drivers/misc/habanalabs/memory.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /*
   4  * Copyright 2016-2019 HabanaLabs, Ltd.
   5  * All Rights Reserved.
   6  */
   7
   8 #include <uapi/misc/habanalabs.h>
   9 #include "habanalabs.h"
  10 #include "include/hw_ip/mmu/mmu_general.h"
  11
  12 #include <linux/uaccess.h>
  13 #include <linux/slab.h>
  14 #include <linux/genalloc.h>
  15
  16 #define PGS_IN_2MB_PAGE (PAGE_SIZE_2MB >> PAGE_SHIFT)
  17 #define HL_MMU_DEBUG    0
  18
  19 /*
  20  * The va ranges in context object contain a list with the available chunks of
  21  * device virtual memory.
  22  * There is one range for host allocations and one for DRAM allocations.
  23  *
  24  * On initialization each range contains one chunk of all of its available
  25  * virtual range which is a half of the total device virtual range.
  26  *
  27  * On each mapping of physical pages, a suitable virtual range chunk (with a
  28  * minimum size) is selected from the list. If the chunk size equals the
  29  * requested size, the chunk is returned. Otherwise, the chunk is split into
  30  * two chunks - one to return as result and a remainder to stay in the list.
  31  *
  32  * On each Unmapping of a virtual address, the relevant virtual chunk is
  33  * returned to the list. The chunk is added to the list and if its edges match
  34  * the edges of the adjacent chunks (means a contiguous chunk can be created),
  35  * the chunks are merged.
  36  *
  37  * On finish, the list is checked to have only one chunk of all the relevant
  38  * virtual range (which is a half of the device total virtual range).
  39  * If not (means not all mappings were unmapped), a warning is printed.
  40  */
  41
  42 /*
  43  * alloc_device_memory - allocate device memory
  44  *
  45  * @ctx                 : current context
  46  * @args                : host parameters containing the requested size
  47  * @ret_handle          : result handle
  48  *
  49  * This function does the following:
  50  * - Allocate the requested size rounded up to 2MB pages
  51  * - Return unique handle
  52  */
  53 static int alloc_device_memory(struct hl_ctx *ctx, struct hl_mem_in *args,
  54                                 u32 *ret_handle)
  55 {
  56         struct hl_device *hdev = ctx->hdev;
  57         struct hl_vm *vm = &hdev->vm;
  58         struct hl_vm_phys_pg_pack *phys_pg_pack;
  59         u64 paddr = 0;
  60         u32 total_size, num_pgs, num_curr_pgs, page_size, page_shift;
  61         int handle, rc, i;
  62         bool contiguous;
  63
  64         num_curr_pgs = 0;
  65         page_size = hdev->asic_prop.dram_page_size;
  66         page_shift = __ffs(page_size);
  67         num_pgs = (args->alloc.mem_size + (page_size - 1)) >> page_shift;
  68         total_size = num_pgs << page_shift;
  69
  70         contiguous = args->flags & HL_MEM_CONTIGUOUS;
  71
  72         if (contiguous) {
  73                 paddr = (u64) gen_pool_alloc(vm->dram_pg_pool, total_size);
  74                 if (!paddr) {
  75                         dev_err(hdev->dev,
  76                                 "failed to allocate %u huge contiguous pages\n",
  77                                 num_pgs);
  78                         return -ENOMEM;
  79                 }
  80         }
  81
  82         phys_pg_pack = kzalloc(sizeof(*phys_pg_pack), GFP_KERNEL);
  83         if (!phys_pg_pack) {
  84                 rc = -ENOMEM;
  85                 goto pages_pack_err;
  86         }
  87
  88         phys_pg_pack->vm_type = VM_TYPE_PHYS_PACK;
  89         phys_pg_pack->asid = ctx->asid;
  90         phys_pg_pack->npages = num_pgs;
  91         phys_pg_pack->page_size = page_size;
  92         phys_pg_pack->total_size = total_size;
  93         phys_pg_pack->flags = args->flags;
  94         phys_pg_pack->contiguous = contiguous;
  95
  96         phys_pg_pack->pages = kcalloc(num_pgs, sizeof(u64), GFP_KERNEL);
  97         if (!phys_pg_pack->pages) {
  98                 rc = -ENOMEM;
  99                 goto pages_arr_err;
 100         }
 101
 102         if (phys_pg_pack->contiguous) {
 103                 for (i = 0 ; i < num_pgs ; i++)
 104                         phys_pg_pack->pages[i] = paddr + i * page_size;
 105         } else {
 106                 for (i = 0 ; i < num_pgs ; i++) {
 107                         phys_pg_pack->pages[i] = (u64) gen_pool_alloc(
 108                                                         vm->dram_pg_pool,
 109                                                         page_size);
 110                         if (!phys_pg_pack->pages[i]) {
 111                                 dev_err(hdev->dev,
 112                                         "ioctl failed to allocate page\n");
 113                                 rc = -ENOMEM;
 114                                 goto page_err;
 115                         }
 116
 117                         num_curr_pgs++;
 118                 }
 119         }
 120
 121         spin_lock(&vm->idr_lock);
 122         handle = idr_alloc(&vm->phys_pg_pack_handles, phys_pg_pack, 1, 0,
 123                                 GFP_ATOMIC);
 124         spin_unlock(&vm->idr_lock);
 125
 126         if (handle < 0) {
 127                 dev_err(hdev->dev, "Failed to get handle for page\n");
 128                 rc = -EFAULT;
 129                 goto idr_err;
 130         }
 131
 132         for (i = 0 ; i < num_pgs ; i++)
 133                 kref_get(&vm->dram_pg_pool_refcount);
 134
 135         phys_pg_pack->handle = handle;
 136
 137         atomic64_add(phys_pg_pack->total_size, &ctx->dram_phys_mem);
 138         atomic64_add(phys_pg_pack->total_size, &hdev->dram_used_mem);
 139
 140         *ret_handle = handle;
 141
 142         return 0;
 143
 144 idr_err:
 145 page_err:
 146         if (!phys_pg_pack->contiguous)
 147                 for (i = 0 ; i < num_curr_pgs ; i++)
 148                         gen_pool_free(vm->dram_pg_pool, phys_pg_pack->pages[i],
 149                                         page_size);
 150
 151         kfree(phys_pg_pack->pages);
 152 pages_arr_err:
 153         kfree(phys_pg_pack);
 154 pages_pack_err:
 155         if (contiguous)
 156                 gen_pool_free(vm->dram_pg_pool, paddr, total_size);
 157
 158         return rc;
 159 }
 160
 161 /*
 162  * get_userptr_from_host_va - initialize userptr structure from given host
 163  *                            virtual address
 164  *
 165  * @hdev                : habanalabs device structure
 166  * @args                : parameters containing the virtual address and size
 167  * @p_userptr           : pointer to result userptr structure
 168  *
 169  * This function does the following:
 170  * - Allocate userptr structure
 171  * - Pin the given host memory using the userptr structure
 172  * - Perform DMA mapping to have the DMA addresses of the pages
 173  */
 174 static int get_userptr_from_host_va(struct hl_device *hdev,
 175                 struct hl_mem_in *args, struct hl_userptr **p_userptr)
 176 {
 177         struct hl_userptr *userptr;
 178         int rc;
 179
 180         userptr = kzalloc(sizeof(*userptr), GFP_KERNEL);
 181         if (!userptr) {
 182                 rc = -ENOMEM;
 183                 goto userptr_err;
 184         }
 185
 186         rc = hl_pin_host_memory(hdev, args->map_host.host_virt_addr,
 187                         args->map_host.mem_size, userptr);
 188         if (rc) {
 189                 dev_err(hdev->dev, "Failed to pin host memory\n");
 190                 goto pin_err;
 191         }
 192
 193         rc = hdev->asic_funcs->asic_dma_map_sg(hdev, userptr->sgt->sgl,
 194                                         userptr->sgt->nents, DMA_BIDIRECTIONAL);
 195         if (rc) {
 196                 dev_err(hdev->dev, "failed to map sgt with DMA region\n");
 197                 goto dma_map_err;
 198         }
 199
 200         userptr->dma_mapped = true;
 201         userptr->dir = DMA_BIDIRECTIONAL;
 202         userptr->vm_type = VM_TYPE_USERPTR;
 203
 204         *p_userptr = userptr;
 205
 206         return 0;
 207
 208 dma_map_err:
 209         hl_unpin_host_memory(hdev, userptr);
 210 pin_err:
 211         kfree(userptr);
 212 userptr_err:
 213
 214         return rc;
 215 }
 216
 217 /*
 218  * free_userptr - free userptr structure
 219  *
 220  * @hdev                : habanalabs device structure
 221  * @userptr             : userptr to free
 222  *
 223  * This function does the following:
 224  * - Unpins the physical pages
 225  * - Frees the userptr structure
 226  */
