2 * Copyright 2018 Red Hat Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
22 #include "nouveau_dmem.h"
23 #include "nouveau_drv.h"
24 #include "nouveau_chan.h"
25 #include "nouveau_dma.h"
26 #include "nouveau_mem.h"
27 #include "nouveau_bo.h"
28 #include "nouveau_svm.h"
30 #include <nvif/class.h>
31 #include <nvif/object.h>
32 #include <nvif/push906f.h>
33 #include <nvif/if000c.h>
34 #include <nvif/if500b.h>
35 #include <nvif/if900b.h>
36 #include <nvif/if000c.h>
38 #include <nvhw/class/cla0b5.h>
40 #include <linux/sched/mm.h>
41 #include <linux/hmm.h>
44 * FIXME: this is ugly right now we are using TTM to allocate vram and we pin
45 * it in vram while in use. We likely want to overhaul memory management for
46 * nouveau to be more page like (not necessarily with system page size but a
47 * bigger page size) at lowest level and have some shim layer on top that would
48 * provide the same functionality as TTM.
50 #define DMEM_CHUNK_SIZE (2UL << 20)
51 #define DMEM_CHUNK_NPAGES (DMEM_CHUNK_SIZE >> PAGE_SHIFT)
59 typedef int (*nouveau_migrate_copy_t)(struct nouveau_drm *drm, u64 npages,
60 enum nouveau_aper, u64 dst_addr,
61 enum nouveau_aper, u64 src_addr);
62 typedef int (*nouveau_clear_page_t)(struct nouveau_drm *drm, u32 length,
63 enum nouveau_aper, u64 dst_addr);
65 struct nouveau_dmem_chunk {
66 struct list_head list;
67 struct nouveau_bo *bo;
68 struct nouveau_drm *drm;
69 unsigned long callocated;
70 struct dev_pagemap pagemap;
73 struct nouveau_dmem_migrate {
74 nouveau_migrate_copy_t copy_func;
75 nouveau_clear_page_t clear_func;
76 struct nouveau_channel *chan;
80 struct nouveau_drm *drm;
81 struct nouveau_dmem_migrate migrate;
82 struct list_head chunks;
84 struct page *free_pages;
88 static struct nouveau_dmem_chunk *nouveau_page_to_chunk(struct page *page)
90 return container_of(page->pgmap, struct nouveau_dmem_chunk, pagemap);
93 static struct nouveau_drm *page_to_drm(struct page *page)
95 struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
100 unsigned long nouveau_dmem_page_addr(struct page *page)
102 struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
103 unsigned long off = (page_to_pfn(page) << PAGE_SHIFT) -
104 chunk->pagemap.res.start;
106 return chunk->bo->offset + off;
109 static void nouveau_dmem_page_free(struct page *page)
111 struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page);
112 struct nouveau_dmem *dmem = chunk->drm->dmem;
114 spin_lock(&dmem->lock);
115 page->zone_device_data = dmem->free_pages;
116 dmem->free_pages = page;
118 WARN_ON(!chunk->callocated);
121 * FIXME when chunk->callocated reach 0 we should add the chunk to
122 * a reclaim list so that it can be freed in case of memory pressure.
