal an atomic context just that the 505 * allocation can't trigger any kind of reclaim. 506 */ 507 might_sleep(); 508 509 gfp_mask |= __GFP_ZERO | __GFP_NORETRY; 510 511 /* 512 * If the system already has a suitable high order page then just use 513 * that, but don't try hard to create one. This max is about 1M, so a 514 * free x86 huge page will satisfy it. 515 */ 516 size = min_t(size_t, ent_size * ALIGN(*nents, xlt_chunk_align), 517 MLX5_MAX_UMR_CHUNK); 518 *nents = size / ent_size; 519 res = (void *)__get_free_pages(gfp_mask | __GFP_NOWARN, 520 get_order(size)); 521 if (res) 522 return res; 523 524 if (size > MLX5_SPARE_UMR_CHUNK) { 525 size = MLX5_SPARE_UMR_CHUNK; 526 *nents = size / ent_size; 527 res = (void *)__get_free_pages(gfp_mask | __GFP_NOWARN, 528 get_order(size)); 529 if (res) 530 return res; 531 } 532 533 *nents = PAGE_SIZE / ent_size; 534 res = (void *)__get_free_page(gfp_mask); 535 if (res) 536 return res; 537 538 mutex_lock(&xlt_emergency_page_mutex); 539 memset(xlt_emergency_page, 0, PAGE_SIZE); 540 return xlt_emergency_page; 541 } 542 543 static void mlx5r_umr_free_xlt(void *xlt, size_t length) 544 { 545 if (xlt == xlt_emergency_page) { 546 mutex_unlock(&xlt_emergency_page_mutex); 547 return; 548 } 549 550 free_pages((unsigned long)xlt, get_order(length)); 551 } 552 553 static void mlx5r_umr_unmap_free_xlt(struct mlx5_ib_dev *dev, void *xlt, 554 struct ib_sge *sg) 555 { 556 struct device *ddev = &dev->mdev->pdev->dev; 557 558 dma_unmap_single(ddev, sg->addr, sg->length, DMA_TO_DEVICE); 559 mlx5r_umr_free_xlt(xlt, sg->length); 560 } 561 562 /* 563 * Create an XLT buffer ready for submission. 564 */ 565 static void *mlx5r_umr_create_xlt(struct mlx5_ib_dev *dev, struct ib_sge *sg, 566 size_t nents, size_t ent_size, 567 unsigned int flags) 568 { 569 struct device *ddev = &dev->mdev->pdev->dev; 570 dma_addr_t dma; 571 void *xlt; 572 573 xlt = mlx5r_umr_alloc_xlt(&nents, ent_size, 574 flags & MLX5_IB_UPD_XLT_ATOMIC ? GFP_ATOMIC : 575 GFP_KERNEL); 576 sg->length = nents * ent_size; 577 dma = dma_map_single(ddev, xlt, sg->length, DMA_TO_DEVICE); 578 if (dma_mapping_error(ddev, dma)) { 579 mlx5_ib_err(dev, "unable to map DMA during XLT update.\n"); 580 mlx5r_umr_free_xlt(xlt, sg->length); 581 return NULL; 582 } 583 sg->addr = dma; 584 sg->lkey = dev->umrc.pd->local_dma_lkey; 585 586 return xlt; 587 } 588 589 static void 590 mlx5r_umr_set_update_xlt_ctrl_seg(struct mlx5_wqe_umr_ctrl_seg *ctrl_seg, 591 unsigned int flags, struct ib_sge *sg) 592 { 593 if (!