 227 static void free_userptr(struct hl_device *hdev, struct hl_userptr *userptr)
 228 {
 229         hl_unpin_host_memory(hdev, userptr);
 230         kfree(userptr);
 231 }
 232
 233 /*
 234  * dram_pg_pool_do_release - free DRAM pages pool
 235  *
 236  * @ref                 : pointer to reference object
 237  *
 238  * This function does the following:
 239  * - Frees the idr structure of physical pages handles
 240  * - Frees the generic pool of DRAM physical pages
 241  */
 242 static void dram_pg_pool_do_release(struct kref *ref)
 243 {
 244         struct hl_vm *vm = container_of(ref, struct hl_vm,
 245                         dram_pg_pool_refcount);
 246
 247         /*
 248          * free the idr here as only here we know for sure that there are no
 249          * allocated physical pages and hence there are no handles in use
 250          */
 251         idr_destroy(&vm->phys_pg_pack_handles);
 252         gen_pool_destroy(vm->dram_pg_pool);
 253 }
 254
 255 /*
 256  * free_phys_pg_pack   - free physical page pack
 257  *
 258  * @hdev               : habanalabs device structure
 259  * @phys_pg_pack       : physical page pack to free
 260  *
 261  * This function does the following:
 262  * - For DRAM memory only, iterate over the pack and free each physical block
 263  *   structure by returning it to the general pool
 264  * - Free the hl_vm_phys_pg_pack structure
 265  */
 266 static void free_phys_pg_pack(struct hl_device *hdev,
 267                 struct hl_vm_phys_pg_pack *phys_pg_pack)
 268 {
 269         struct hl_vm *vm = &hdev->vm;
 270         int i;
 271
 272         if (!phys_pg_pack->created_from_userptr) {
 273                 if (phys_pg_pack->contiguous) {
 274                         gen_pool_free(vm->dram_pg_pool, phys_pg_pack->pages[0],
 275                                         phys_pg_pack->total_size);
 276
 277                         for (i = 0; i < phys_pg_pack->npages ; i++)
 278                                 kref_put(&vm->dram_pg_pool_refcount,
 279                                         dram_pg_pool_do_release);
 280                 } else {
 281                         for (i = 0 ; i < phys_pg_pack->npages ; i++) {
 282                                 gen_pool_free(vm->dram_pg_pool,
 283                                                 phys_pg_pack->pages[i],
 284                                                 phys_pg_pack->page_size);
 285                                 kref_put(&vm->dram_pg_pool_refcount,
 286                                         dram_pg_pool_do_release);
 287                         }
 288                 }
 289         }
 290
 291         kfree(phys_pg_pack->pages);
 292         kfree(phys_pg_pack);
 293 }
 294
 295 /*
 296  * free_device_memory - free device memory
 297  *
 298  * @ctx                  : current context
 299  * @handle              : handle of the memory chunk to free
 300  *
 301  * This function does the following:
 302  * - Free the device memory related to the given handle
 303  */
 304 static int free_device_memory(struct hl_ctx *ctx, u32 handle)
 305 {
 306         struct hl_device *hdev = ctx->hdev;
 307         struct hl_vm *vm = &hdev->vm;
 308         struct hl_vm_phys_pg_pack *phys_pg_pack;
 309
 310         spin_lock(&vm->idr_lock);
 311         phys_pg_pack = idr_find(&vm->phys_pg_pack_handles, handle);
 312         if (phys_pg_pack) {
 313                 if (atomic_read(&phys_pg_pack->mapping_cnt) > 0) {
 314                         dev_err(hdev->dev, "handle %u is mapped, cannot free\n",
 315                                 handle);
 316                         spin_unlock(&vm->idr_lock);
 317                         return -EINVAL;
 318                 }
 319
 320                 /*
 321                  * must remove from idr before the freeing of the physical
 322                  * pages as the refcount of the pool is also the trigger of the
 323                  * idr destroy
 324                  */
 325                 idr_remove(&vm->phys_pg_pack_handles, handle);
 326                 spin_unlock(&vm->idr_lock);
 327
 328                 atomic64_sub(phys_pg_pack->total_size, &ctx->dram_phys_mem);
 329                 atomic64_sub(phys_pg_pack->total_size, &hdev->dram_used_mem);
 330
 331                 free_phys_pg_pack(hdev, phys_pg_pack);
 332         } else {
 333                 spin_unlock(&vm->idr_lock);
 334                 dev_err(hdev->dev,
 335                         "free device memory failed, no match for handle %u\n",
 336                         handle);
 337                 return -EINVAL;
 338         }
 339
 340         return 0;
 341 }
 342
 343 /*
 344  * clear_va_list_locked - free virtual addresses list
 345  *
 346  * @hdev                : habanalabs device structure
 347  * @va_list             : list of virtual addresses to free
 348  *
 349  * This function does the following:
 350  * - Iterate over the list and free each virtual addresses block
 351  *
 352  * This function should be called only when va_list lock is taken
 353  */
 354 static void clear_va_list_locked(struct hl_device *hdev,
 355                 struct list_head *va_list)
 356 {
 357         struct hl_vm_va_block *va_block, *tmp;
 358
 359         list_for_each_entry_safe(va_block, tmp, va_list, node) {
 360                 list_del(&va_block->node);
 361                 kfree(va_block);
 362         }
 363 }
 364
 365 /*
 366  * print_va_list_locked    - print virtual addresses list
 367  *
 368  * @hdev                : habanalabs device structure
 369  * @va_list             : list of virtual addresses to print
 370  *
 371  * This function does the following:
 372  * - Iterate over the list and print each virtual addresses block
 373  *
 374  * This function should be called only when va_list lock is taken
 375  */
 376 static void print_va_list_locked(struct hl_device *hdev,
 377                 struct list_head *va_list)
 378 {
 379 #if HL_MMU_DEBUG
 380         struct hl_vm_va_block *va_block;
 381
 382         dev_dbg(hdev->dev, "print va list:\n");
 383
 384         list_for_each_entry(va_block, va_list, node)
 385                 dev_dbg(hdev->dev,
 386                         "va block, start: 0x%llx, end: 0x%llx, size: %llu\n",
 387                         va_block->start, va_block->end, va_block->size);
 388 #endif
 389 }
 390
 391 /*
 392  * merge_va_blocks_locked - merge a virtual block if possible
 393  *
 394  * @hdev                : pointer to the habanalabs device structure
 395  * @va_list             : pointer to the virtual addresses block list
 396  * @va_block            : virtual block to merge with adjacent blocks
 397  *
 398  * This function does the following:
 399  * - Merge the given blocks with the adjacent blocks if their virtual ranges
 400  *   create a contiguous virtual range
 401  *
 402  * This Function should be called only when va_list lock is taken
 403  */
 404 static void merge_va_blocks_locked(struct hl_device *hdev,
 405                 struct list_head *va_list, struct hl_vm_va_block *va_block)
 406 {
 407         struct hl_vm_va_block *prev, *next;
 408
 409         prev = list_prev_entry(va_block, node);
 410         if (&prev->node != va_list && prev->end + 1 == va_block->start) {
 411                 prev->end = va_block->end;
 412                 prev->size = prev->end - prev->start;
 413                 list_del(&va_block->node);
 414                 kfree(va_block);
 415                 va_block = prev;
 416         }
 417
 418         next = list_next_entry(va_block, node);
 419         if (&next->node != va_list && va_block->end + 1 == next->start) {
 420                 next->start = va_block->start;
 421                 next->size = next->end - next->start;
 422                 list_del(&va_block->node);
 423                 kfree(va_block);
 424         }
 425 }
 426
 427 /*
 428  * add_va_block_locked - add a virtual block to the virtual addresses list
 429  *
 430  * @hdev                : pointer to the habanalabs device structure
 431  * @va_list             : pointer to the virtual addresses block list
 432  * @start               : start virtual address
 433  * @end                 : end virtual address
 434  *
 435  * This function does the following:
 436  * - Add the given block to the virtual blocks list and merge with other
 437  * blocks if a contiguous virtual block can be created
 438  *
 439  * This Function should be called only when va_list lock is taken
 440  */
 441 static int add_va_block_locked(struct hl_device *hdev,
 442                 struct list_head *va_list, u64 start, u64 end)
