From 2733ea144dcce789de20988c1056e228a07b1bff Mon Sep 17 00:00:00 2001 From: Jason Gunthorpe Date: Fri, 1 May 2020 15:20:48 -0300 Subject: mm/hmm: remove the customizable pfn format from hmm_range_fault Presumably the intent here was that hmm_range_fault() could put the data into some HW specific format and thus avoid some work. However, nothing actually does that, and it isn't clear how anything actually could do that as hmm_range_fault() provides CPU addresses which must be DMA mapped. Perhaps there is some special HW that does not need DMA mapping, but we don't have any examples of this, and the theoretical performance win of avoiding an extra scan over the pfns array doesn't seem worth the complexity. Plus pfns needs to be scanned anyhow to sort out any DEVICE_PRIVATE pages. This version replaces the uint64_t with an usigned long containing a pfn and fixed flags. On input flags is filled with the HMM_PFN_REQ_* values, on successful output it is filled with HMM_PFN_* values, describing the state of the pages. amdgpu is simple to convert, it doesn't use snapshot and doesn't use per-page flags. nouveau uses only 16 hmm_pte entries at most (ie fits in a few cache lines), and it sweeps over its pfns array a couple of times anyhow. It also has a nasty call chain before it reaches the dma map and hardware suggesting performance isn't important: nouveau_svm_fault(): args.i.m.method = NVIF_VMM_V0_PFNMAP nouveau_range_fault() nvif_object_ioctl() client->driver->ioctl() struct nvif_driver nvif_driver_nvkm: .ioctl = nvkm_client_ioctl nvkm_ioctl() nvkm_ioctl_path() nvkm_ioctl_v0[type].func(..) nvkm_ioctl_mthd() nvkm_object_mthd() struct nvkm_object_func nvkm_uvmm: .mthd = nvkm_uvmm_mthd nvkm_uvmm_mthd() nvkm_uvmm_mthd_pfnmap() nvkm_vmm_pfn_map() nvkm_vmm_ptes_get_map() func == gp100_vmm_pgt_pfn struct nvkm_vmm_desc_func gp100_vmm_desc_spt: .pfn = gp100_vmm_pgt_pfn nvkm_vmm_iter() REF_PTES == func == gp100_vmm_pgt_pfn() dma_map_page() Link: https://lore.kernel.org/r/5-v2-b4e84f444c7d+24f57-hmm_no_flags_jgg@mellanox.com Acked-by: Felix Kuehling Tested-by: Ralph Campbell Signed-off-by: Christoph Hellwig Reviewed-by: Christoph Hellwig Signed-off-by: Jason Gunthorpe --- drivers/gpu/drm/nouveau/nouveau_dmem.c | 27 +-------------------------- 1 file changed, 1 insertion(+), 26 deletions(-) (limited to 'drivers/gpu/drm/nouveau/nouveau_dmem.c') diff --git a/drivers/gpu/drm/nouveau/nouveau_dmem.c b/drivers/gpu/drm/nouveau/nouveau_dmem.c index ad89e09a0be3..3364904eccff 100644 --- a/drivers/gpu/drm/nouveau/nouveau_dmem.c +++ b/drivers/gpu/drm/nouveau/nouveau_dmem.c @@ -85,7 +85,7 @@ static inline struct nouveau_dmem *page_to_dmem(struct page *page) return container_of(page->pgmap, struct nouveau_dmem, pagemap); } -static unsigned long nouveau_dmem_page_addr(struct page *page) +unsigned long nouveau_dmem_page_addr(struct page *page) { struct nouveau_dmem_chunk *chunk = page->zone_device_data; unsigned long idx = page_to_pfn(page) - chunk->pfn_first; @@ -671,28 +671,3 @@ out_free_src: out: return ret; } - -void -nouveau_dmem_convert_pfn(struct nouveau_drm *drm, - struct hmm_range *range) -{ - unsigned long i, npages; - - npages = (range->end - range->start) >> PAGE_SHIFT; - for (i = 0; i < npages; ++i) { - struct page *page; - uint64_t addr; - - page = hmm_device_entry_to_page(range, range->pfns[i]); - if (page == NULL) - continue; - - if (!is_device_private_page(page)) - continue; - - addr = nouveau_dmem_page_addr(page); - range->pfns[i] &= ((1UL << range->pfn_shift) - 1); - range->pfns[i] |= (addr >> PAGE_SHIFT) << range->pfn_shift; - range->pfns[i] |= NVIF_VMM_PFNMAP_V0_VRAM; - } -} -- cgit v1.2.3 From e3d8b08904694e9ccae5163d0bb7d35fa66e5bdc Mon Sep 17 00:00:00 2001 From: Ralph Campbell