hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. This series adds a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available. Nouveau and the HMM tests are updated to use the new flag.
Note that this series depends on a patch queued in Ben Skeggs' nouveau tree ("nouveau/hmm: map pages after migration") and the patches queued in Jason's HMM tree. There is also a patch outstanding ("nouveau/hmm: fix nouveau_dmem_chunk allocations") that is independent of the above and could be applied before or after.
Ralph Campbell (6): nouveau/hmm: map pages after migration nouveau: make nvkm_vmm_ctor() and nvkm_mmu_ptp_get() static nouveau/hmm: fault one page at a time mm/hmm: add output flag for compound page mapping nouveau/hmm: support mapping large sysmem pages hmm: add tests for HMM_PFN_COMPOUND flag
drivers/gpu/drm/nouveau/nouveau_dmem.c | 46 ++- drivers/gpu/drm/nouveau/nouveau_dmem.h | 2 + drivers/gpu/drm/nouveau/nouveau_svm.c | 288 +++++++++--------- drivers/gpu/drm/nouveau/nouveau_svm.h | 5 + .../gpu/drm/nouveau/nvkm/subdev/mmu/base.c | 6 +- .../gpu/drm/nouveau/nvkm/subdev/mmu/priv.h | 2 + drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c | 12 +- drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h | 3 - .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c | 29 +- include/linux/hmm.h | 4 +- lib/test_hmm.c | 2 + lib/test_hmm_uapi.h | 2 + mm/hmm.c | 10 +- tools/testing/selftests/vm/hmm-tests.c | 76 +++++ 14 files changed, 311 insertions(+), 176 deletions(-)
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 rcampbell@nvidia.com Cc: Christoph Hellwig hch@lst.de Cc: Jason Gunthorpe jgg@mellanox.com Cc: "Jérôme Glisse" jglisse@redhat.com Cc: Ben Skeggs bskeggs@redhat.com --- 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(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_dmem.c b/drivers/gpu/drm/nouveau/nouveau_dmem.c index 3364904eccff..b29b01a56d07 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 <nvif/class.h> #include <nvif/object.h> #include <nvif/if000c.h> #include <nvif/if500b.h> #include <nvif/if900b.h> +#include <nvif/if000c.h>
#include <linux/sched/mm.h> #include <linux/hmm.h> @@ -561,10 +563,11 @@ nouveau_dmem_init(struct nouveau_drm *drm) }
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 @@ static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm, 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 db3b59b210af..64da5d3635c8 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 407e34a5c0ab..22f054f7ee3e 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; }
@@ -814,6 +821,56 @@ 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) {}
The functions nvkm_vmm_ctor() and nvkm_mmu_ptp_get() are not called outside of the file defining them so make them static. Also, remove a useless semicolon after a {} statement block.
Signed-off-by: Ralph Campbell rcampbell@nvidia.com --- drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c | 2 +- drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c | 4 ++-- drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h | 3 --- 3 files changed, 3 insertions(+), 6 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c index ee11ccaf0563..de91e9a26172 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c @@ -61,7 +61,7 @@ nvkm_mmu_ptp_put(struct nvkm_mmu *mmu, bool force, struct nvkm_mmu_pt *pt) kfree(pt); }
-struct nvkm_mmu_pt * +static struct nvkm_mmu_pt * nvkm_mmu_ptp_get(struct nvkm_mmu *mmu, u32 size, bool zero) { struct nvkm_mmu_pt *pt; diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c index 41640e0584ac..67b00dcef4b8 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c @@ -580,7 +580,7 @@ nvkm_vmm_iter(struct nvkm_vmm *vmm, const struct nvkm_vmm_page *page, it.pte[it.lvl]++; } } - }; + }
nvkm_vmm_flush(&it); return ~0ULL; @@ -1030,7 +1030,7 @@ nvkm_vmm_ctor_managed(struct nvkm_vmm *vmm, u64 addr, u64 size) return 0; }
-int +static int nvkm_vmm_ctor(const struct nvkm_vmm_func *func, struct nvkm_mmu *mmu, u32 pd_header, bool managed, u64 addr, u64 size, struct lock_class_key *key, const char *name, diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h index 5e55ecbd8005..fd722bdc4c78 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h @@ -163,9 +163,6 @@ int nvkm_vmm_new_(const struct nvkm_vmm_func *, struct nvkm_mmu *, u32 pd_header, bool managed, u64 addr, u64 size, struct lock_class_key *, const char *name, struct nvkm_vmm **); -int nvkm_vmm_ctor(const struct nvkm_vmm_func *, struct nvkm_mmu *, - u32 pd_header, bool managed, u64 addr, u64 size, - struct lock_class_key *, const char *name, struct nvkm_vmm *); struct nvkm_vma *nvkm_vmm_node_search(struct nvkm_vmm *, u64 addr); struct nvkm_vma *nvkm_vmm_node_split(struct nvkm_vmm *, struct nvkm_vma *, u64 addr, u64 size);
The SVM page fault handler groups faults into a range of contiguous virtual addresses and requests hmm_range_fault() to populate and return the page frame number of system memory mapped by the CPU. In preparation for supporting large pages to be mapped by the GPU, process faults one page at a time. In addition, use the hmm_range default_flags to fix a corner case where the input hmm_pfns array is not reinitialized after hmm_range_fault() returns -EBUSY and must be called again.