 443 {
 444         struct hl_vm_va_block *va_block, *res = NULL;
 445         u64 size = end - start;
 446
 447         print_va_list_locked(hdev, va_list);
 448
 449         list_for_each_entry(va_block, va_list, node) {
 450                 /* TODO: remove upon matureness */
 451                 if (hl_mem_area_crosses_range(start, size, va_block->start,
 452                                 va_block->end)) {
 453                         dev_err(hdev->dev,
 454                                 "block crossing ranges at start 0x%llx, end 0x%llx\n",
 455                                 va_block->start, va_block->end);
 456                         return -EINVAL;
 457                 }
 458
 459                 if (va_block->end < start)
 460                         res = va_block;
 461         }
 462
 463         va_block = kmalloc(sizeof(*va_block), GFP_KERNEL);
 464         if (!va_block)
 465                 return -ENOMEM;
 466
 467         va_block->start = start;
 468         va_block->end = end;
 469         va_block->size = size;
 470
 471         if (!res)
 472                 list_add(&va_block->node, va_list);
 473         else
 474                 list_add(&va_block->node, &res->node);
 475
 476         merge_va_blocks_locked(hdev, va_list, va_block);
 477
 478         print_va_list_locked(hdev, va_list);
 479
 480         return 0;
 481 }
 482
 483 /*
 484  * add_va_block - wrapper for add_va_block_locked
 485  *
 486  * @hdev                : pointer to the habanalabs device structure
 487  * @va_list             : pointer to the virtual addresses block list
 488  * @start               : start virtual address
 489  * @end                 : end virtual address
 490  *
 491  * This function does the following:
 492  * - Takes the list lock and calls add_va_block_locked
 493  */
 494 static inline int add_va_block(struct hl_device *hdev,
 495                 struct hl_va_range *va_range, u64 start, u64 end)
 496 {
 497         int rc;
 498
 499         mutex_lock(&va_range->lock);
 500         rc = add_va_block_locked(hdev, &va_range->list, start, end);
 501         mutex_unlock(&va_range->lock);
 502
 503         return rc;
 504 }
 505
 506 /*
 507  * get_va_block - get a virtual block with the requested size
 508  *
 509  * @hdev            : pointer to the habanalabs device structure
 510  * @va_range        : pointer to the virtual addresses range
 511  * @size            : requested block size
 512  * @hint_addr       : hint for request address by the user
 513  * @is_userptr      : is host or DRAM memory
 514  *
 515  * This function does the following:
 516  * - Iterate on the virtual block list to find a suitable virtual block for the
 517  *   requested size
 518  * - Reserve the requested block and update the list
 519  * - Return the start address of the virtual block
 520  */
 521 static u64 get_va_block(struct hl_device *hdev,
 522                 struct hl_va_range *va_range, u32 size, u64 hint_addr,
 523                 bool is_userptr)
 524 {
 525         struct hl_vm_va_block *va_block, *new_va_block = NULL;
 526         u64 valid_start, valid_size, prev_start, prev_end, page_mask,
 527                 res_valid_start = 0, res_valid_size = 0;
 528         u32 page_size;
 529         bool add_prev = false;
 530
 531         if (is_userptr) {
 532                 /*
 533                  * We cannot know if the user allocated memory with huge pages
 534                  * or not, hence we continue with the biggest possible
 535                  * granularity.
 536                  */
 537                 page_size = PAGE_SIZE_2MB;
 538                 page_mask = PAGE_MASK_2MB;
 539         } else {
 540                 page_size = hdev->asic_prop.dram_page_size;
 541                 page_mask = ~((u64)page_size - 1);
 542         }
 543
 544         mutex_lock(&va_range->lock);
 545
 546         print_va_list_locked(hdev, &va_range->list);
 547
 548         list_for_each_entry(va_block, &va_range->list, node) {
 549                 /* calc the first possible aligned addr */
 550                 valid_start = va_block->start;
 551
 552
 553                 if (valid_start & (page_size - 1)) {
 554                         valid_start &= page_mask;
 555                         valid_start += page_size;
 556                         if (valid_start > va_block->end)
 557                                 continue;
 558                 }
 559
 560                 valid_size = va_block->end - valid_start;
 561
 562                 if (valid_size >= size &&
 563                         (!new_va_block || valid_size < res_valid_size)) {
 564
 565                         new_va_block = va_block;
 566                         res_valid_start = valid_start;
 567                         res_valid_size = valid_size;
 568                 }
 569
 570                 if (hint_addr && hint_addr >= valid_start &&
 571                                 ((hint_addr + size) <= va_block->end)) {
 572                         new_va_block = va_block;
 573                         res_valid_start = hint_addr;
 574                         res_valid_size = valid_size;
 575                         break;
 576                 }
 577         }
 578
 579         if (!new_va_block) {
 580                 dev_err(hdev->dev, "no available va block for size %u\n", size);
 581                 goto out;
 582         }
 583
 584         if (res_valid_start > new_va_block->start) {
 585                 prev_start = new_va_block->start;
 586                 prev_end = res_valid_start - 1;
 587
 588                 new_va_block->start = res_valid_start;
 589                 new_va_block->size = res_valid_size;
 590
 591                 add_prev = true;
 592         }
 593
 594         if (new_va_block->size > size) {
 595                 new_va_block->start += size;
 596                 new_va_block->size = new_va_block->end - new_va_block->start;
 597         } else {
 598                 list_del(&new_va_block->node);
 599                 kfree(new_va_block);
 600         }
 601
 602         if (add_prev)
 603                 add_va_block_locked(hdev, &va_range->list, prev_start,
 604                                 prev_end);
 605
 606         print_va_list_locked(hdev, &va_range->list);
 607 out:
 608         mutex_unlock(&va_range->lock);
 609
 610         return res_valid_start;
 611 }
 612
 613 /*
 614  * get_sg_info - get number of pages and the DMA address from SG list
 615  *
 616  * @sg                 : the SG list
 617  * @dma_addr           : pointer to DMA address to return
 618  *
 619  * Calculate the number of consecutive pages described by the SG list. Take the
 620  * offset of the address in the first page, add to it the length and round it up
 621  * to the number of needed pages.
 622  */
 623 static u32 get_sg_info(struct scatterlist *sg, dma_addr_t *dma_addr)
 624 {
 625         *dma_addr = sg_dma_address(sg);
 626
 627         return ((((*dma_addr) & (PAGE_SIZE - 1)) + sg_dma_len(sg)) +
 628                         (PAGE_SIZE - 1)) >> PAGE_SHIFT;
 629 }
 630
 631 /*
 632  * init_phys_pg_pack_from_userptr - initialize physical page pack from host
 633  *                                   memory
 634  *
 635  * @ctx                : current context
 636  * @userptr            : userptr to initialize from
 637  * @pphys_pg_pack      : res pointer
 638  *
 639  * This function does the following:
 640  * - Pin the physical pages related to the given virtual block
 641  * - Create a physical page pack from the physical pages related to the given
 642  *   virtual block
 643  */
 644 static int init_phys_pg_pack_from_userptr(struct hl_ctx *ctx,
 645                 struct hl_userptr *userptr,
 646                 struct hl_vm_phys_pg_pack **pphys_pg_pack)
 647 {
 648         struct hl_vm_phys_pg_pack *phys_pg_pack;
 649         struct scatterlist *sg;
 650         dma_addr_t dma_addr;
 651         u64 page_mask;
 652         u32 npages, total_npages, page_size = PAGE_SIZE;
 653         bool first = true, is_huge_page_opt = true;
 654         int rc, i, j;
 655
 656         phys_pg_pack = kzalloc(sizeof(*phys_pg_pack), GFP_KERNEL);
 657         if (!phys_pg_pack)
 658                 return -ENOMEM;
 659
 660         phys_pg_pack->vm_type = userptr->vm_type;
 661         phys_pg_pack->created_from_userptr = true;
 662         phys_pg_pack->asid = ctx->asid;
 663         atomic_set(&phys_pg_pack->mapping_cnt, 1);
 664
 665         /* Only if all dma_addrs are aligned to 2MB and their
 666          * sizes is at least 2MB, we can use huge page mapping.