124 spin_unlock(&dmem->lock);
127 static void nouveau_dmem_fence_done(struct nouveau_fence **fence)
130 nouveau_fence_wait(*fence, true, false);
131 nouveau_fence_unref(fence);
134 * FIXME wait for channel to be IDLE before calling finalizing
140 static vm_fault_t nouveau_dmem_fault_copy_one(struct nouveau_drm *drm,
141 struct vm_fault *vmf, struct migrate_vma *args,
142 dma_addr_t *dma_addr)
144 struct device *dev = drm->dev->dev;
145 struct page *dpage, *spage;
146 struct nouveau_svmm *svmm;
148 spage = migrate_pfn_to_page(args->src[0]);
149 if (!spage || !(args->src[0] & MIGRATE_PFN_MIGRATE))
152 dpage = alloc_page_vma(GFP_HIGHUSER, vmf->vma, vmf->address);
154 return VM_FAULT_SIGBUS;
157 *dma_addr = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
158 if (dma_mapping_error(dev, *dma_addr))
159 goto error_free_page;
161 svmm = spage->zone_device_data;
162 mutex_lock(&svmm->mutex);
163 nouveau_svmm_invalidate(svmm, args->start, args->end);
164 if (drm->dmem->migrate.copy_func(drm, 1, NOUVEAU_APER_HOST, *dma_addr,
165 NOUVEAU_APER_VRAM, nouveau_dmem_page_addr(spage)))
166 goto error_dma_unmap;
167 mutex_unlock(&svmm->mutex);
169 args->dst[0] = migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED;
173 mutex_unlock(&svmm->mutex);
174 dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
177 return VM_FAULT_SIGBUS;
180 static vm_fault_t nouveau_dmem_migrate_to_ram(struct vm_fault *vmf)
182 struct nouveau_drm *drm = page_to_drm(vmf->page);
183 struct nouveau_dmem *dmem = drm->dmem;
184 struct nouveau_fence *fence;
185 unsigned long src = 0, dst = 0;
186 dma_addr_t dma_addr = 0;
188 struct migrate_vma args = {
190 .start = vmf->address,
191 .end = vmf->address + PAGE_SIZE,
194 .pgmap_owner = drm->dev,
195 .flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE,
199 * FIXME what we really want is to find some heuristic to migrate more
200 * than just one page on CPU fault. When such fault happens it is very
201 * likely that more surrounding page will CPU fault too.
203 if (migrate_vma_setup(&args) < 0)
204 return VM_FAULT_SIGBUS;
208 ret = nouveau_dmem_fault_copy_one(drm, vmf, &args, &dma_addr);
212 nouveau_fence_new(dmem->migrate.chan, false, &fence);
213 migrate_vma_pages(&args);
214 nouveau_dmem_fence_done(&fence);
215 dma_unmap_page(drm->dev->dev, dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
217 migrate_vma_finalize(&args);
221 static const struct dev_pagemap_ops nouveau_dmem_pagemap_ops = {
222 .page_free = nouveau_dmem_page_free,
223 .migrate_to_ram = nouveau_dmem_migrate_to_ram,
227 nouveau_dmem_chunk_alloc(struct nouveau_drm *drm, struct page **ppage)
229 struct nouveau_dmem_chunk *chunk;
230 struct resource *res;
233 unsigned long i, pfn_first;
236 chunk = kzalloc(sizeof(*chunk), GFP_KERNEL);
242 /* Allocate unused physical address space for device private pages. */
243 res = request_free_mem_region(&iomem_resource, DMEM_CHUNK_SIZE,
251 chunk->pagemap.type = MEMORY_DEVICE_PRIVATE;
252 chunk->pagemap.res = *res;
253 chunk->pagemap.ops = &nouveau_dmem_pagemap_ops;
254 chunk->pagemap.owner = drm->dev;
256 ret = nouveau_bo_new(&drm->client, DMEM_CHUNK_SIZE, 0,
257 TTM_PL_FLAG_VRAM, 0, 0, NULL, NULL,
262 ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false);