Signed-off-by: Ralph Campbell rcampbell@nvidia.com --- drivers/gpu/drm/nouveau/nouveau_svm.c | 198 +++++++++----------------- 1 file changed, 65 insertions(+), 133 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_svm.c b/drivers/gpu/drm/nouveau/nouveau_svm.c index 22f054f7ee3e..a87b9347d6ce 100644 --- a/drivers/gpu/drm/nouveau/nouveau_svm.c +++ b/drivers/gpu/drm/nouveau/nouveau_svm.c @@ -516,7 +516,7 @@ static const struct mmu_interval_notifier_ops nouveau_svm_mni_ops = { static void nouveau_hmm_convert_pfn(struct nouveau_drm *drm, struct hmm_range *range, u64 *ioctl_addr) { - unsigned long i, npages; + struct page *page;
/* * The ioctl_addr prepared here is passed through nvif_object_ioctl() @@ -525,42 +525,38 @@ static void nouveau_hmm_convert_pfn(struct nouveau_drm *drm, * This is all just encoding the internal hmm representation into a * different nouveau internal representation. */ - npages = (range->end - range->start) >> PAGE_SHIFT; - for (i = 0; i < npages; ++i) { - struct page *page; - - if (!(range->hmm_pfns[i] & HMM_PFN_VALID)) { - ioctl_addr[i] = 0; - continue; - } - - page = hmm_pfn_to_page(range->hmm_pfns[i]); - if (is_device_private_page(page)) - ioctl_addr[i] = nouveau_dmem_page_addr(page) | - NVIF_VMM_PFNMAP_V0_V | - NVIF_VMM_PFNMAP_V0_VRAM; - else - ioctl_addr[i] = page_to_phys(page) | - NVIF_VMM_PFNMAP_V0_V | - NVIF_VMM_PFNMAP_V0_HOST; - if (range->hmm_pfns[i] & HMM_PFN_WRITE) - ioctl_addr[i] |= NVIF_VMM_PFNMAP_V0_W; + if (!(range->hmm_pfns[0] & HMM_PFN_VALID)) { + ioctl_addr[0] = 0; + return; } + + page = hmm_pfn_to_page(range->hmm_pfns[0]); + if (is_device_private_page(page)) + ioctl_addr[0] = nouveau_dmem_page_addr(page) | + NVIF_VMM_PFNMAP_V0_V | + NVIF_VMM_PFNMAP_V0_VRAM; + else + ioctl_addr[0] = page_to_phys(page) | + NVIF_VMM_PFNMAP_V0_V | + NVIF_VMM_PFNMAP_V0_HOST; + if (range->hmm_pfns[0] & HMM_PFN_WRITE) + ioctl_addr[0] |= NVIF_VMM_PFNMAP_V0_W; }
static int nouveau_range_fault(struct nouveau_svmm *svmm, struct nouveau_drm *drm, void *data, u32 size, - unsigned long hmm_pfns[], u64 *ioctl_addr, + u64 *ioctl_addr, unsigned long hmm_flags, struct svm_notifier *notifier) { unsigned long timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT); /* Have HMM fault pages within the fault window to the GPU. */ + unsigned long hmm_pfns[1]; struct hmm_range range = { .notifier = ¬ifier->notifier, .start = notifier->notifier.interval_tree.start, .end = notifier->notifier.interval_tree.last + 1, - .pfn_flags_mask = HMM_PFN_REQ_FAULT | HMM_PFN_REQ_WRITE, + .default_flags = hmm_flags, .hmm_pfns = hmm_pfns, }; struct mm_struct *mm = notifier->notifier.mm; @@ -575,11 +571,6 @@ static int nouveau_range_fault(struct nouveau_svmm *svmm, ret = hmm_range_fault(&range); up_read(&mm->mmap_sem); if (ret) { - /* - * FIXME: the input PFN_REQ flags are destroyed on - * -EBUSY, we need to regenerate them, also for the - * other continue below - */ if (ret == -EBUSY) continue; return ret; @@ -614,17 +605,12 @@ nouveau_svm_fault(struct nvif_notify *notify) struct nvif_object *device = &svm->drm->client.device.object; struct nouveau_svmm *svmm; struct { - struct { - struct nvif_ioctl_v0 i; - struct nvif_ioctl_mthd_v0 m; - struct nvif_vmm_pfnmap_v0 p; - } i; - u64 phys[16]; + struct nouveau_pfnmap_args i; + u64 phys[1]; } args; - unsigned long hmm_pfns[ARRAY_SIZE(args.phys)]; - struct vm_area_struct *vma; + unsigned long hmm_flags; u64 inst, start, limit; - int fi, fn, pi, fill; + int fi, fn; int replay = 0, ret;