(flags & MLX5_IB_UPD_XLT_ENABLE)) 594 /* fail if free */ 595 ctrl_seg->flags = MLX5_UMR_CHECK_FREE; 596 else 597 /* fail if not free */ 598 ctrl_seg->flags = MLX5_UMR_CHECK_NOT_FREE; 599 ctrl_seg->xlt_octowords = 600 cpu_to_be16(mlx5r_umr_get_xlt_octo(sg->length)); 601 } 602 603 static void mlx5r_umr_set_update_xlt_mkey_seg(struct mlx5_ib_dev *dev, 604 struct mlx5_mkey_seg *mkey_seg, 605 struct mlx5_ib_mr *mr, 606 unsigned int page_shift) 607 { 608 mlx5r_umr_set_access_flags(dev, mkey_seg, mr->access_flags); 609 MLX5_SET(mkc, mkey_seg, pd, to_mpd(mr->ibmr.pd)->pdn); 610 MLX5_SET64(mkc, mkey_seg, start_addr, mr->ibmr.iova); 611 MLX5_SET64(mkc, mkey_seg, len, mr->ibmr.length); 612 MLX5_SET(mkc, mkey_seg, log_page_size, page_shift); 613 MLX5_SET(mkc, mkey_seg, qpn, 0xffffff); 614 MLX5_SET(mkc, mkey_seg, mkey_7_0, mlx5_mkey_variant(mr->mmkey.key)); 615 } 616 617 static void 618 mlx5r_umr_set_update_xlt_data_seg(struct mlx5_wqe_data_seg *data_seg, 619 struct ib_sge *sg) 620 { 621 data_seg->byte_count = cpu_to_be32(sg->length); 622 data_seg->lkey = cpu_to_be32(sg->lkey); 623 data_seg->addr = cpu_to_be64(sg->addr); 624 } 625 626 static void mlx5r_umr_update_offset(struct mlx5_wqe_umr_ctrl_seg *ctrl_seg, 627 u64 offset) 628 { 629 u64 octo_offset = mlx5r_umr_get_xlt_octo(offset); 630 631 ctrl_seg->xlt_offset = cpu_to_be16(octo_offset & 0xffff); 632 ctrl_seg->xlt_offset_47_16 = cpu_to_be32(octo_offset >> 16); 633 ctrl_seg->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN; 634 } 635 636 static void mlx5r_umr_final_update_xlt(struct mlx5_ib_dev *dev, 637 struct mlx5r_umr_wqe *wqe, 638 struct mlx5_ib_mr *mr, struct ib_sge *sg, 639 unsigned int flags) 640 { 641 bool update_pd_access, update_translation; 642 643 if (flags & MLX5_IB_UPD_XLT_ENABLE) 644 wqe->ctrl_seg.mkey_mask |= get_umr_enable_mr_mask(); 645 646 update_pd_access = flags & MLX5_IB_UPD_XLT_ENABLE || 647 flags & MLX5_IB_UPD_XLT_PD || 648 flags & MLX5_IB_UPD_XLT_ACCESS; 649 650 if (update_pd_access) { 651 wqe->ctrl_seg.mkey_mask |= get_umr_update_access_mask(dev); 652 wqe->ctrl_seg.mkey_mask |= get_umr_update_pd_mask(); 653 } 654 655 update_translation = 656 flags & MLX5_IB_UPD_XLT_ENABLE || flags & MLX5_IB_UPD_XLT_ADDR; 657 658 if (update_translation) { 659 wqe->ctrl_seg.mkey_mask |= get_umr_update_translation_mask(dev); 660 if (!mr->ibmr.length) 661 MLX5_SET(mkc, &wqe->mkey_seg, length64, 1); 662 if (flags & MLX5_IB_UPD_XLT_KEEP_PGSZ) 663 wqe->ctrl_seg.mkey_mask &= ~MLX5_MKEY_MASK_PAGE_SIZE; 664 } 665 666 wqe->ctrl_seg.xlt_octowords = 667 cpu_to_be16(mlx5r_umr_get_xlt_octo(sg->length)); 668 wqe->data_seg.byte_count = cpu_to_be32(sg->length); 669 } 670 671 static void 672 _mlx5r_umr_init_wqe(struct mlx5_ib_mr *mr, struct mlx5r_umr_wqe *wqe, 673 struct ib_sge *sg, unsigned int flags, 674 unsigned int page_shift, bool dd) 675 { 676 struct mlx5_ib_dev *dev = mr_to_mdev(mr); 677 678 mlx5r_umr_set_update_xlt_ctrl_seg(&wqe->ctrl_seg, flags, sg); 679 mlx5r_umr_set_update_xlt_mkey_seg(dev, &wqe->mkey_seg, mr, page_shift); 680 if (dd) /* Use the data direct internal kernel PD */ 681 MLX5_SET(mkc, &wqe->mkey_seg, pd, dev->ddr.pdn); 682 mlx5r_umr_set_update_xlt_data_seg(&wqe->data_seg, sg); 683 } 684 685 static int 686 _mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags, bool