266 ptr = memremap_pages(&chunk->pagemap, numa_node_id());
272 mutex_lock(&drm->dmem->mutex);
273 list_add(&chunk->list, &drm->dmem->chunks);
274 mutex_unlock(&drm->dmem->mutex);
276 pfn_first = chunk->pagemap.res.start >> PAGE_SHIFT;
277 page = pfn_to_page(pfn_first);
278 spin_lock(&drm->dmem->lock);
279 for (i = 0; i < DMEM_CHUNK_NPAGES - 1; ++i, ++page) {
280 page->zone_device_data = drm->dmem->free_pages;
281 drm->dmem->free_pages = page;
285 spin_unlock(&drm->dmem->lock);
287 NV_INFO(drm, "DMEM: registered %ldMB of device memory\n",
288 DMEM_CHUNK_SIZE >> 20);
293 nouveau_bo_unpin(chunk->bo);
295 nouveau_bo_ref(NULL, &chunk->bo);
297 release_mem_region(chunk->pagemap.res.start,
298 resource_size(&chunk->pagemap.res));
306 nouveau_dmem_page_alloc_locked(struct nouveau_drm *drm)
308 struct nouveau_dmem_chunk *chunk;
309 struct page *page = NULL;
312 spin_lock(&drm->dmem->lock);
313 if (drm->dmem->free_pages) {
314 page = drm->dmem->free_pages;
315 drm->dmem->free_pages = page->zone_device_data;
316 chunk = nouveau_page_to_chunk(page);
318 spin_unlock(&drm->dmem->lock);
320 spin_unlock(&drm->dmem->lock);
321 ret = nouveau_dmem_chunk_alloc(drm, &page);
332 nouveau_dmem_page_free_locked(struct nouveau_drm *drm, struct page *page)
339 nouveau_dmem_resume(struct nouveau_drm *drm)
341 struct nouveau_dmem_chunk *chunk;
344 if (drm->dmem == NULL)
347 mutex_lock(&drm->dmem->mutex);
348 list_for_each_entry(chunk, &drm->dmem->chunks, list) {
349 ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false);
350 /* FIXME handle pin failure */
353 mutex_unlock(&drm->dmem->mutex);
357 nouveau_dmem_suspend(struct nouveau_drm *drm)
359 struct nouveau_dmem_chunk *chunk;
361 if (drm->dmem == NULL)
364 mutex_lock(&drm->dmem->mutex);
365 list_for_each_entry(chunk, &drm->dmem->chunks, list)
366 nouveau_bo_unpin(chunk->bo);
367 mutex_unlock(&drm->dmem->mutex);
371 nouveau_dmem_fini(struct nouveau_drm *drm)
373 struct nouveau_dmem_chunk *chunk, *tmp;
375 if (drm->dmem == NULL)
378 mutex_lock(&drm->dmem->mutex);
380 list_for_each_entry_safe(chunk, tmp, &drm->dmem->chunks, list) {
381 nouveau_bo_unpin(chunk->bo);
382 nouveau_bo_ref(NULL, &chunk->bo);
383 list_del(&chunk->list);
384 memunmap_pages(&chunk->pagemap);
385 release_mem_region(chunk->pagemap.res.start,
386 resource_size(&chunk->pagemap.res));
390 mutex_unlock(&drm->dmem->mutex);
394 nvc0b5_migrate_copy(struct nouveau_drm *drm, u64 npages,
395 enum nouveau_aper dst_aper, u64 dst_addr,
396 enum nouveau_aper src_aper, u64 src_addr)
398 struct nvif_push *push = drm->dmem->migrate.chan->chan.push;
402 ret = PUSH_WAIT(push, 13);
406 if (src_aper != NOUVEAU_APER_VIRT) {
408 case NOUVEAU_APER_VRAM:
409 PUSH_IMMD(push, NVA0B5, SET_SRC_PHYS_MODE,
410 NVDEF(NVA0B5, SET_SRC_PHYS_MODE, TARGET, LOCAL_FB));
412 case NOUVEAU_APER_HOST:
413 PUSH_IMMD(push, NVA0B5, SET_SRC_PHYS_MODE,
414 NVDEF(NVA0B5, SET_SRC_PHYS_MODE, TARGET, COHERENT_SYSMEM));
420 launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, SRC_TYPE, PHYSICAL);
423 if (dst_aper != NOUVEAU_APER_VIRT) {
425 case NOUVEAU_APER_VRAM:
426 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
427 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, LOCAL_FB));
429 case NOUVEAU_APER_HOST:
430 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