 667          * We limit the 2MB optimization to this condition,
 668          * since later on we acquire the related VA range as one
 669          * consecutive block.
 670          */
 671         total_npages = 0;
 672         for_each_sg(userptr->sgt->sgl, sg, userptr->sgt->nents, i) {
 673                 npages = get_sg_info(sg, &dma_addr);
 674
 675                 total_npages += npages;
 676
 677                 if (first) {
 678                         first = false;
 679                         dma_addr &= PAGE_MASK_2MB;
 680                 }
 681
 682                 if ((npages % PGS_IN_2MB_PAGE) ||
 683                                         (dma_addr & (PAGE_SIZE_2MB - 1)))
 684                         is_huge_page_opt = false;
 685         }
 686
 687         if (is_huge_page_opt) {
 688                 page_size = PAGE_SIZE_2MB;
 689                 total_npages /= PGS_IN_2MB_PAGE;
 690         }
 691
 692         page_mask = ~(((u64) page_size) - 1);
 693
 694         phys_pg_pack->pages = kcalloc(total_npages, sizeof(u64), GFP_KERNEL);
 695         if (!phys_pg_pack->pages) {
 696                 rc = -ENOMEM;
 697                 goto page_pack_arr_mem_err;
 698         }
 699
 700         phys_pg_pack->npages = total_npages;
 701         phys_pg_pack->page_size = page_size;
 702         phys_pg_pack->total_size = total_npages * page_size;
 703
 704         j = 0;
 705         first = true;
 706         for_each_sg(userptr->sgt->sgl, sg, userptr->sgt->nents, i) {
 707                 npages = get_sg_info(sg, &dma_addr);
 708
 709                 /* align down to physical page size and save the offset */
 710                 if (first) {
 711                         first = false;
 712                         phys_pg_pack->offset = dma_addr & (page_size - 1);
 713                         dma_addr &= page_mask;
 714                 }
 715
 716                 while (npages) {
 717                         phys_pg_pack->pages[j++] = dma_addr;
 718                         dma_addr += page_size;
 719
 720                         if (is_huge_page_opt)
 721                                 npages -= PGS_IN_2MB_PAGE;
 722                         else
 723                                 npages--;
 724                 }
 725         }
 726
 727         *pphys_pg_pack = phys_pg_pack;
 728
 729         return 0;
 730
 731 page_pack_arr_mem_err:
 732         kfree(phys_pg_pack);
 733
 734         return rc;
 735 }
 736
 737 /*
 738  * map_phys_page_pack - maps the physical page pack
 739  *
 740  * @ctx                : current context
 741  * @vaddr              : start address of the virtual area to map from
 742  * @phys_pg_pack       : the pack of physical pages to map to
 743  *
 744  * This function does the following:
 745  * - Maps each chunk of virtual memory to matching physical chunk
 746  * - Stores number of successful mappings in the given argument
 747  * - Returns 0 on success, error code otherwise.
 748  */
 749 static int map_phys_page_pack(struct hl_ctx *ctx, u64 vaddr,
 750                 struct hl_vm_phys_pg_pack *phys_pg_pack)
 751 {
 752         struct hl_device *hdev = ctx->hdev;
 753         u64 next_vaddr = vaddr, paddr;
 754         u32 page_size = phys_pg_pack->page_size;
 755         int i, rc = 0, mapped_pg_cnt = 0;
 756
 757         for (i = 0 ; i < phys_pg_pack->npages ; i++) {
 758                 paddr = phys_pg_pack->pages[i];
 759
 760                 /* For accessing the host we need to turn on bit 39 */
 761                 if (phys_pg_pack->created_from_userptr)
 762                         paddr += hdev->asic_prop.host_phys_base_address;
 763
 764                 rc = hl_mmu_map(ctx, next_vaddr, paddr, page_size);
 765                 if (rc) {
 766                         dev_err(hdev->dev,
 767                                 "map failed for handle %u, npages: %d, mapped: %d",
 768                                 phys_pg_pack->handle, phys_pg_pack->npages,
 769                                 mapped_pg_cnt);
 770                         goto err;
 771                 }
 772
 773                 mapped_pg_cnt++;
 774                 next_vaddr += page_size;
 775         }
 776
 777         return 0;
 778
 779 err:
 780         next_vaddr = vaddr;
 781         for (i = 0 ; i < mapped_pg_cnt ; i++) {
 782                 if (hl_mmu_unmap(ctx, next_vaddr, page_size))
 783                         dev_warn_ratelimited(hdev->dev,
 784                                 "failed to unmap handle %u, va: 0x%llx, pa: 0x%llx, page size: %u\n",
 785                                         phys_pg_pack->handle, next_vaddr,
 786                                         phys_pg_pack->pages[i], page_size);
 787
 788                 next_vaddr += page_size;
 789         }
 790
 791         return rc;
 792 }
 793
 794 static int get_paddr_from_handle(struct hl_ctx *ctx, struct hl_mem_in *args,
 795                                 u64 *paddr)
 796 {
 797         struct hl_device *hdev = ctx->hdev;
 798         struct hl_vm *vm = &hdev->vm;
 799         struct hl_vm_phys_pg_pack *phys_pg_pack;
 800         u32 handle;
 801
 802         handle = lower_32_bits(args->map_device.handle);
 803         spin_lock(&vm->idr_lock);
 804         phys_pg_pack = idr_find(&vm->phys_pg_pack_handles, handle);
 805         if (!phys_pg_pack) {
 806                 spin_unlock(&vm->idr_lock);
 807                 dev_err(hdev->dev, "no match for handle %u\n", handle);
 808                 return -EINVAL;
 809         }
 810
 811         *paddr = phys_pg_pack->pages[0];
 812
 813         spin_unlock(&vm->idr_lock);
 814
 815         return 0;
 816 }
 817
 818 /*
 819  * map_device_va - map the given memory
 820  *
 821  * @ctx          : current context
 822  * @args         : host parameters with handle/host virtual address
 823  * @device_addr  : pointer to result device virtual address
 824  *
 825  * This function does the following:
 826  * - If given a physical device memory handle, map to a device virtual block
 827  *   and return the start address of this block
 828  * - If given a host virtual address and size, find the related physical pages,
 829  *   map a device virtual block to this pages and return the start address of
 830  *   this block
 831  */
 832 static int map_device_va(struct hl_ctx *ctx, struct hl_mem_in *args,
 833                 u64 *device_addr)
 834 {
 835         struct hl_device *hdev = ctx->hdev;
 836         struct hl_vm *vm = &hdev->vm;
 837         struct hl_vm_phys_pg_pack *phys_pg_pack;
 838         struct hl_userptr *userptr = NULL;
 839         struct hl_vm_hash_node *hnode;
 840         enum vm_type_t *vm_type;
 841         u64 ret_vaddr, hint_addr;
 842         u32 handle = 0;
 843         int rc;
 844         bool is_userptr = args->flags & HL_MEM_USERPTR;
 845
 846         /* Assume failure */
 847         *device_addr = 0;
 848
 849         if (is_userptr) {
 850                 rc = get_userptr_from_host_va(hdev, args, &userptr);
 851                 if (rc) {
 852                         dev_err(hdev->dev, "failed to get userptr from va\n");
 853                         return rc;
 854                 }
 855
 856                 rc = init_phys_pg_pack_from_userptr(ctx, userptr,
 857                                 &phys_pg_pack);
 858                 if (rc) {