dd, 687 size_t start_block, size_t nblocks) 688 { 689 size_t ent_size = dd ? sizeof(struct mlx5_ksm) : sizeof(struct mlx5_mtt); 690 struct mlx5_ib_dev *dev = mr_to_mdev(mr); 691 struct device *ddev = &dev->mdev->pdev->dev; 692 struct mlx5r_umr_wqe wqe = {}; 693 size_t processed_blocks = 0; 694 struct ib_block_iter biter; 695 size_t cur_block_idx = 0; 696 struct mlx5_ksm *cur_ksm; 697 struct mlx5_mtt *cur_mtt; 698 size_t orig_sg_length; 699 size_t total_blocks; 700 size_t final_size; 701 void *curr_entry; 702 struct ib_sge sg; 703 void *entry; 704 u64 offset; 705 int err = 0; 706 707 total_blocks = ib_umem_num_dma_blocks(mr->umem, 1UL << mr->page_shift); 708 if (start_block > total_blocks) 709 return -EINVAL; 710 711 /* nblocks 0 means update all blocks starting from start_block */ 712 if (nblocks) 713 total_blocks = nblocks; 714 715 entry = mlx5r_umr_create_xlt(dev, &sg, total_blocks, ent_size, flags); 716 if (!entry) 717 return -ENOMEM; 718 719 orig_sg_length = sg.length; 720 721 _mlx5r_umr_init_wqe(mr, &wqe, &sg, flags, mr->page_shift, dd); 722 723 /* Set initial translation offset to start_block */ 724 offset = (u64)start_block * ent_size; 725 mlx5r_umr_update_offset(&wqe.ctrl_seg, offset); 726 727 if (dd) 728 cur_ksm = entry; 729 else 730 cur_mtt = entry; 731 732 curr_entry = entry; 733 734 rdma_umem_for_each_dma_block(mr->umem, &biter, BIT(mr->page_shift)) { 735 if (cur_block_idx < start_block) { 736 cur_block_idx++; 737 continue; 738 } 739 740 if (nblocks && processed_blocks >= nblocks) 741 break; 742 743 if (curr_entry == entry + sg.length) { 744 dma_sync_single_for_device(ddev, sg.addr, sg.length, 745 DMA_TO_DEVICE); 746 747 err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, 748 true); 749 if (err) 750 goto err; 751 dma_sync_single_for_cpu(ddev, sg.addr, sg.length, 752 DMA_TO_DEVICE); 753 offset += sg.length; 754 mlx5r_umr_update_offset(&wqe.ctrl_seg, offset); 755 if (dd) 756 cur_ksm = entry; 757 else 758 cur_mtt = entry; 759 } 760 761 if (dd) { 762 cur_ksm->va = cpu_to_be64(rdma_block_iter_dma_address(&biter)); 763 cur_ksm->key = cpu_to_be32(dev->ddr.mkey); 764 if (mr->umem->is_dmabuf && 765 (flags & MLX5_IB_UPD_XLT_ZAP)) { 766 cur_ksm->va = 0; 767 cur_ksm->key = 0; 768 } 769 cur_ksm++; 770 curr_entry = cur_ksm; 771 } else { 772 cur_mtt->ptag = 773 cpu_to_be64(rdma_block_iter_dma_address(&biter) | 774 MLX5_IB_MTT_PRESENT); 775 if (mr->umem->is_dmabuf && (flags & MLX5_IB_UPD_XLT_ZAP)) 776 cur_mtt->ptag = 0; 777 cur_mtt++; 778 curr_entry = cur_mtt; 779 } 780 781 processed_blocks++; 782 } 783 784 final_size = curr_entry - entry; 785 sg.length = ALIGN(final_size, MLX5_UMR_FLEX_ALIGNMENT); 786 memset(curr_entry, 0, sg.length - final_size); 787 mlx5r_umr_final_update_xlt(dev, &wqe, mr, &sg, flags); 788 789 dma_sync_single_for_device(ddev, sg.addr, sg.length, DMA_TO_DEVICE); 790 err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, true); 791 792 err: 793 sg.length = orig_sg_length; 794 mlx5r_umr_unmap_free_xlt(dev, entry, &sg); 795 return err; 796 } 797 798 int mlx5r_umr_update_data_direct_ksm_pas_range(struct mlx5_ib_mr *mr, 799 unsigned int flags, 800 size_t start_block, 801 size_t nblocks) 802 { 803 /* No invalidation flow is expected */ 804 if (WARN_ON(!mr->umem->is_dmabuf) || ((flags & MLX5_IB_UPD_XLT_ZAP) && 805 !