431 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, COHERENT_SYSMEM));
437 launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, DST_TYPE, PHYSICAL);
440 PUSH_MTHD(push, NVA0B5, OFFSET_IN_UPPER,
441 NVVAL(NVA0B5, OFFSET_IN_UPPER, UPPER, upper_32_bits(src_addr)),
443 OFFSET_IN_LOWER, lower_32_bits(src_addr),
446 NVVAL(NVA0B5, OFFSET_OUT_UPPER, UPPER, upper_32_bits(dst_addr)),
448 OFFSET_OUT_LOWER, lower_32_bits(dst_addr),
450 PITCH_OUT, PAGE_SIZE,
451 LINE_LENGTH_IN, PAGE_SIZE,
454 PUSH_MTHD(push, NVA0B5, LAUNCH_DMA, launch_dma |
455 NVDEF(NVA0B5, LAUNCH_DMA, DATA_TRANSFER_TYPE, NON_PIPELINED) |
456 NVDEF(NVA0B5, LAUNCH_DMA, FLUSH_ENABLE, TRUE) |
457 NVDEF(NVA0B5, LAUNCH_DMA, SEMAPHORE_TYPE, NONE) |
458 NVDEF(NVA0B5, LAUNCH_DMA, INTERRUPT_TYPE, NONE) |
459 NVDEF(NVA0B5, LAUNCH_DMA, SRC_MEMORY_LAYOUT, PITCH) |
460 NVDEF(NVA0B5, LAUNCH_DMA, DST_MEMORY_LAYOUT, PITCH) |
461 NVDEF(NVA0B5, LAUNCH_DMA, MULTI_LINE_ENABLE, TRUE) |
462 NVDEF(NVA0B5, LAUNCH_DMA, REMAP_ENABLE, FALSE) |
463 NVDEF(NVA0B5, LAUNCH_DMA, BYPASS_L2, USE_PTE_SETTING));
468 nvc0b5_migrate_clear(struct nouveau_drm *drm, u32 length,
469 enum nouveau_aper dst_aper, u64 dst_addr)
471 struct nvif_push *push = drm->dmem->migrate.chan->chan.push;
475 ret = PUSH_WAIT(push, 12);
480 case NOUVEAU_APER_VRAM:
481 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
482 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, LOCAL_FB));
484 case NOUVEAU_APER_HOST:
485 PUSH_IMMD(push, NVA0B5, SET_DST_PHYS_MODE,
486 NVDEF(NVA0B5, SET_DST_PHYS_MODE, TARGET, COHERENT_SYSMEM));
492 launch_dma |= NVDEF(NVA0B5, LAUNCH_DMA, DST_TYPE, PHYSICAL);
494 PUSH_MTHD(push, NVA0B5, SET_REMAP_CONST_A, 0,
495 SET_REMAP_CONST_B, 0,
497 SET_REMAP_COMPONENTS,
498 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, DST_X, CONST_A) |
499 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, DST_Y, CONST_B) |
500 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, COMPONENT_SIZE, FOUR) |
501 NVDEF(NVA0B5, SET_REMAP_COMPONENTS, NUM_DST_COMPONENTS, TWO));
503 PUSH_MTHD(push, NVA0B5, OFFSET_OUT_UPPER,
504 NVVAL(NVA0B5, OFFSET_OUT_UPPER, UPPER, upper_32_bits(dst_addr)),
506 OFFSET_OUT_LOWER, lower_32_bits(dst_addr));
508 PUSH_MTHD(push, NVA0B5, LINE_LENGTH_IN, length >> 3);
510 PUSH_MTHD(push, NVA0B5, LAUNCH_DMA, launch_dma |
511 NVDEF(NVA0B5, LAUNCH_DMA, DATA_TRANSFER_TYPE, NON_PIPELINED) |
512 NVDEF(NVA0B5, LAUNCH_DMA, FLUSH_ENABLE, TRUE) |
513 NVDEF(NVA0B5, LAUNCH_DMA, SEMAPHORE_TYPE, NONE) |
514 NVDEF(NVA0B5, LAUNCH_DMA, INTERRUPT_TYPE, NONE) |
515 NVDEF(NVA0B5, LAUNCH_DMA, SRC_MEMORY_LAYOUT, PITCH) |
516 NVDEF(NVA0B5, LAUNCH_DMA, DST_MEMORY_LAYOUT, PITCH) |
517 NVDEF(NVA0B5, LAUNCH_DMA, MULTI_LINE_ENABLE, FALSE) |
518 NVDEF(NVA0B5, LAUNCH_DMA, REMAP_ENABLE, TRUE) |
519 NVDEF(NVA0B5, LAUNCH_DMA, BYPASS_L2, USE_PTE_SETTING));
524 nouveau_dmem_migrate_init(struct nouveau_drm *drm)
526 switch (drm->ttm.copy.oclass) {
527 case PASCAL_DMA_COPY_A:
528 case PASCAL_DMA_COPY_B:
529 case VOLTA_DMA_COPY_A:
530 case TURING_DMA_COPY_A:
531 drm->dmem->migrate.copy_func = nvc0b5_migrate_copy;
532 drm->dmem->migrate.clear_func = nvc0b5_migrate_clear;
533 drm->dmem->migrate.chan = drm->ttm.chan;
542 nouveau_dmem_init(struct nouveau_drm *drm)