 859                         dev_err(hdev->dev,
 860                                 "unable to init page pack for vaddr 0x%llx\n",
 861                                 args->map_host.host_virt_addr);
 862                         goto init_page_pack_err;
 863                 }
 864
 865                 vm_type = (enum vm_type_t *) userptr;
 866                 hint_addr = args->map_host.hint_addr;
 867         } else {
 868                 handle = lower_32_bits(args->map_device.handle);
 869
 870                 spin_lock(&vm->idr_lock);
 871                 phys_pg_pack = idr_find(&vm->phys_pg_pack_handles, handle);
 872                 if (!phys_pg_pack) {
 873                         spin_unlock(&vm->idr_lock);
 874                         dev_err(hdev->dev,
 875                                 "no match for handle %u\n", handle);
 876                         return -EINVAL;
 877                 }
 878
 879                 /* increment now to avoid freeing device memory while mapping */
 880                 atomic_inc(&phys_pg_pack->mapping_cnt);
 881
 882                 spin_unlock(&vm->idr_lock);
 883
 884                 vm_type = (enum vm_type_t *) phys_pg_pack;
 885
 886                 hint_addr = args->map_device.hint_addr;
 887         }
 888
 889         /*
 890          * relevant for mapping device physical memory only, as host memory is
 891          * implicitly shared
 892          */
 893         if (!is_userptr && !(phys_pg_pack->flags & HL_MEM_SHARED) &&
 894                         phys_pg_pack->asid != ctx->asid) {
 895                 dev_err(hdev->dev,
 896                         "Failed to map memory, handle %u is not shared\n",
 897                         handle);
 898                 rc = -EPERM;
 899                 goto shared_err;
 900         }
 901
 902         hnode = kzalloc(sizeof(*hnode), GFP_KERNEL);
 903         if (!hnode) {
 904                 rc = -ENOMEM;
 905                 goto hnode_err;
 906         }
 907
 908         ret_vaddr = get_va_block(hdev,
 909                         is_userptr ? &ctx->host_va_range : &ctx->dram_va_range,
 910                         phys_pg_pack->total_size, hint_addr, is_userptr);
 911         if (!ret_vaddr) {
 912                 dev_err(hdev->dev, "no available va block for handle %u\n",
 913                                 handle);
 914                 rc = -ENOMEM;
 915                 goto va_block_err;
 916         }
 917
 918         mutex_lock(&ctx->mmu_lock);
 919
 920         rc = map_phys_page_pack(ctx, ret_vaddr, phys_pg_pack);
 921         if (rc) {
 922                 mutex_unlock(&ctx->mmu_lock);
 923                 dev_err(hdev->dev, "mapping page pack failed for handle %u\n",
 924                                 handle);
 925                 goto map_err;
 926         }
 927
 928         hdev->asic_funcs->mmu_invalidate_cache(hdev, false);
 929
 930         mutex_unlock(&ctx->mmu_lock);
 931
 932         ret_vaddr += phys_pg_pack->offset;
 933
 934         hnode->ptr = vm_type;
 935         hnode->vaddr = ret_vaddr;
 936
 937         mutex_lock(&ctx->mem_hash_lock);
 938         hash_add(ctx->mem_hash, &hnode->node, ret_vaddr);
 939         mutex_unlock(&ctx->mem_hash_lock);
 940
 941         *device_addr = ret_vaddr;
 942
 943         if (is_userptr)
 944                 free_phys_pg_pack(hdev, phys_pg_pack);
 945
 946         return 0;
 947
 948 map_err:
 949         if (add_va_block(hdev,
 950                         is_userptr ? &ctx->host_va_range : &ctx->dram_va_range,
 951                         ret_vaddr,
 952                         ret_vaddr + phys_pg_pack->total_size - 1))
 953                 dev_warn(hdev->dev,
 954                         "release va block failed for handle 0x%x, vaddr: 0x%llx\n",
 955                                 handle, ret_vaddr);
 956
 957 va_block_err:
 958         kfree(hnode);
 959 hnode_err:
 960 shared_err:
 961         atomic_dec(&phys_pg_pack->mapping_cnt);
 962         if (is_userptr)
 963                 free_phys_pg_pack(hdev, phys_pg_pack);
 964 init_page_pack_err:
 965         if (is_userptr)
 966                 free_userptr(hdev, userptr);
 967
 968         return rc;
 969 }
 970
 971 /*
 972  * unmap_device_va      - unmap the given device virtual address
 973  *
 974  * @ctx                 : current context
 975  * @vaddr               : device virtual address to unmap
 976  *
 977  * This function does the following:
 978  * - Unmap the physical pages related to the given virtual address
 979  * - return the device virtual block to the virtual block list
 980  */
 981 static int unmap_device_va(struct hl_ctx *ctx, u64 vaddr)
 982 {
 983         struct hl_device *hdev = ctx->hdev;
 984         struct hl_vm_phys_pg_pack *phys_pg_pack = NULL;
 985         struct hl_vm_hash_node *hnode = NULL;
 986         struct hl_userptr *userptr = NULL;
 987         enum vm_type_t *vm_type;
 988         u64 next_vaddr;
 989         u32 page_size;
 990         bool is_userptr;
 991         int i, rc;
 992
 993         /* protect from double entrance */
 994         mutex_lock(&ctx->mem_hash_lock);
 995         hash_for_each_possible(ctx->mem_hash, hnode, node, (unsigned long)vaddr)
 996                 if (vaddr == hnode->vaddr)
 997                         break;
 998
 999         if (!hnode) {
1000                 mutex_unlock(&ctx->mem_hash_lock);
1001                 dev_err(hdev->dev,
1002                         "unmap failed, no mem hnode for vaddr 0x%llx\n",
1003                         vaddr);
1004                 return -EINVAL;
1005         }
1006
1007         hash_del(&hnode->node);
1008         mutex_unlock(&ctx->mem_hash_lock);
1009
1010         vm_type = hnode->ptr;
1011
1012         if (*vm_type == VM_TYPE_USERPTR) {
1013                 is_userptr = true;
1014                 userptr = hnode->ptr;
1015                 rc = init_phys_pg_pack_from_userptr(ctx, userptr,
1016                                 &phys_pg_pack);
1017                 if (rc) {
1018                         dev_err(hdev->dev,
1019                                 "unable to init page pack for vaddr 0x%llx\n",
1020                                 vaddr);
1021                         goto vm_type_err;
1022                 }
1023         } else if (*vm_type == VM_TYPE_PHYS_PACK) {
1024                 is_userptr = false;
1025                 phys_pg_pack = hnode->ptr;
1026         } else {
1027                 dev_warn(hdev->dev,
1028                         "unmap failed, unknown vm desc for vaddr 0x%llx\n",
1029                                 vaddr);
1030                 rc = -EFAULT;
1031                 goto vm_type_err;
1032         }
1033
1034         if (atomic_read(&phys_pg_pack->mapping_cnt) == 0) {
1035                 dev_err(hdev->dev, "vaddr 0x%llx is not mapped\n", vaddr);
1036                 rc = -EINVAL;
1037                 goto mapping_cnt_err;
1038         }
1039
1040         page_size = phys_pg_pack->page_size;
1041         vaddr &= ~(((u64) page_size) - 1);
1042
1043         next_vaddr = vaddr;
1044
1045         mutex_lock(&ctx->mmu_lock);
1046
1047         for (i = 0 ; i < phys_pg_pack->npages ; i++, next_vaddr += page_size)
1048                 if (hl_mmu_unmap(ctx, next_vaddr, page_size))
1049                         dev_warn_ratelimited(hdev->dev,