(flags & MLX5_IB_UPD_XLT_KEEP_PGSZ))) 806 return -EINVAL; 807 808 return _mlx5r_umr_update_mr_pas(mr, flags, true, start_block, nblocks); 809 } 810 811 int mlx5r_umr_update_data_direct_ksm_pas(struct mlx5_ib_mr *mr, 812 unsigned int flags) 813 { 814 return mlx5r_umr_update_data_direct_ksm_pas_range(mr, flags, 0, 0); 815 } 816 817 int mlx5r_umr_update_mr_pas_range(struct mlx5_ib_mr *mr, unsigned int flags, 818 size_t start_block, size_t nblocks) 819 { 820 if (WARN_ON(mr->umem->is_odp)) 821 return -EINVAL; 822 823 return _mlx5r_umr_update_mr_pas(mr, flags, false, start_block, nblocks); 824 } 825 826 /* 827 * Send the DMA list to the HW for a normal MR using UMR. 828 * Dmabuf MR is handled in a similar way, except that the MLX5_IB_UPD_XLT_ZAP 829 * flag may be used. 830 */ 831 int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags) 832 { 833 return mlx5r_umr_update_mr_pas_range(mr, flags, 0, 0); 834 } 835 836 static bool umr_can_use_indirect_mkey(struct mlx5_ib_dev *dev) 837 { 838 return !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled); 839 } 840 841 int mlx5r_umr_update_xlt(struct mlx5_ib_mr *mr, u64 idx, int npages, 842 int page_shift, int flags) 843 { 844 int desc_size = (flags & MLX5_IB_UPD_XLT_INDIRECT) 845 ? sizeof(struct mlx5_klm) 846 : sizeof(struct mlx5_mtt); 847 const int page_align = MLX5_UMR_FLEX_ALIGNMENT / desc_size; 848 struct mlx5_ib_dev *dev = mr_to_mdev(mr); 849 struct device *ddev = &dev->mdev->pdev->dev; 850 const int page_mask = page_align - 1; 851 struct mlx5r_umr_wqe wqe = {}; 852 size_t pages_mapped = 0; 853 size_t pages_to_map = 0; 854 size_t size_to_map = 0; 855 size_t orig_sg_length; 856 size_t pages_iter; 857 struct ib_sge sg; 858 int err = 0; 859 void *xlt; 860 861 if ((flags & MLX5_IB_UPD_XLT_INDIRECT) && 862 !umr_can_use_indirect_mkey(dev)) 863 return -EPERM; 864 865 if (WARN_ON(!mr->umem->is_odp)) 866 return -EINVAL; 867 868 /* UMR copies MTTs in units of MLX5_UMR_FLEX_ALIGNMENT bytes, 869 * so we need to align the offset and length accordingly 870 */ 871 if (idx & page_mask) { 872 npages += idx & page_mask; 873 idx &= ~page_mask; 874 } 875 pages_to_map = ALIGN(npages, page_align); 876 877 xlt = mlx5r_umr_create_xlt(dev, &sg, npages, desc_size, flags); 878 if (!xlt) 879 return -ENOMEM; 880 881 pages_iter = sg.length / desc_size; 882 orig_sg_length = sg.length; 883 884 if (!