/* Parse available fault buffer entries into a cache, and update @@ -691,66 +677,52 @@ nouveau_svm_fault(struct nvif_notify *notify) * window into a single update. */ start = buffer->fault[fi]->addr; - limit = start + (ARRAY_SIZE(args.phys) << PAGE_SHIFT); + limit = start + PAGE_SIZE; if (start < svmm->unmanaged.limit) limit = min_t(u64, limit, svmm->unmanaged.start); - SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit); - - mm = svmm->notifier.mm; - if (!mmget_not_zero(mm)) { - nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]); - continue; - }
- /* Intersect fault window with the CPU VMA, cancelling - * the fault if the address is invalid. + /* Prepare the GPU-side update of all pages within the + * fault window, determining required pages and access + * permissions based on pending faults. */ - down_read(&mm->mmap_sem); - vma = find_vma_intersection(mm, start, limit); - if (!vma) { - SVMM_ERR(svmm, "wndw %016llx-%016llx", start, limit); - up_read(&mm->mmap_sem); - mmput(mm); - nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]); - continue; + args.i.p.addr = start; + args.i.p.page = PAGE_SHIFT; + args.i.p.size = PAGE_SIZE; + /* + * Determine required permissions based on GPU fault + * access flags. + * XXX: atomic? + */ + switch (buffer->fault[fi]->access) { + case 0: /* READ. */ + hmm_flags = HMM_PFN_REQ_FAULT; + break; + case 3: /* PREFETCH. */ + hmm_flags = 0; + break; + default: + hmm_flags = HMM_PFN_REQ_FAULT | HMM_PFN_REQ_WRITE; + break; } - start = max_t(u64, start, vma->vm_start); - limit = min_t(u64, limit, vma->vm_end); - up_read(&mm->mmap_sem); - SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit);
- if (buffer->fault[fi]->addr != start) { - SVMM_ERR(svmm, "addr %016llx", buffer->fault[fi]->addr); - mmput(mm); + mm = svmm->notifier.mm; + if (!mmget_not_zero(mm)) { nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]); continue; }
- /* Prepare the GPU-side update of all pages within the - * fault window, determining required pages and access - * permissions based on pending faults. - */ - args.i.p.page = PAGE_SHIFT; - args.i.p.addr = start; - for (fn = fi, pi = 0;;) { - /* Determine required permissions based on GPU fault - * access flags. - *XXX: atomic? - */ - switch (buffer->fault[fn]->access) { - case 0: /* READ. */ - hmm_pfns[pi++] = HMM_PFN_REQ_FAULT; - break; - case 3: /* PREFETCH. */ - hmm_pfns[pi++] = 0; - break; - default: - hmm_pfns[pi++] = HMM_PFN_REQ_FAULT | - HMM_PFN_REQ_WRITE; - break; - } - args.i.p.size = pi << PAGE_SHIFT; + notifier.svmm = svmm; + ret = mmu_interval_notifier_insert(¬ifier.notifier, mm, + args.i.p.addr, args.i.p.size, + &nouveau_svm_mni_ops); + if (!ret) { + ret = nouveau_range_fault(svmm, svm->drm, &args, + sizeof(args), args.phys, hmm_flags, ¬ifier); + mmu_interval_notifier_remove(¬ifier.notifier); + } + mmput(mm);
+ for (fn = fi; ++fn < buffer->fault_nr; ) { /* It's okay to skip over duplicate addresses from the * same SVMM as faults are ordered by access type such * that only the first one needs to be handled. @@ -758,61 +730,21 @@ nouveau_svm_fault(struct nvif_notify *notify) * ie. WRITE faults appear first, thus any handling of * pending READ faults will already be satisfied. */ - while (++fn < buffer->fault_nr && - buffer->fault[fn]->svmm == svmm && - buffer->fault[fn ]->addr == - buffer->fault[fn - 1]->addr); - - /* If the next fault is outside the window, or all GPU - * faults have been dealt with, we're done here. - */ - if (fn >= buffer->fault_nr || - buffer->fault[fn]->svmm != svmm || + if (buffer->fault[fn]->svmm != svmm || buffer->fault[fn]->addr >= limit) break; - - /* Fill in the gap between this fault and the next. */ - fill = (buffer->fault[fn ]->addr - - buffer->fault[fn - 1]->addr) >> PAGE_SHIFT; - while (--fill) - hmm_pfns[pi++] = 0; }