1050                                 "unmap failed for vaddr: 0x%llx\n", next_vaddr);
1051
1052         hdev->asic_funcs->mmu_invalidate_cache(hdev, true);
1053
1054         mutex_unlock(&ctx->mmu_lock);
1055
1056         if (add_va_block(hdev,
1057                         is_userptr ? &ctx->host_va_range : &ctx->dram_va_range,
1058                         vaddr,
1059                         vaddr + phys_pg_pack->total_size - 1))
1060                 dev_warn(hdev->dev, "add va block failed for vaddr: 0x%llx\n",
1061                                 vaddr);
1062
1063         atomic_dec(&phys_pg_pack->mapping_cnt);
1064         kfree(hnode);
1065
1066         if (is_userptr) {
1067                 free_phys_pg_pack(hdev, phys_pg_pack);
1068                 free_userptr(hdev, userptr);
1069         }
1070
1071         return 0;
1072
1073 mapping_cnt_err:
1074         if (is_userptr)
1075                 free_phys_pg_pack(hdev, phys_pg_pack);
1076 vm_type_err:
1077         mutex_lock(&ctx->mem_hash_lock);
1078         hash_add(ctx->mem_hash, &hnode->node, vaddr);
1079         mutex_unlock(&ctx->mem_hash_lock);
1080
1081         return rc;
1082 }
1083
1084 int hl_mem_ioctl(struct hl_fpriv *hpriv, void *data)
1085 {
1086         union hl_mem_args *args = data;
1087         struct hl_device *hdev = hpriv->hdev;
1088         struct hl_ctx *ctx = hpriv->ctx;
1089         u64 device_addr = 0;
1090         u32 handle = 0;
1091         int rc;
1092
1093         if (hl_device_disabled_or_in_reset(hdev)) {
1094                 dev_warn_ratelimited(hdev->dev,
1095                         "Device is disabled or in reset. Can't execute memory IOCTL\n");
1096                 return -EBUSY;
1097         }
1098
1099         if (hdev->mmu_enable) {
1100                 switch (args->in.op) {
1101                 case HL_MEM_OP_ALLOC:
1102                         if (!hdev->dram_supports_virtual_memory) {
1103                                 dev_err(hdev->dev,
1104                                         "DRAM alloc is not supported\n");
1105                                 rc = -EINVAL;
1106                                 goto out;
1107                         }
1108                         if (args->in.alloc.mem_size == 0) {
1109                                 dev_err(hdev->dev,
1110                                         "alloc size must be larger than 0\n");
1111                                 rc = -EINVAL;
1112                                 goto out;
1113                         }
1114                         rc = alloc_device_memory(ctx, &args->in, &handle);
1115
1116                         memset(args, 0, sizeof(*args));
1117                         args->out.handle = (__u64) handle;
1118                         break;
1119
1120                 case HL_MEM_OP_FREE:
1121                         if (!hdev->dram_supports_virtual_memory) {
1122                                 dev_err(hdev->dev,
1123                                         "DRAM free is not supported\n");
1124                                 rc = -EINVAL;
1125                                 goto out;
1126                         }
1127                         rc = free_device_memory(ctx, args->in.free.handle);
1128                         break;
1129
1130                 case HL_MEM_OP_MAP:
1131                         rc = map_device_va(ctx, &args->in, &device_addr);
1132
1133                         memset(args, 0, sizeof(*args));
1134                         args->out.device_virt_addr = device_addr;
1135                         break;
1136
1137                 case HL_MEM_OP_UNMAP:
1138                         rc = unmap_device_va(ctx,
1139                                         args->in.unmap.device_virt_addr);
1140                         break;
1141
1142                 default:
1143                         dev_err(hdev->dev, "Unknown opcode for memory IOCTL\n");
1144                         rc = -ENOTTY;
1145                         break;
1146                 }
1147         } else {
1148                 switch (args->in.op) {
1149                 case HL_MEM_OP_ALLOC:
1150                         if (args->in.alloc.mem_size == 0) {
1151                                 dev_err(hdev->dev,
1152                                         "alloc size must be larger than 0\n");
1153                                 rc = -EINVAL;
1154                                 goto out;
1155                         }
1156
1157                         /* Force contiguous as there are no real MMU
1158                          * translations to overcome physical memory gaps
1159                          */
1160                         args->in.flags |= HL_MEM_CONTIGUOUS;
1161                         rc = alloc_device_memory(ctx, &args->in, &handle);
1162
1163                         memset(args, 0, sizeof(*args));
1164                         args->out.handle = (__u64) handle;
1165                         break;
1166
1167                 case HL_MEM_OP_FREE:
1168                         rc = free_device_memory(ctx, args->in.free.handle);
1169                         break;
1170
1171                 case HL_MEM_OP_MAP:
1172                         if (args->in.flags & HL_MEM_USERPTR) {
1173                                 device_addr = args->in.map_host.host_virt_addr;
1174                                 rc = 0;
1175                         } else {
1176                                 rc = get_paddr_from_handle(ctx, &args->in,
1177                                                 &device_addr);
1178                         }
1179
1180                         memset(args, 0, sizeof(*args));
1181                         args->out.device_virt_addr = device_addr;
1182                         break;
1183
1184                 case HL_MEM_OP_UNMAP:
1185                         rc = 0;
1186                         break;
1187
1188                 default:
1189                         dev_err(hdev->dev, "Unknown opcode for memory IOCTL\n");
1190                         rc = -ENOTTY;
1191                         break;
1192                 }
1193         }
1194
1195 out:
1196         return rc;
1197 }
1198
1199 /*
1200  * hl_pin_host_memory - pins a chunk of host memory
1201  *
1202  * @hdev                : pointer to the habanalabs device structure
1203  * @addr                : the user-space virtual address of the memory area
1204  * @size                : the size of the memory area
1205  * @userptr             : pointer to hl_userptr structure
1206  *
1207  * This function does the following:
1208  * - Pins the physical pages
1209  * - Create a SG list from those pages
1210  */
1211 int hl_pin_host_memory(struct hl_device *hdev, u64 addr, u64 size,
1212                         struct hl_userptr *userptr)
1213 {
1214         u64 start, end;
1215         u32 npages, offset;
1216         int rc;
1217
1218         if (!size) {
1219                 dev_err(hdev->dev, "size to pin is invalid - %llu\n", size);
1220                 return -EINVAL;
1221         }
1222
1223         if (!access_ok((void __user *) (uintptr_t) addr, size)) {
1224                 dev_err(hdev->dev, "user pointer is invalid - 0x%llx\n", addr);
1225                 return -EFAULT;
1226         }
1227
1228         /*
1229          * If the combination of the address and size requested for this memory
1230          * region causes an integer overflow, return error.