(flags & MLX5_IB_UPD_XLT_INDIRECT)) { 885 struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem); 886 size_t max_pages = ib_umem_odp_num_pages(odp) - idx; 887 888 pages_to_map = min_t(size_t, pages_to_map, max_pages); 889 } 890 891 mlx5r_umr_set_update_xlt_ctrl_seg(&wqe.ctrl_seg, flags, &sg); 892 mlx5r_umr_set_update_xlt_mkey_seg(dev, &wqe.mkey_seg, mr, page_shift); 893 mlx5r_umr_set_update_xlt_data_seg(&wqe.data_seg, &sg); 894 895 for (pages_mapped = 0; 896 pages_mapped < pages_to_map && !err; 897 pages_mapped += pages_iter, idx += pages_iter) { 898 npages = min_t(int, pages_iter, pages_to_map - pages_mapped); 899 size_to_map = npages * desc_size; 900 dma_sync_single_for_cpu(ddev, sg.addr, sg.length, 901 DMA_TO_DEVICE); 902 /* 903 * npages is the maximum number of pages to map, but we 904 * can't guarantee that all pages are actually mapped. 905 * 906 * For example, if page is p2p of type which is not supported 907 * for mapping, the number of pages mapped will be less than 908 * requested. 909 */ 910 err = mlx5_odp_populate_xlt(xlt, idx, npages, mr, flags); 911 if (err) 912 return err; 913 dma_sync_single_for_device(ddev, sg.addr, sg.length, 914 DMA_TO_DEVICE); 915 sg.length = ALIGN(size_to_map, MLX5_UMR_FLEX_ALIGNMENT); 916 917 if (pages_mapped + pages_iter >= pages_to_map) 918 mlx5r_umr_final_update_xlt(dev, &wqe, mr, &sg, flags); 919 mlx5r_umr_update_offset(&wqe.ctrl_seg, idx * desc_size); 920 err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, true); 921 } 922 sg.length = orig_sg_length; 923 mlx5r_umr_unmap_free_xlt(dev, xlt, &sg); 924 return err; 925 } 926 927 /* 928 * Update only the page-size (log_page_size) field of an existing memory key 929 * using UMR. This is useful when the MR's physical layout stays the same 930 * but the optimal page shift has changed (e.g. dmabuf after pages are 931 * pinned and the HW can switch from 4K to huge-page alignment). 932 */ 933 int mlx5r_umr_update_mr_page_shift(struct mlx5_ib_mr *mr, 934 unsigned int page_shift, 935 bool dd) 936 { 937 struct mlx5_ib_dev *dev = mr_to_mdev(mr); 938 struct mlx5r_umr_wqe wqe = {}; 939 int err; 940 941 /* Build UMR wqe: we touch only PAGE_SIZE, so use the dedicated mask */ 942 wqe.ctrl_seg.mkey_mask = get_umr_update_translation_mask(dev); 943 944 /* MR must be free while page size is modified */ 945 wqe.ctrl_seg.flags = MLX5_UMR_CHECK_FREE | MLX5_UMR_INLINE; 946 947 /* Fill mkey segment with the new page size, keep the rest unchanged */ 948 MLX5_SET(mkc, &wqe.mkey_seg, log_page_size, page_shift); 949 950 if (dd) 951 MLX5_SET(mkc, &wqe.mkey_seg, pd, dev->ddr.pdn); 952 else 953 MLX5_SET(mkc, &wqe.mkey_seg, pd, to_mpd(mr->ibmr.pd)->pdn); 954 955 MLX5_SET64(mkc, &wqe.mkey_seg, start_addr, mr->ibmr.iova); 956 MLX5_SET64(mkc, &wqe.mkey_seg, len, mr->ibmr.length); 957 MLX5_SET(mkc, &wqe.mkey_seg, qpn, 0xffffff); 958 MLX5_SET(mkc, &wqe.mkey_seg, mkey_7_0, 959 mlx5_mkey_variant(mr->mmkey.key)); 960 961 err = mlx5r_umr_post_send_wait(dev, mr->mmkey.key, &wqe, false); 