Date: Tue, 3 Mar 2020 16:13:39 -0800 Subject: drm/nouveau/svm: map pages after migration MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit When memory is migrated to the GPU, it is likely to be accessed by GPU code soon afterwards. Instead of waiting for a GPU fault, map the migrated memory into the GPU page tables with the same access permissions as the source CPU page table entries. This preserves copy on write semantics. Signed-off-by: Ralph Campbell Cc: Christoph Hellwig Cc: Jason Gunthorpe Cc: "Jérôme Glisse" Cc: Ben Skeggs Signed-off-by: Ben Skeggs --- drivers/gpu/drm/nouveau/nouveau_dmem.c | 46 +++++++++++++++++--------- drivers/gpu/drm/nouveau/nouveau_dmem.h | 2 ++ drivers/gpu/drm/nouveau/nouveau_svm.c | 59 +++++++++++++++++++++++++++++++++- drivers/gpu/drm/nouveau/nouveau_svm.h | 5 +++ 4 files changed, 95 insertions(+), 17 deletions(-) (limited to 'drivers/gpu/drm/nouveau/nouveau_dmem.c') diff --git a/drivers/gpu/drm/nouveau/nouveau_dmem.c b/drivers/gpu/drm/nouveau/nouveau_dmem.c index ad89e09a0be3..a2ef8d562867 100644 --- a/drivers/gpu/drm/nouveau/nouveau_dmem.c +++ b/drivers/gpu/drm/nouveau/nouveau_dmem.c @@ -25,12 +25,14 @@ #include "nouveau_dma.h" #include "nouveau_mem.h" #include "nouveau_bo.h" +#include "nouveau_svm.h" #include #include #include #include #include +#include #include #include @@ -561,10 +563,11 @@ out_free: } static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm, - unsigned long src, dma_addr_t *dma_addr) + unsigned long src, dma_addr_t *dma_addr, u64 *pfn) { struct device *dev = drm->dev->dev; struct page *dpage, *spage; + unsigned long paddr; spage = migrate_pfn_to_page(src); if (!spage || !(src & MIGRATE_PFN_MIGRATE)) @@ -572,17 +575,21 @@ static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm, dpage = nouveau_dmem_page_alloc_locked(drm); if (!dpage) - return 0; + goto out; *dma_addr = dma_map_page(dev, spage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL); if (dma_mapping_error(dev, *dma_addr)) goto out_free_page; + paddr = nouveau_dmem_page_addr(dpage); if (drm->dmem->migrate.copy_func(drm, 1, NOUVEAU_APER_VRAM, - nouveau_dmem_page_addr(dpage), NOUVEAU_APER_HOST, - *dma_addr)) + paddr, NOUVEAU_APER_HOST, *dma_addr)) goto out_dma_unmap; + *pfn = NVIF_VMM_PFNMAP_V0_V | NVIF_VMM_PFNMAP_V0_VRAM | + ((paddr >> PAGE_SHIFT) << NVIF_VMM_PFNMAP_V0_ADDR_SHIFT); + if (src & MIGRATE_PFN_WRITE) + *pfn |= NVIF_VMM_PFNMAP_V0_W; return migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED; out_dma_unmap: @@ -590,18 +597,20 @@ out_dma_unmap: out_free_page: nouveau_dmem_page_free_locked(drm, dpage); out: + *pfn = NVIF_VMM_PFNMAP_V0_NONE; return 0; } static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm, - struct migrate_vma *args, dma_addr_t *dma_addrs) + struct nouveau_svmm *svmm, struct migrate_vma *args, + dma_addr_t *dma_addrs, u64 *pfns) { struct nouveau_fence *fence; unsigned long addr = args->start, nr_dma = 0, i; for (i = 0; addr < args->end; i++) { args->dst[i] = nouveau_dmem_migrate_copy_one(drm, args->src[i], - dma_addrs + nr_dma); + dma_addrs + nr_dma, pfns + i); if (args->dst[i]) nr_dma++; addr += PAGE_SIZE; @@ -610,20 +619,18 @@ static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm, nouveau_fence_new(drm->dmem->migrate.chan, false, &fence); migrate_vma_pages(args); nouveau_dmem_fence_done(&fence); + nouveau_pfns_map(svmm, args->vma->vm_mm, args->start, pfns, i); while (nr_dma--) { dma_unmap_page(drm->dev->dev, dma_addrs[nr_dma], PAGE_SIZE, DMA_BIDIRECTIONAL); } - /* - * FIXME optimization: update GPU page table to point to newly migrated - * memory. - */ migrate_vma_finalize(args); } int nouveau_dmem_migrate_vma(struct nouveau_drm *drm, + struct nouveau_svmm *svmm, struct vm_area_struct *vma, unsigned long start, unsigned long end) @@ -635,7 +642,8 @@ nouveau_dmem_migrate_vma(struct nouveau_drm *drm, .vma = vma, .start = start, }; - unsigned long c, i; + unsigned long i; + u64 *pfns; int ret = -ENOMEM; args.src = kcalloc(max, sizeof(*args.src), GFP_KERNEL); @@ -649,19 +657,25 @@ nouveau_dmem_migrate_vma(struct nouveau_drm *drm, if (!dma_addrs) goto out_free_dst; - for (i = 0; i < npages; i += c) { - c = min(SG_MAX_SINGLE_ALLOC, npages); - args.end = start + (c << PAGE_SHIFT); + pfns = nouveau_pfns_alloc(max); + if (!pfns) + goto out_free_dma; + + for (i = 0; i < npages; i += max) { + args.end = start + (max << PAGE_SHIFT); ret = migrate_vma_setup(&args); if (ret) - goto out_free_dma; + goto out_free_pfns; if (args.cpages) - nouveau_dmem_migrate_chunk(drm, &args, dma_addrs); + nouveau_dmem_migrate_chunk(drm, svmm, &args, dma_addrs, + pfns); args.start = args.end; } ret = 0; +out_free_pfns: + nouveau_pfns_free(pfns); out_free_dma: kfree(dma_addrs); out_free_dst: diff --git a/drivers/gpu/drm/nouveau/nouveau_dmem.h b/drivers/gpu/drm/nouveau/nouveau_dmem.h index 92394be5d649..3e03d9629a38 100644 --- a/drivers/gpu/drm/nouveau/nouveau_dmem.h +++ b/drivers/gpu/drm/nouveau/nouveau_dmem.h @@ -25,6 +25,7 @@ struct drm_device; struct drm_file; struct nouveau_drm; +struct nouveau_svmm; struct hmm_range; #if IS_ENABLED(CONFIG_DRM_NOUVEAU_SVM) @@ -34,6 +35,7 @@ void nouveau_dmem_suspend(struct nouveau_drm *); void nouveau_dmem_resume(struct nouveau_drm *); int nouveau_dmem_migrate_vma(struct nouveau_drm *drm, + struct nouveau_svmm *svmm, struct vm_area_struct *vma, unsigned long start, unsigned long end); diff --git a/drivers/gpu/drm/nouveau/nouveau_svm.c b/drivers/gpu/drm/nouveau/nouveau_svm.c index 645fedd77e21..fe89abf237a8 100644 --- a/drivers/gpu/drm/nouveau/nouveau_svm.c +++ b/drivers/gpu/drm/nouveau/nouveau_svm.c @@ -70,6 +70,12 @@ struct nouveau_svm { #define SVM_DBG(s,f,a...) NV_DEBUG((s)->drm, "svm: "f"\n", ##a) #define SVM_ERR(s,f,a...) NV_WARN((s)->drm, "svm: "f"\n", ##a) +struct nouveau_pfnmap_args { + struct nvif_ioctl_v0 i; + struct nvif_ioctl_mthd_v0 m; + struct nvif_vmm_pfnmap_v0 p; +}; + struct nouveau_ivmm { struct nouveau_svmm *svmm; u64 inst; @@ -187,7 +193,8 @@ nouveau_svmm_bind(struct drm_device *dev, void *data, addr = max(addr, vma->vm_start); next = min(vma->vm_end, end); /* This is a best effort so we ignore errors */ - nouveau_dmem_migrate_vma(cli->drm, vma, addr, next); + nouveau_dmem_migrate_vma(cli->drm, cli->svm.svmm, vma, addr, + next); addr = next; } @@ -784,6 +791,56 @@ nouveau_svm_fault(struct nvif_notify *notify) return NVIF_NOTIFY_KEEP; } +static struct nouveau_pfnmap_args * +nouveau_pfns_to_args(void *pfns) +{ + return container_of(pfns, struct nouveau_pfnmap_args, p.phys); +} + +u64 * +nouveau_pfns_alloc(unsigned long npages) +{ + struct nouveau_pfnmap_args *args; + + args = kzalloc(struct_size(args, p.phys, npages), GFP_KERNEL); + if (!args) + return NULL; + + args->i.type = NVIF_IOCTL_V0_MTHD; + args->m.method = NVIF_VMM_V0_PFNMAP; + args->p.page = PAGE_SHIFT; + + return args->p.phys; +} + +void +nouveau_pfns_free(u64 *pfns) +{ + struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns); + + kfree(args); +} + +void +nouveau_pfns_map(struct nouveau_svmm *svmm, struct mm_struct *mm, + unsigned long addr, u64 *pfns, unsigned long npages) +{ + struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns); + int ret; + + args->p.addr = addr; + args->p.size = npages << PAGE_SHIFT; + + mutex_lock(&svmm->mutex); + + svmm->vmm->vmm.object.client->super = true; + ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, sizeof(*args) + + npages * sizeof(args->p.phys[0]), NULL); + svmm->vmm->vmm.object.client->super = false; + + mutex_unlock(&svmm->mutex); +} + static void nouveau_svm_fault_buffer_fini(struct nouveau_svm *svm, int id) { diff --git a/drivers/gpu/drm/nouveau/nouveau_svm.h b/drivers/gpu/drm/nouveau/nouveau_svm.h index e839d8189461..f0fcd1b72e8b 100644 --- a/drivers/gpu/drm/nouveau/nouveau_svm.h +++ b/drivers/gpu/drm/nouveau/nouveau_svm.h @@ -18,6 +18,11 @@ void nouveau_svmm_fini(struct nouveau_svmm **); int nouveau_svmm_join(struct nouveau_svmm *, u64 inst); void nouveau_svmm_part(struct nouveau_svmm *, u64 inst); int nouveau_svmm_bind(struct drm_device *, void *, struct drm_file *); + +u64 *nouveau_pfns_alloc(unsigned long npages); +void nouveau_pfns_free(u64 *pfns); +void nouveau_pfns_map(struct nouveau_svmm *svmm, struct mm_struct *mm, + unsigned long addr, u64 *pfns, unsigned long npages); #else /* IS_ENABLED(CONFIG_DRM_NOUVEAU_SVM) */ static inline void nouveau_svm_init(struct nouveau_drm *drm) {} static inline void nouveau_svm_fini(struct nouveau_drm *drm) {} -- cgit v1.2.3 From 1d7f940c3a16623adb25706b10d81f8b822d53f9 Mon Sep 17 00:00:00 2001 From: Ralph Campbell Date: Tue, 21 Apr 2020 16:11:07 -0700 Subject: drm/nouveau/nouveau/hmm: fix nouveau_dmem_chunk allocations In nouveau_dmem_init(), a number of struct nouveau_dmem_chunk are allocated and put on the dmem->chunk_empty list. Then in nouveau_dmem_pages_alloc(), a nouveau_dmem_chunk is removed from the list and GPU memory is allocated. However, the nouveau_dmem_chunk is never removed from the chunk_empty list nor placed on the chunk_free or chunk_full lists. This results in only one chunk ever being actually used (2MB) and quickly leads to migration to device private memory failures. Fix this by having just one list of free device private pages and if no pages are free, allocate a chunk of device private pages and GPU memory. Signed-off-by: Ralph Campbell Signed-off-by: Ben Skeggs --- drivers/gpu/drm/nouveau/nouveau_dmem.c | 304 ++++++++++++--------------------- 1 file changed, 112 insertions(+), 192 deletions(-) (limited to 'drivers/gpu/drm/nouveau/nouveau_dmem.c') diff --git a/drivers/gpu/drm/nouveau/nouveau_dmem.c b/drivers/gpu/drm/nouveau/nouveau_dmem.c index a2ef8d562867..c39500a84b17 100644 --- a/drivers/gpu/drm/nouveau/nouveau_dmem.c +++ b/drivers/gpu/drm/nouveau/nouveau_dmem.c @@ -61,10 +61,8 @@ struct nouveau_dmem_chunk { struct list_head list; struct nouveau_bo *bo; struct nouveau_drm *drm; - unsigned long pfn_first; unsigned long callocated; - unsigned long bitmap[BITS_TO_LONGS(DMEM_CHUNK_NPAGES)]; - spinlock_t lock; + struct dev_pagemap pagemap; }; struct nouveau_dmem_migrate { @@ -74,48 +72,50 @@ struct nouveau_dmem_migrate { struct nouveau_dmem { struct nouveau_drm *drm; - struct dev_pagemap pagemap; struct nouveau_dmem_migrate migrate; - struct list_head chunk_free; - struct list_head chunk_full; - struct list_head chunk_empty; + struct list_head chunks; struct mutex mutex; + struct page *free_pages; + spinlock_t lock; }; -static inline struct nouveau_dmem *page_to_dmem(struct page *page) +static struct nouveau_dmem_chunk *nouveau_page_to_chunk(struct page *page) +{ + return container_of(page->pgmap, struct nouveau_dmem_chunk, pagemap); +} + +static struct nouveau_drm *page_to_drm(struct page *page) { - return container_of(page->pgmap, struct nouveau_dmem, pagemap); + struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page); + + return chunk->drm; } static unsigned long nouveau_dmem_page_addr(struct page *page) { - struct nouveau_dmem_chunk *chunk = page->zone_device_data; - unsigned long idx = page_to_pfn(page) - chunk->pfn_first; + struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page); + unsigned long off = (page_to_pfn(page) << PAGE_SHIFT) - + chunk->pagemap.res.start; - return (idx << PAGE_SHIFT) + chunk->bo->bo.offset; + return chunk->bo->bo.offset + off; } static void nouveau_dmem_page_free(struct page *page) { - struct nouveau_dmem_chunk *chunk = page->zone_device_data; - unsigned long idx = page_to_pfn(page) - chunk->pfn_first; + struct nouveau_dmem_chunk *chunk = nouveau_page_to_chunk(page); + struct nouveau_dmem *dmem = chunk->drm->dmem; + + spin_lock(&dmem->lock); + page->zone_device_data = dmem->free_pages; + dmem->free_pages = page; - /* - * FIXME: - * - * This is really a bad example, we need to overhaul nouveau memory - * management to be more page focus and allow lighter locking scheme - * to be use in the process. - */ - spin_lock(&chunk->lock); - clear_bit(idx, chunk->bitmap); WARN_ON(!chunk->callocated); chunk->callocated--; /* * FIXME when chunk->callocated reach 0 we should add the chunk to * a reclaim list so that it can be freed in case of memory pressure. */ - spin_unlock(&chunk->lock); + spin_unlock(&dmem->lock); } static void nouveau_dmem_fence_done(struct nouveau_fence **fence) @@ -167,8 +167,8 @@ error_free_page: static vm_fault_t nouveau_dmem_migrate_to_ram(struct vm_fault *vmf) { - struct nouveau_dmem *dmem = page_to_dmem(vmf->page); - struct nouveau_drm *drm = dmem->drm; + struct nouveau_drm *drm = page_to_drm(vmf->page); + struct nouveau_dmem *dmem = drm->dmem; struct nouveau_fence *fence; unsigned long src = 0, dst = 0; dma_addr_t dma_addr = 0; @@ -211,131 +211,105 @@ static const struct dev_pagemap_ops nouveau_dmem_pagemap_ops = { }; static int -nouveau_dmem_chunk_alloc(struct nouveau_drm *drm) +nouveau_dmem_chunk_alloc(struct nouveau_drm *drm, struct page **ppage) { struct nouveau_dmem_chunk *chunk; + struct resource *res; + struct page *page; + void *ptr; + unsigned long i, pfn_first; int ret; - if (drm->dmem == NULL) - return -EINVAL; - - mutex_lock(&drm->dmem->mutex); - chunk = list_first_entry_or_null(&drm->dmem->chunk_empty, - struct nouveau_dmem_chunk, - list); + chunk = kzalloc(sizeof(*chunk), GFP_KERNEL); if (chunk == NULL) { - mutex_unlock(&drm->dmem->mutex); - return -ENOMEM; + ret = -ENOMEM; + goto out; } - list_del(&chunk->list); - mutex_unlock(&drm->dmem->mutex); + /* Allocate unused physical address space for device private pages. */ + res = request_free_mem_region(&iomem_resource, DMEM_CHUNK_SIZE, + "nouveau_dmem"); + if (IS_ERR(res)) { + ret = PTR_ERR(res); + goto out_free; + } + + chunk->drm = drm; + chunk->pagemap.type = MEMORY_DEVICE_PRIVATE; + chunk->pagemap.res = *res; + chunk->pagemap.ops = &nouveau_dmem_pagemap_ops; + chunk->pagemap.owner = drm->dev; ret = nouveau_bo_new(&drm->client, DMEM_CHUNK_SIZE, 0, TTM_PL_FLAG_VRAM, 0, 0, NULL, NULL, &chunk->bo); if (ret) - goto out; + goto out_release; ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false); - if (ret) { - nouveau_bo_ref(NULL, &chunk->bo); - goto out; - } + if (ret) + goto out_bo_free; - bitmap_zero(chunk->bitmap, DMEM_CHUNK_NPAGES); - spin_lock_init(&chunk->lock); + ptr = memremap_pages(&chunk->pagemap, numa_node_id()); + if (IS_ERR(ptr)) { + ret = PTR_ERR(ptr); + goto out_bo_unpin; + } -out: mutex_lock(&drm->dmem->mutex); - if (chunk->bo) - list_add(&chunk->list, &drm->dmem->chunk_empty); - else - list_add_tail(&chunk->list, &drm->dmem->chunk_empty); + list_add(&chunk->list, &drm->dmem->chunks); mutex_unlock(&drm->dmem->mutex); - return ret; -} - -static struct nouveau_dmem_chunk * -nouveau_dmem_chunk_first_free_locked(struct nouveau_drm *drm) -{ - struct nouveau_dmem_chunk *chunk; - - chunk = list_first_entry_or_null(&drm->dmem->chunk_free, - struct nouveau_dmem_chunk, - list); - if (chunk) - return chunk; - - chunk = list_first_entry_or_null(&drm->dmem->chunk_empty, - struct nouveau_dmem_chunk, - list); - if (chunk->bo) - return chunk; - - return