1231          */
1232         if (((addr + size) < addr) ||
1233                         PAGE_ALIGN(addr + size) < (addr + size)) {
1234                 dev_err(hdev->dev,
1235                         "user pointer 0x%llx + %llu causes integer overflow\n",
1236                         addr, size);
1237                 return -EINVAL;
1238         }
1239
1240         start = addr & PAGE_MASK;
1241         offset = addr & ~PAGE_MASK;
1242         end = PAGE_ALIGN(addr + size);
1243         npages = (end - start) >> PAGE_SHIFT;
1244
1245         userptr->size = size;
1246         userptr->addr = addr;
1247         userptr->dma_mapped = false;
1248         INIT_LIST_HEAD(&userptr->job_node);
1249
1250         userptr->vec = frame_vector_create(npages);
1251         if (!userptr->vec) {
1252                 dev_err(hdev->dev, "Failed to create frame vector\n");
1253                 return -ENOMEM;
1254         }
1255
1256         rc = get_vaddr_frames(start, npages, FOLL_FORCE | FOLL_WRITE,
1257                                 userptr->vec);
1258
1259         if (rc != npages) {
1260                 dev_err(hdev->dev,
1261                         "Failed to map host memory, user ptr probably wrong\n");
1262                 if (rc < 0)
1263                         goto destroy_framevec;
1264                 rc = -EFAULT;
1265                 goto put_framevec;
1266         }
1267
1268         if (frame_vector_to_pages(userptr->vec) < 0) {
1269                 dev_err(hdev->dev,
1270                         "Failed to translate frame vector to pages\n");
1271                 rc = -EFAULT;
1272                 goto put_framevec;
1273         }
1274
1275         userptr->sgt = kzalloc(sizeof(*userptr->sgt), GFP_ATOMIC);
1276         if (!userptr->sgt) {
1277                 rc = -ENOMEM;
1278                 goto put_framevec;
1279         }
1280
1281         rc = sg_alloc_table_from_pages(userptr->sgt,
1282                                         frame_vector_pages(userptr->vec),
1283                                         npages, offset, size, GFP_ATOMIC);
1284         if (rc < 0) {
1285                 dev_err(hdev->dev, "failed to create SG table from pages\n");
1286                 goto free_sgt;
1287         }
1288
1289         hl_debugfs_add_userptr(hdev, userptr);
1290
1291         return 0;
1292
1293 free_sgt:
1294         kfree(userptr->sgt);
1295 put_framevec:
1296         put_vaddr_frames(userptr->vec);
1297 destroy_framevec:
1298         frame_vector_destroy(userptr->vec);
1299         return rc;
1300 }
1301
1302 /*
1303  * hl_unpin_host_memory - unpins a chunk of host memory
1304  *
1305  * @hdev                : pointer to the habanalabs device structure
1306  * @userptr             : pointer to hl_userptr structure
1307  *
1308  * This function does the following:
1309  * - Unpins the physical pages related to the host memory
1310  * - Free the SG list
1311  */
1312 int hl_unpin_host_memory(struct hl_device *hdev, struct hl_userptr *userptr)
1313 {
1314         struct page **pages;
1315
1316         hl_debugfs_remove_userptr(hdev, userptr);
1317
1318         if (userptr->dma_mapped)
1319                 hdev->asic_funcs->hl_dma_unmap_sg(hdev,
1320                                 userptr->sgt->sgl,
1321                                 userptr->sgt->nents,
1322                                 userptr->dir);
1323
1324         pages = frame_vector_pages(userptr->vec);
1325         if (!IS_ERR(pages)) {
1326                 int i;
1327
1328                 for (i = 0; i < frame_vector_count(userptr->vec); i++)
1329                         set_page_dirty_lock(pages[i]);
1330         }
1331         put_vaddr_frames(userptr->vec);
1332         frame_vector_destroy(userptr->vec);
1333
1334         list_del(&userptr->job_node);
1335
1336         sg_free_table(userptr->sgt);
1337         kfree(userptr->sgt);
1338
1339         return 0;
1340 }
1341
1342 /*
1343  * hl_userptr_delete_list - clear userptr list
1344  *
1345  * @hdev                : pointer to the habanalabs device structure
1346  * @userptr_list        : pointer to the list to clear
1347  *
1348  * This function does the following:
1349  * - Iterates over the list and unpins the host memory and frees the userptr
1350  *   structure.
1351  */
1352 void hl_userptr_delete_list(struct hl_device *hdev,
1353                                 struct list_head *userptr_list)
1354 {
1355         struct hl_userptr *userptr, *tmp;
1356
1357         list_for_each_entry_safe(userptr, tmp, userptr_list, job_node) {
1358                 hl_unpin_host_memory(hdev, userptr);
1359                 kfree(userptr);
1360         }
1361
1362         INIT_LIST_HEAD(userptr_list);
1363 }
1364
1365 /*
1366  * hl_userptr_is_pinned - returns whether the given userptr is pinned
1367  *
1368  * @hdev                : pointer to the habanalabs device structure
1369  * @userptr_list        : pointer to the list to clear
1370  * @userptr             : pointer to userptr to check
1371  *
1372  * This function does the following:
1373  * - Iterates over the list and checks if the given userptr is in it, means is
1374  *   pinned. If so, returns true, otherwise returns false.
1375  */
1376 bool hl_userptr_is_pinned(struct hl_device *hdev, u64 addr,
1377                                 u32 size, struct list_head *userptr_list,
1378                                 struct hl_userptr **userptr)
1379 {
1380         list_for_each_entry((*userptr), userptr_list, job_node) {
1381                 if ((addr == (*userptr)->addr) && (size == (*userptr)->size))
1382                         return true;
1383         }
1384
1385         return false;
1386 }
1387
1388 /*
1389  * hl_va_range_init - initialize virtual addresses range
1390  *
1391  * @hdev                : pointer to the habanalabs device structure
1392  * @va_range            : pointer to the range to initialize
1393  * @start               : range start address
1394  * @end                 : range end address
1395  *
1396  * This function does the following:
1397  * - Initializes the virtual addresses list of the given range with the given
1398  *   addresses.
1399  */
1400 static int hl_va_range_init(struct hl_device *hdev,
1401                 struct hl_va_range *va_range, u64 start, u64 end)
1402 {
1403         int rc;
1404
1405         INIT_LIST_HEAD(&va_range->list);
1406
1407         /* PAGE_SIZE alignment */
1408
1409         if (start & (PAGE_SIZE - 1)) {
1410                 start &= PAGE_MASK;
1411                 start += PAGE_SIZE;
1412         }
1413
1414         if (end & (PAGE_SIZE - 1))
1415                 end &= PAGE_MASK;
1416
1417         if (start >= end) {
1418                 dev_err(hdev->dev, "too small vm range for va list\n");
1419                 return -EFAULT;
1420         }
1421
1422         rc = add_va_block(hdev, va_range, start, end);
1423
1424         if (rc) {
1425                 dev_err(hdev->dev, "Failed to init host va list\n");
1426                 return rc;
1427         }
1428
1429         va_range->start_addr = start;
1430         va_range->end_addr = end;
1431
1432         return 0;
1433 }
1434
1435 /*
1436  * hl_vm_ctx_init_with_ranges - initialize virtual memory for context
1437  *
1438  * @ctx                 : pointer to the habanalabs context structure
1439  * @host_range_start    : host virtual addresses range start
1440  * @host_range_end      : host virtual addresses range end
1441  * @dram_range_start    : dram virtual addresses range start
1442  * @dram_range_end      : dram virtual addresses range end
1443  *
1444  * This function initializes the following:
1445  * - MMU for context
1446  * - Virtual address to area descriptor hashtable
1447  * - Virtual block list of available virtual memory
1448  */
1449 static int hl_vm_ctx_init_with_ranges(struct hl_ctx *ctx, u64 host_range_start,
1450                                 u64 host_range_end, u64 dram_range_start,
1451                                 u64 dram_range_end)
1452 {
1453         struct hl_device *hdev = ctx->hdev;
1454         int rc;
1455
1456         rc = hl_mmu_ctx_init(ctx);
1457         if (rc) {
1458                 dev_err(hdev->dev, "failed to init context %d\n", ctx->asid);
1459                 return rc;
1460         }
1461
1462         mutex_init(&ctx->mem_hash_lock);
1463         hash_init(ctx->mem_hash);
1464
1465         mutex_init(&ctx->host_va_range.lock);
1466
1467         rc = hl_va_range_init(hdev, &ctx->host_va_range, host_range_start,
1468                         host_range_end);
1469         if (rc) {
1470                 dev_err(hdev->dev, "failed to init host vm range\n");
1471                 goto host_vm_err;
1472         }
1473
1474         mutex_init(&ctx->dram_va_range.lock);
1475
1476         rc = hl_va_range_init(hdev, &ctx->dram_va_range, dram_range_start,
1477                         dram_range_end);
1478         if (rc) {
1479                 dev_err(hdev->dev, "failed to init dram vm range\n");
1480                 goto dram_vm_err;
1481         }
1482
1483         hl_debugfs_add_ctx_mem_hash(hdev, ctx);
1484
1485         return 0;
1486
1487 dram_vm_err:
1488         mutex_destroy(&ctx->dram_va_range.lock);
1489
1490         mutex_lock(&ctx->host_va_range.lock);
1491         clear_va_list_locked(hdev, &ctx->host_va_range.list);
1492         mutex_unlock(&ctx->host_va_range.lock);
1493 host_vm_err:
1494         mutex_destroy(&ctx->host_va_range.lock);
1495         mutex_destroy(&ctx->mem_hash_lock);
1496         hl_mmu_ctx_fini(ctx);
1497
1498         return rc;
1499 }
1500
1501 int hl_vm_ctx_init(struct hl_ctx *ctx)
1502 {
1503         struct asic_fixed_properties *prop = &ctx->hdev->asic_prop;
1504         u64 host_range_start, host_range_end, dram_range_start,
1505                 dram_range_end;
1506
1507         atomic64_set(&ctx->dram_phys_mem, 0);
1508
1509         /*
1510          * - If MMU is enabled, init the ranges as usual.
1511          * - If MMU is disabled, in case of host mapping, the returned address
1512          *   is the given one.
1513          *   In case of DRAM mapping, the returned address is the physical
1514          *   address of the memory related to the given handle.