962 if (!err) 963 mr->page_shift = page_shift; 964 965 return err; 966 } 967 968 static inline int 969 _mlx5r_dmabuf_umr_update_pas(struct mlx5_ib_mr *mr, unsigned int flags, 970 size_t start_block, size_t nblocks, bool dd) 971 { 972 if (dd) 973 return mlx5r_umr_update_data_direct_ksm_pas_range(mr, flags, 974 start_block, 975 nblocks); 976 else 977 return mlx5r_umr_update_mr_pas_range(mr, flags, start_block, 978 nblocks); 979 } 980 981 /** 982 * This function makes an mkey non-present by zapping the translation entries of 983 * the mkey by zapping (zeroing out) the first N entries, where N is determined 984 * by the largest page size supported by the device and the MR length. 985 * It then updates the mkey's page size to the largest possible value, ensuring 986 * the MR is completely non-present and safe for further updates. 987 * It is useful to update the page size of a dmabuf MR on a page fault. 988 * 989 * Return: On success, returns the number of entries that were zapped. 990 * On error, returns a negative error code. 991 */ 992 static int _mlx5r_umr_zap_mkey(struct mlx5_ib_mr *mr, 993 unsigned int flags, 994 unsigned int page_shift, 995 bool dd) 996 { 997 unsigned int old_page_shift = mr->page_shift; 998 struct mlx5_ib_dev *dev = mr_to_mdev(mr); 999 unsigned int max_page_shift; 1000 size_t page_shift_nblocks; 1001 unsigned int max_log_size; 1002 int access_mode; 1003 size_t nblocks; 1004 int err; 1005 1006 access_mode = dd ? MLX5_MKC_ACCESS_MODE_KSM : MLX5_MKC_ACCESS_MODE_MTT; 1007 flags |= MLX5_IB_UPD_XLT_KEEP_PGSZ | MLX5_IB_UPD_XLT_ZAP | 1008 MLX5_IB_UPD_XLT_ATOMIC; 1009 max_log_size = get_max_log_entity_size_cap(dev, access_mode); 1010 max_page_shift = order_base_2(mr->ibmr.length); 1011 max_page_shift = min(max(max_page_shift, page_shift), max_log_size); 1012 /* Count blocks in units of max_page_shift, we will zap exactly this 1013 * many to make the whole MR non-present. 1014 * Block size must be aligned to MLX5_UMR_FLEX_ALIGNMENT since it may 1015 * be used as offset into the XLT later on. 1016 */ 1017 nblocks = ib_umem_num_dma_blocks(mr->umem, 1UL << max_page_shift); 1018 if (dd) 1019 nblocks = ALIGN(nblocks, MLX5_UMR_KSM_NUM_ENTRIES_ALIGNMENT); 1020 else 1021 nblocks = ALIGN(nblocks, MLX5_UMR_MTT_NUM_ENTRIES_ALIGNMENT); 1022 page_shift_nblocks = ib_umem_num_dma_blocks(mr->umem, 1023 1UL << page_shift); 1024 /* If the number of blocks at max possible page shift is greater than 1025 * the number of blocks at the new page size, we should just go over the 1026 * whole mkey entries. 1027 */ 1028 if (nblocks >= page_shift_nblocks) 1029 nblocks = 0; 1030 1031 /* Make the first nblocks entries non-present without changing 1032 * page size yet. 1033 */ 1034 if (nblocks) 1035 mr->page_shift = max_page_shift; 1036 err = _mlx5r_dmabuf_umr_update_pas(mr, flags, 0, nblocks, dd); 1037 if (err) { 1038 mr->page_shift = old_page_shift; 1039 return err; 1040 } 1041 1042 /* Change page size to the max page size now that the MR is completely 1043 * non-present. 