NULL; -} - -static int -nouveau_dmem_pages_alloc(struct nouveau_drm *drm, - unsigned long npages, - unsigned long *pages) -{ - struct nouveau_dmem_chunk *chunk; - unsigned long c; - int ret; - - memset(pages, 0xff, npages * sizeof(*pages)); - - mutex_lock(&drm->dmem->mutex); - for (c = 0; c < npages;) { - unsigned long i; - - chunk = nouveau_dmem_chunk_first_free_locked(drm); - if (chunk == NULL) { - mutex_unlock(&drm->dmem->mutex); - ret = nouveau_dmem_chunk_alloc(drm); - if (ret) { - if (c) - return 0; - return ret; - } - mutex_lock(&drm->dmem->mutex); - continue; - } - - spin_lock(&chunk->lock); - i = find_first_zero_bit(chunk->bitmap, DMEM_CHUNK_NPAGES); - while (i < DMEM_CHUNK_NPAGES && c < npages) { - pages[c] = chunk->pfn_first + i; - set_bit(i, chunk->bitmap); - chunk->callocated++; - c++; - - i = find_next_zero_bit(chunk->bitmap, - DMEM_CHUNK_NPAGES, i); - } - spin_unlock(&chunk->lock); + pfn_first = chunk->pagemap.res.start >> PAGE_SHIFT; + page = pfn_to_page(pfn_first); + spin_lock(&drm->dmem->lock); + for (i = 0; i < DMEM_CHUNK_NPAGES - 1; ++i, ++page) { + page->zone_device_data = drm->dmem->free_pages; + drm->dmem->free_pages = page; } - mutex_unlock(&drm->dmem->mutex); + *ppage = page; + chunk->callocated++; + spin_unlock(&drm->dmem->lock); + + NV_INFO(drm, "DMEM: registered %ldMB of device memory\n", + DMEM_CHUNK_SIZE >> 20); return 0; + +out_bo_unpin: + nouveau_bo_unpin(chunk->bo); +out_bo_free: + nouveau_bo_ref(NULL, &chunk->bo); +out_release: + release_mem_region(chunk->pagemap.res.start, + resource_size(&chunk->pagemap.res)); +out_free: + kfree(chunk); +out: + return ret; } static struct page * nouveau_dmem_page_alloc_locked(struct nouveau_drm *drm) { - unsigned long pfns[1]; - struct page *page; + struct nouveau_dmem_chunk *chunk; + struct page *page = NULL; int ret; - /* FIXME stop all the miss-match API ... */ - ret = nouveau_dmem_pages_alloc(drm, 1, pfns); - if (ret) - return NULL; + spin_lock(&drm->dmem->lock); + if (drm->dmem->free_pages) { + page = drm->dmem->free_pages; + drm->dmem->free_pages = page->zone_device_data; + chunk = nouveau_page_to_chunk(page); + chunk->callocated++; + spin_unlock(&drm->dmem->lock); + } else { + spin_unlock(&drm->dmem->lock); + ret = nouveau_dmem_chunk_alloc(drm, &page); + if (ret) + return NULL; + } - page = pfn_to_page(pfns[0]); get_page(page); lock_page(page); return page; @@ -358,12 +332,7 @@ nouveau_dmem_resume(struct nouveau_drm *drm) return; mutex_lock(&drm->dmem->mutex); - list_for_each_entry (chunk, &drm->dmem->chunk_free, list) { - ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false); - /* FIXME handle pin failure */ - WARN_ON(ret); - } - list_for_each_entry (chunk, &drm->dmem->chunk_full, list) { + list_for_each_entry(chunk, &drm->dmem->chunks, list) { ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false); /* FIXME handle pin failure */ WARN_ON(ret); @@ -380,12 +349,8 @@ nouveau_dmem_suspend(struct nouveau_drm *drm) return; mutex_lock(&drm->dmem->mutex); - list_for_each_entry (chunk, &drm->dmem->chunk_free, list) { - nouveau_bo_unpin(chunk->bo); - } - list_for_each_entry (chunk, &drm->dmem->chunk_full, list) { + list_for_each_entry(chunk, &drm->dmem->chunks, list) nouveau_bo_unpin(chunk->bo); - } mutex_unlock(&drm->dmem->mutex); } @@ -399,15 +364,13 @@ nouveau_dmem_fini(struct nouveau_drm *drm) mutex_lock(&drm->dmem->mutex); - WARN_ON(!list_empty(&drm->dmem->chunk_free)); - WARN_ON(!list_empty(&drm->dmem->chunk_full)); - - list_for_each_entry_safe (chunk, tmp, &drm->dmem->chunk_empty, list) { - if (chunk->bo) { - nouveau_bo_unpin(chunk->bo); - nouveau_bo_ref(NULL, &chunk->bo); - } + list_for_each_entry_safe(chunk, tmp, &drm->dmem->chunks, list) { + nouveau_bo_unpin(chunk->bo); + nouveau_bo_ref(NULL, &chunk->bo); list_del(&chunk->list); + memunmap_pages(&chunk->pagemap); + release_mem_region(chunk->pagemap.res.start, + resource_size(&chunk->pagemap.res)); kfree(chunk); } @@ -493,9 +456,6 @@ nouveau_dmem_migrate_init(struct nouveau_drm *drm) void nouveau_dmem_init(struct nouveau_drm *drm) { - struct device *device = drm->dev->dev; - struct resource *res; - unsigned long i, size, pfn_first; int ret; /* This only make sense on PASCAL or newer */ @@ -507,59 +467,16 @@ nouveau_dmem_init(struct nouveau_drm *drm) drm->dmem->drm = drm; mutex_init(&drm->dmem->mutex); - INIT_LIST_HEAD(&drm->dmem->chunk_free); - INIT_LIST_HEAD(&drm->dmem->chunk_full); - INIT_LIST_HEAD(&drm->dmem->chunk_empty); - - size = ALIGN(drm->client.device.info.ram_user, DMEM_CHUNK_SIZE); + INIT_LIST_HEAD(&drm->dmem->chunks); + mutex_init(&drm->dmem->mutex); + spin_lock_init(&drm->dmem->lock); /* Initialize migration dma helpers before registering memory */ ret = nouveau_dmem_migrate_init(drm); - if (ret) - goto out_free; - - /* - * FIXME we need some kind of policy to decide how much VRAM we - * want to register with HMM. For now just register everything - * and latter if we want to do thing like over commit then we - * could revisit this. - */ - res = devm_request_free_mem_region(device, &iomem_resource, size); - if (IS_ERR(res)) - goto out_free; - drm->dmem->pagemap.type = MEMORY_DEVICE_PRIVATE; - drm->dmem->pagemap.res = *res; - drm->dmem->pagemap.ops = &nouveau_dmem_pagemap_ops; - drm->dmem->pagemap.owner = drm->dev; - if (IS_ERR(devm_memremap_pages(device, &drm->dmem->pagemap))) - goto out_free; - - pfn_first = res->start >> PAGE_SHIFT; - for (i = 0; i < (size / DMEM_CHUNK_SIZE); ++i) { - struct nouveau_dmem_chunk *chunk; - struct page *page; - unsigned long j; - - chunk = kzalloc(sizeof(*chunk), GFP_KERNEL); - if (chunk == NULL) { - nouveau_dmem_fini(drm); - return; - } - - chunk->drm = drm; - chunk->pfn_first = pfn_first + (i * DMEM_CHUNK_NPAGES); - list_add_tail(&chunk->list, &drm->dmem->chunk_empty); - - page = pfn_to_page(chunk->pfn_first); - for (j = 0; j < DMEM_CHUNK_NPAGES; ++j, ++page) - page->zone_device_data = chunk; + if (ret) { + kfree(drm->dmem); + drm->dmem = NULL; } - - NV_INFO(drm, "DMEM: registered %ldMB of device memory\n", size >> 20); - return; -out_free: - kfree(drm->dmem); - drm->dmem = NULL; } static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm, @@ -646,6 +563,9 @@ nouveau_dmem_migrate_vma(struct nouveau_drm *drm, u64 *pfns; int ret = -ENOMEM; + if (drm->dmem == NULL) + return -ENODEV; + args.src = kcalloc(max, sizeof(*args.src), GFP_KERNEL); if (!args.src) goto out; -- cgit v1.2.3 From 9d4296a7d4b329016e6b5b59a7fb3f8afa36ac05 Mon Sep 17 00:00:00 2001 From: Ralph Campbell Date: Wed, 20 May 2020 11:36:52 -0700 Subject: drm/nouveau/nouveau/hmm: fix migrate zero page to GPU When calling OpenCL clEnqueueSVMMigrateMem() on a region of memory that is backed by pte_none() or zero pages, migrate_vma_setup() will fill the source PFN array with an entry indicating the source page is zero. Use this to optimize migration to device private memory by allocating GPU memory and zero filling it instead of failing to migrate the page. Signed-off-by: Ralph Campbell Signed-off-by: Ben Skeggs --- drivers/gpu/drm/nouveau/nouveau_dmem.c | 75 ++++++++++++++++++++++++++++++---- 1 file changed, 66 insertions(+), 9 deletions(-) (limited to 'drivers/gpu/drm/nouveau/nouveau_dmem.c') diff --git a/drivers/gpu/drm/nouveau/nouveau_dmem.c b/drivers/gpu/drm/nouveau/nouveau_dmem.c index c39500a84b17..339a0c387eae 100644 --- a/drivers/gpu/drm/nouveau/nouveau_dmem.c +++ b/drivers/gpu/drm/nouveau/nouveau_dmem.c @@ -56,6 +56,8 @@ enum nouveau_aper { typedef int (*nouveau_migrate_copy_t)(struct