- SVMM_DBG(svmm, "wndw %016llx-%016llx covering %d fault(s)", - args.i.p.addr, - args.i.p.addr + args.i.p.size, fn - fi); - - notifier.svmm = svmm; - ret = mmu_interval_notifier_insert(¬ifier.notifier, - svmm->notifier.mm, - args.i.p.addr, args.i.p.size, - &nouveau_svm_mni_ops); - if (!ret) { - ret = nouveau_range_fault( - svmm, svm->drm, &args, - sizeof(args.i) + pi * sizeof(args.phys[0]), - hmm_pfns, args.phys, ¬ifier); - mmu_interval_notifier_remove(¬ifier.notifier); - } - mmput(mm); + /* If handling failed completely, cancel all faults. */ + if (ret) { + while (fi < fn) { + struct nouveau_svm_fault *fault = + buffer->fault[fi++];
- /* Cancel any faults in the window whose pages didn't manage - * to keep their valid bit, or stay writeable when required. - * - * If handling failed completely, cancel all faults. - */ - while (fi < fn) { - struct nouveau_svm_fault *fault = buffer->fault[fi++]; - pi = (fault->addr - args.i.p.addr) >> PAGE_SHIFT; - if (ret || - !(args.phys[pi] & NVIF_VMM_PFNMAP_V0_V) || - (!(args.phys[pi] & NVIF_VMM_PFNMAP_V0_W) && - fault->access != 0 && fault->access != 3)) { nouveau_svm_fault_cancel_fault(svm, fault); - continue; } + } else replay++; - } }
/* Issue fault replay to the GPU. */
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. Add a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available.
Signed-off-by: Ralph Campbell rcampbell@nvidia.com --- include/linux/hmm.h | 4 +++- mm/hmm.c | 10 +++++++--- 2 files changed, 10 insertions(+), 4 deletions(-)
diff --git a/include/linux/hmm.h b/include/linux/hmm.h index e912b9dc4633..f2d38af421e7 100644 --- a/include/linux/hmm.h +++ b/include/linux/hmm.h @@ -41,12 +41,14 @@ enum hmm_pfn_flags { HMM_PFN_VALID = 1UL << (BITS_PER_LONG - 1), HMM_PFN_WRITE = 1UL << (BITS_PER_LONG - 2), HMM_PFN_ERROR = 1UL << (BITS_PER_LONG - 3), + HMM_PFN_COMPOUND = 1UL << (BITS_PER_LONG - 4),
/* Input flags */ HMM_PFN_REQ_FAULT = HMM_PFN_VALID, HMM_PFN_REQ_WRITE = HMM_PFN_WRITE,
- HMM_PFN_FLAGS = HMM_PFN_VALID | HMM_PFN_WRITE | HMM_PFN_ERROR, + HMM_PFN_FLAGS = HMM_PFN_VALID | HMM_PFN_WRITE | HMM_PFN_ERROR | + HMM_PFN_COMPOUND, };
/* diff --git a/mm/hmm.c b/mm/hmm.c index 41673a6d8d46..a9dd06e190a1 100644 --- a/mm/hmm.c +++ b/mm/hmm.c @@ -170,7 +170,9 @@ static inline unsigned long pmd_to_hmm_pfn_flags(struct hmm_range *range, { if (pmd_protnone(pmd)) return 0; - return pmd_write(pmd) ? (HMM_PFN_VALID | HMM_PFN_WRITE) : HMM_PFN_VALID; + return pmd_write(pmd) ? + (HMM_PFN_VALID | HMM_PFN_COMPOUND | HMM_PFN_WRITE) : + (HMM_PFN_VALID | HMM_PFN_COMPOUND); }
#ifdef CONFIG_TRANSPARENT_HUGEPAGE @@ -389,7 +391,9 @@ static inline unsigned long pud_to_hmm_pfn_flags(struct hmm_range *range, { if (!pud_present(pud)) return 0; - return pud_write(pud) ? (HMM_PFN_VALID | HMM_PFN_WRITE) : HMM_PFN_VALID; + return pud_write(pud) ? + (HMM_PFN_VALID | HMM_PFN_COMPOUND | HMM_PFN_WRITE) : + (HMM_PFN_VALID | HMM_PFN_COMPOUND); }
static int hmm_vma_walk_pud(pud_t *pudp, unsigned long start, unsigned long end, @@ -484,7 +488,7 @@ static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