1044 */ 1045 if (nblocks) { 1046 err = mlx5r_umr_update_mr_page_shift(mr, max_page_shift, dd); 1047 if (err) { 1048 mr->page_shift = old_page_shift; 1049 return err; 1050 } 1051 } 1052 1053 return err ? err : nblocks; 1054 } 1055 1056 /** 1057 * mlx5r_umr_dmabuf_update_pgsz - Safely update DMABUF MR page size and its 1058 * entries accordingly 1059 * @mr: The memory region to update 1060 * @xlt_flags: Translation table update flags 1061 * @page_shift: The new (optimized) page shift to use 1062 * 1063 * This function updates the page size and mkey translation entries for a DMABUF 1064 * MR in a safe, multi-step process to avoid exposing partially updated mappings 1065 * The update is performed in 5 steps: 1066 * 1. Make the first X entries non-present, while X is calculated to be 1067 * minimal according to a large page shift that can be used to cover the 1068 * MR length. 1069 * 2. Update the page size to the large supported page size 1070 * 3. Load the remaining N-X entries according to the (optimized) page_shift 1071 * 4. Update the page size according to the (optimized) page_shift 1072 * 5. Load the first X entries with the correct translations 1073 * 1074 * This ensures that at no point is the MR accessible with a partially updated 1075 * translation table, maintaining correctness and preventing access to stale or 1076 * inconsistent mappings. 1077 * 1078 * Returns 0 on success or a negative error code on failure. 1079 */ 1080 int mlx5r_umr_dmabuf_update_pgsz(struct mlx5_ib_mr *mr, u32 xlt_flags, 1081 unsigned int page_shift) 1082 { 1083 unsigned int old_page_shift = mr->page_shift; 1084 size_t zapped_blocks; 1085 size_t total_blocks; 1086 int err; 1087 1088 zapped_blocks = _mlx5r_umr_zap_mkey(mr, xlt_flags, page_shift, 1089 mr->data_direct); 1090 if (zapped_blocks < 0) 1091 return zapped_blocks; 1092 1093 /* _mlx5r_umr_zap_mkey already enables the mkey */ 1094 xlt_flags &= ~MLX5_IB_UPD_XLT_ENABLE; 1095 mr->page_shift = page_shift; 1096 total_blocks = ib_umem_num_dma_blocks(mr->umem, 1UL << mr->page_shift); 1097 if (zapped_blocks && zapped_blocks < total_blocks) { 1098 /* Update PAS according to the new page size but don't update 1099 * the page size in the mkey yet. 1100 */ 1101 err = _mlx5r_dmabuf_umr_update_pas( 1102 mr, 1103 xlt_flags | MLX5_IB_UPD_XLT_KEEP_PGSZ, 1104 zapped_blocks, 1105 total_blocks - zapped_blocks, 1106 mr->data_direct); 1107 if (err) 1108 goto err; 1109 } 1110 1111 err = mlx5r_umr_update_mr_page_shift(mr, mr->page_shift, 1112 mr->data_direct); 1113 if (err) 1114 goto err; 1115 err = _mlx5r_dmabuf_umr_update_pas(mr, xlt_flags, 0, zapped_blocks, 1116 mr->data_direct); 1117 if (err) 1118 goto err; 1119 1120 return 0; 1121 err: 1122 mr->page_shift = old_page_shift; 1123 return err; 1124 } 1125