nouveau_drm *drm, u64 npages, enum nouveau_aper, u64 dst_addr, enum nouveau_aper, u64 src_addr); +typedef int (*nouveau_clear_page_t)(struct nouveau_drm *drm, u32 length, + enum nouveau_aper, u64 dst_addr); struct nouveau_dmem_chunk { struct list_head list; @@ -67,6 +69,7 @@ struct nouveau_dmem_chunk { struct nouveau_dmem_migrate { nouveau_migrate_copy_t copy_func; + nouveau_clear_page_t clear_func; struct nouveau_channel *chan; }; @@ -436,6 +439,52 @@ nvc0b5_migrate_copy(struct nouveau_drm *drm, u64 npages, return 0; } +static int +nvc0b5_migrate_clear(struct nouveau_drm *drm, u32 length, + enum nouveau_aper dst_aper, u64 dst_addr) +{ + struct nouveau_channel *chan = drm->dmem->migrate.chan; + u32 launch_dma = (1 << 10) /* REMAP_ENABLE_TRUE */ | + (1 << 8) /* DST_MEMORY_LAYOUT_PITCH. */ | + (1 << 7) /* SRC_MEMORY_LAYOUT_PITCH. */ | + (1 << 2) /* FLUSH_ENABLE_TRUE. */ | + (2 << 0) /* DATA_TRANSFER_TYPE_NON_PIPELINED. */; + u32 remap = (4 << 0) /* DST_X_CONST_A */ | + (5 << 4) /* DST_Y_CONST_B */ | + (3 << 16) /* COMPONENT_SIZE_FOUR */ | + (1 << 24) /* NUM_DST_COMPONENTS_TWO */; + int ret; + + ret = RING_SPACE(chan, 12); + if (ret) + return ret; + + switch (dst_aper) { + case NOUVEAU_APER_VRAM: + BEGIN_IMC0(chan, NvSubCopy, 0x0264, 0); + break; + case NOUVEAU_APER_HOST: + BEGIN_IMC0(chan, NvSubCopy, 0x0264, 1); + break; + default: + return -EINVAL; + } + launch_dma |= 0x00002000; /* DST_TYPE_PHYSICAL. */ + + BEGIN_NVC0(chan, NvSubCopy, 0x0700, 3); + OUT_RING(chan, 0); + OUT_RING(chan, 0); + OUT_RING(chan, remap); + BEGIN_NVC0(chan, NvSubCopy, 0x0408, 2); + OUT_RING(chan, upper_32_bits(dst_addr)); + OUT_RING(chan, lower_32_bits(dst_addr)); + BEGIN_NVC0(chan, NvSubCopy, 0x0418, 1); + OUT_RING(chan, length >> 3); + BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1); + OUT_RING(chan, launch_dma); + return 0; +} + static int nouveau_dmem_migrate_init(struct nouveau_drm *drm) { @@ -445,6 +494,7 @@ nouveau_dmem_migrate_init(struct nouveau_drm *drm) case VOLTA_DMA_COPY_A: case TURING_DMA_COPY_A: drm->dmem->migrate.copy_func = nvc0b5_migrate_copy; + drm->dmem->migrate.clear_func = nvc0b5_migrate_clear; drm->dmem->migrate.chan = drm->ttm.chan; return 0; default: @@ -487,21 +537,28 @@ static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm, unsigned long paddr; spage = migrate_pfn_to_page(src); - if (!spage || !(src & MIGRATE_PFN_MIGRATE)) + if (!(src & MIGRATE_PFN_MIGRATE)) goto out; dpage = nouveau_dmem_page_alloc_locked(drm); if (!dpage) goto out; - *dma_addr = dma_map_page(dev, spage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL); - if (dma_mapping_error(dev, *dma_addr)) - goto out_free_page; - paddr = nouveau_dmem_page_addr(dpage); - if (drm->dmem->migrate.copy_func(drm, 1, NOUVEAU_APER_VRAM, - paddr, NOUVEAU_APER_HOST, *dma_addr)) - goto out_dma_unmap; + if (spage) { + *dma_addr = dma_map_page(dev, spage, 0, page_size(spage), + DMA_BIDIRECTIONAL); + if (dma_mapping_error(dev, *dma_addr)) + goto out_free_page; + if (drm->dmem->migrate.copy_func(drm, page_size(spage), + NOUVEAU_APER_VRAM, paddr, NOUVEAU_APER_HOST, *dma_addr)) + goto out_dma_unmap; + } else { + *dma_addr = DMA_MAPPING_ERROR; + if (drm->dmem->migrate.clear_func(drm, page_size(dpage), + NOUVEAU_APER_VRAM, paddr)) + goto out_free_page; + } *pfn = NVIF_VMM_PFNMAP_V0_V | NVIF_VMM_PFNMAP_V0_VRAM | ((paddr >> PAGE_SHIFT) << NVIF_VMM_PFNMAP_V0_ADDR_SHIFT); @@ -528,7 +585,7 @@ static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm, for (i = 0; addr < args->end; i++) { args->dst[i] = nouveau_dmem_migrate_copy_one(drm, args->src[i], dma_addrs + nr_dma, pfns + i); - if (args->dst[i]) + if (!dma_mapping_error(drm->dev->dev, dma_addrs[nr_dma])) nr_dma++; addr += PAGE_SIZE; } -- cgit v1.2.3