1515          */
1516         if (ctx->hdev->mmu_enable) {
1517                 dram_range_start = prop->va_space_dram_start_address;
1518                 dram_range_end = prop->va_space_dram_end_address;
1519                 host_range_start = prop->va_space_host_start_address;
1520                 host_range_end = prop->va_space_host_end_address;
1521         } else {
1522                 dram_range_start = prop->dram_user_base_address;
1523                 dram_range_end = prop->dram_end_address;
1524                 host_range_start = prop->dram_user_base_address;
1525                 host_range_end = prop->dram_end_address;
1526         }
1527
1528         return hl_vm_ctx_init_with_ranges(ctx, host_range_start, host_range_end,
1529                         dram_range_start, dram_range_end);
1530 }
1531
1532 /*
1533  * hl_va_range_fini     - clear a virtual addresses range
1534  *
1535  * @hdev                : pointer to the habanalabs structure
1536  * va_range             : pointer to virtual addresses range
1537  *
1538  * This function initializes the following:
1539  * - Checks that the given range contains the whole initial range
1540  * - Frees the virtual addresses block list and its lock
1541  */
1542 static void hl_va_range_fini(struct hl_device *hdev,
1543                 struct hl_va_range *va_range)
1544 {
1545         struct hl_vm_va_block *va_block;
1546
1547         if (list_empty(&va_range->list)) {
1548                 dev_warn(hdev->dev,
1549                                 "va list should not be empty on cleanup!\n");
1550                 goto out;
1551         }
1552
1553         if (!list_is_singular(&va_range->list)) {
1554                 dev_warn(hdev->dev,
1555                         "va list should not contain multiple blocks on cleanup!\n");
1556                 goto free_va_list;
1557         }
1558
1559         va_block = list_first_entry(&va_range->list, typeof(*va_block), node);
1560
1561         if (va_block->start != va_range->start_addr ||
1562                 va_block->end != va_range->end_addr) {
1563                 dev_warn(hdev->dev,
1564                         "wrong va block on cleanup, from 0x%llx to 0x%llx\n",
1565                                 va_block->start, va_block->end);
1566                 goto free_va_list;
1567         }
1568
1569 free_va_list:
1570         mutex_lock(&va_range->lock);
1571         clear_va_list_locked(hdev, &va_range->list);
1572         mutex_unlock(&va_range->lock);
1573
1574 out:
1575         mutex_destroy(&va_range->lock);
1576 }
1577
1578 /*
1579  * hl_vm_ctx_fini       - virtual memory teardown of context
1580  *
1581  * @ctx                 : pointer to the habanalabs context structure
1582  *
1583  * This function perform teardown the following:
1584  * - Virtual block list of available virtual memory
1585  * - Virtual address to area descriptor hashtable
1586  * - MMU for context
1587  *
1588  * In addition this function does the following:
1589  * - Unmaps the existing hashtable nodes if the hashtable is not empty. The
1590  *   hashtable should be empty as no valid mappings should exist at this
1591  *   point.
1592  * - Frees any existing physical page list from the idr which relates to the
1593  *   current context asid.
1594  * - This function checks the virtual block list for correctness. At this point
1595  *   the list should contain one element which describes the whole virtual
1596  *   memory range of the context. Otherwise, a warning is printed.
1597  */
1598 void hl_vm_ctx_fini(struct hl_ctx *ctx)
1599 {
1600         struct hl_device *hdev = ctx->hdev;
1601         struct hl_vm *vm = &hdev->vm;
1602         struct hl_vm_phys_pg_pack *phys_pg_list;
1603         struct hl_vm_hash_node *hnode;
1604         struct hlist_node *tmp_node;
1605         int i;
1606
1607         hl_debugfs_remove_ctx_mem_hash(hdev, ctx);
1608
1609         if (!hash_empty(ctx->mem_hash))
1610                 dev_notice(hdev->dev, "ctx is freed while it has va in use\n");
1611
1612         hash_for_each_safe(ctx->mem_hash, i, tmp_node, hnode, node) {
1613                 dev_dbg(hdev->dev,
1614                         "hl_mem_hash_node of vaddr 0x%llx of asid %d is still alive\n",
1615                         hnode->vaddr, ctx->asid);
1616                 unmap_device_va(ctx, hnode->vaddr);
1617         }
1618
1619         spin_lock(&vm->idr_lock);
1620         idr_for_each_entry(&vm->phys_pg_pack_handles, phys_pg_list, i)
1621                 if (phys_pg_list->asid == ctx->asid) {
1622                         dev_dbg(hdev->dev,
1623                                 "page list 0x%p of asid %d is still alive\n",
1624                                 phys_pg_list, ctx->asid);
1625                         free_phys_pg_pack(hdev, phys_pg_list);
1626                         idr_remove(&vm->phys_pg_pack_handles, i);
1627                 }
1628         spin_unlock(&vm->idr_lock);
1629
1630         hl_va_range_fini(hdev, &ctx->dram_va_range);
1631         hl_va_range_fini(hdev, &ctx->host_va_range);
1632
1633         mutex_destroy(&ctx->mem_hash_lock);
1634         hl_mmu_ctx_fini(ctx);
1635 }
1636
1637 /*
1638  * hl_vm_init           - initialize virtual memory module
1639  *
1640  * @hdev                : pointer to the habanalabs device structure
1641  *
1642  * This function initializes the following:
1643  * - MMU module
1644  * - DRAM physical pages pool of 2MB
1645  * - Idr for device memory allocation handles
1646  */
1647 int hl_vm_init(struct hl_device *hdev)
1648 {
1649         struct asic_fixed_properties *prop = &hdev->asic_prop;
1650         struct hl_vm *vm = &hdev->vm;
1651         int rc;
1652
1653         rc = hl_mmu_init(hdev);
1654         if (rc) {
1655                 dev_err(hdev->dev, "Failed to init MMU\n");
1656                 return rc;
1657         }
1658
1659         vm->dram_pg_pool = gen_pool_create(__ffs(prop->dram_page_size), -1);
1660         if (!vm->dram_pg_pool) {
1661                 dev_err(hdev->dev, "Failed to create dram page pool\n");
1662                 rc = -ENOMEM;
1663                 goto pool_create_err;
1664         }
1665
1666         kref_init(&vm->dram_pg_pool_refcount);
1667
1668         rc = gen_pool_add(vm->dram_pg_pool, prop->dram_user_base_address,
1669                         prop->dram_end_address - prop->dram_user_base_address,
1670                         -1);
1671
1672         if (rc) {
1673                 dev_err(hdev->dev,
1674                         "Failed to add memory to dram page pool %d\n", rc);
1675                 goto pool_add_err;
1676         }
1677
1678         spin_lock_init(&vm->idr_lock);
1679         idr_init(&vm->phys_pg_pack_handles);
1680
1681         atomic64_set(&hdev->dram_used_mem, 0);
1682
1683         vm->init_done = true;
1684
1685         return 0;
1686
1687 pool_add_err:
1688         gen_pool_destroy(vm->dram_pg_pool);
1689 pool_create_err:
1690         hl_mmu_fini(hdev);
1691
1692         return rc;
1693 }
1694
1695 /*
1696  * hl_vm_fini           - virtual memory module teardown
1697  *
1698  * @hdev                : pointer to the habanalabs device structure
1699  *
1700  * This function perform teardown to the following:
1701  * - Idr for device memory allocation handles
1702  * - DRAM physical pages pool of 2MB
1703  * - MMU module
1704  */
1705 void hl_vm_fini(struct hl_device *hdev)
1706 {
1707         struct hl_vm *vm = &hdev->vm;
1708
1709         if (!vm->init_done)
1710                 return;
1711
1712         /*
1713          * At this point all the contexts should be freed and hence no DRAM
1714          * memory should be in use. Hence the DRAM pool should be freed here.
1715          */
1716         if (kref_put(&vm->dram_pg_pool_refcount, dram_pg_pool_do_release) != 1)
1717                 dev_warn(hdev->dev, "dram_pg_pool was not destroyed on %s\n",
1718                                 __func__);
1719
1720         hl_mmu_fini(hdev);
1721
1722         vm->init_done = false;
1723 }