546 /* This only make sense on PASCAL or newer */
547 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_PASCAL)
550 if (!(drm->dmem = kzalloc(sizeof(*drm->dmem), GFP_KERNEL)))
553 drm->dmem->drm = drm;
554 mutex_init(&drm->dmem->mutex);
555 INIT_LIST_HEAD(&drm->dmem->chunks);
556 mutex_init(&drm->dmem->mutex);
557 spin_lock_init(&drm->dmem->lock);
559 /* Initialize migration dma helpers before registering memory */
560 ret = nouveau_dmem_migrate_init(drm);
567 static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm,
568 struct nouveau_svmm *svmm, unsigned long src,
569 dma_addr_t *dma_addr, u64 *pfn)
571 struct device *dev = drm->dev->dev;
572 struct page *dpage, *spage;
575 spage = migrate_pfn_to_page(src);
576 if (!(src & MIGRATE_PFN_MIGRATE))
579 dpage = nouveau_dmem_page_alloc_locked(drm);
583 paddr = nouveau_dmem_page_addr(dpage);
585 *dma_addr = dma_map_page(dev, spage, 0, page_size(spage),
587 if (dma_mapping_error(dev, *dma_addr))
589 if (drm->dmem->migrate.copy_func(drm, 1,
590 NOUVEAU_APER_VRAM, paddr, NOUVEAU_APER_HOST, *dma_addr))
593 *dma_addr = DMA_MAPPING_ERROR;
594 if (drm->dmem->migrate.clear_func(drm, page_size(dpage),
595 NOUVEAU_APER_VRAM, paddr))
599 dpage->zone_device_data = svmm;
600 *pfn = NVIF_VMM_PFNMAP_V0_V | NVIF_VMM_PFNMAP_V0_VRAM |
601 ((paddr >> PAGE_SHIFT) << NVIF_VMM_PFNMAP_V0_ADDR_SHIFT);
602 if (src & MIGRATE_PFN_WRITE)
603 *pfn |= NVIF_VMM_PFNMAP_V0_W;
604 return migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED;
607 dma_unmap_page(dev, *dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
609 nouveau_dmem_page_free_locked(drm, dpage);
611 *pfn = NVIF_VMM_PFNMAP_V0_NONE;
615 static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm,
616 struct nouveau_svmm *svmm, struct migrate_vma *args,
617 dma_addr_t *dma_addrs, u64 *pfns)
619 struct nouveau_fence *fence;
620 unsigned long addr = args->start, nr_dma = 0, i;
622 for (i = 0; addr < args->end; i++) {
623 args->dst[i] = nouveau_dmem_migrate_copy_one(drm, svmm,
624 args->src[i], dma_addrs + nr_dma, pfns + i);
625 if (!dma_mapping_error(drm->dev->dev, dma_addrs[nr_dma]))
630 nouveau_fence_new(drm->dmem->migrate.chan, false, &fence);
631 migrate_vma_pages(args);
632 nouveau_dmem_fence_done(&fence);
633 nouveau_pfns_map(svmm, args->vma->vm_mm, args->start, pfns, i);
636 dma_unmap_page(drm->dev->dev, dma_addrs[nr_dma], PAGE_SIZE,
639 migrate_vma_finalize(args);
643 nouveau_dmem_migrate_vma(struct nouveau_drm *drm,
644 struct nouveau_svmm *svmm,
645 struct vm_area_struct *vma,
649 unsigned long npages = (end - start) >> PAGE_SHIFT;
650 unsigned long max = min(SG_MAX_SINGLE_ALLOC, npages);
651 dma_addr_t *dma_addrs;
652 struct migrate_vma args = {
655 .pgmap_owner = drm->dev,
656 .flags = MIGRATE_VMA_SELECT_SYSTEM,
662 if (drm->dmem == NULL)
665 args.src = kcalloc(max, sizeof(*args.src), GFP_KERNEL);
668 args.dst = kcalloc(max, sizeof(*args.dst), GFP_KERNEL);
672 dma_addrs = kmalloc_array(max, sizeof(*dma_addrs), GFP_KERNEL);
676 pfns = nouveau_pfns_alloc(max);
680 for (i = 0; i < npages; i += max) {
681 args.end = start + (max << PAGE_SHIFT);
682 ret = migrate_vma_setup(&args);
687 nouveau_dmem_migrate_chunk(drm, svmm, &args, dma_addrs,
689 args.start = args.end;
694 nouveau_pfns_free(pfns);