pfn = pte_pfn(entry) + ((start & ~hmask) >> PAGE_SHIFT); for (; addr < end; addr += PAGE_SIZE, i++, pfn++) - range->hmm_pfns[i] = pfn | cpu_flags; + range->hmm_pfns[i] = pfn | cpu_flags | HMM_PFN_COMPOUND;
spin_unlock(ptl); return 0;
On Fri, May 08, 2020 at 12:20:07PM -0700, Ralph Campbell wrote:
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. Add a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available.
Why do you need the flag? The caller should be able to just use page_size() (or willys new thp_size helper).
On 5/8/20 12:51 PM, Christoph Hellwig wrote:
On Fri, May 08, 2020 at 12:20:07PM -0700, Ralph Campbell wrote:
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. Add a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available.
Why do you need the flag? The caller should be able to just use page_size() (or willys new thp_size helper).
The question is whether or not a large page can be mapped with smaller page table entries with different permissions. If one process has a 2MB page mapped with 4K PTEs with different read/write permissions, I don't think it would be OK for a device to map the whole 2MB with write access enabled. The flag is supposed to indicate that the whole page can be mapped by the device with the indicated read/write permissions.
On 8 May 2020, at 16:06, Ralph Campbell wrote:
On 5/8/20 12:51 PM, Christoph Hellwig wrote:
On Fri, May 08, 2020 at 12:20:07PM -0700, Ralph Campbell wrote:
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. Add a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available.
Why do you need the flag? The caller should be able to just use page_size() (or willys new thp_size helper).
The question is whether or not a large page can be mapped with smaller page table entries with different permissions. If one process has a 2MB page mapped with 4K PTEs with different read/write permissions, I don't think it would be OK for a device to map the whole 2MB with write access enabled. The flag is supposed to indicate that the whole page can be mapped by the device with the indicated read/write permissions.
If hmm_range_fault() only walks one VMA at a time, you would not have this permission issue, right? Since all pages from one VMA should have the same permission. But it seems that hmm_range_fault() deals with pages across multiple VMAs. Maybe we should make hmm_range_fault() bail out early when it encounters a VMA with a different permission than the existing ones?
— Best Regards, Yan Zi
On 5/26/20 3:29 PM, Zi Yan wrote:
On 8 May 2020, at 16:06, Ralph Campbell wrote:
On 5/8/20 12:51 PM, Christoph Hellwig wrote:
On Fri, May 08, 2020 at 12:20:07PM -0700, Ralph Campbell wrote:
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. Add a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available.
Why do you need the flag? The caller should be able to just use page_size() (or willys new thp_size helper).
The question is whether or not a large page can be mapped with smaller page table entries with different permissions. If one process has a 2MB page mapped with 4K PTEs with different read/write permissions, I don't think it would be OK for a device to map the whole 2MB with write access enabled. The flag is supposed to indicate that the whole page can be mapped by the device with the indicated read/write permissions.
If hmm_range_fault() only walks one VMA at a time, you would not have this permission issue, right? Since all pages from one VMA should have the same permission. But it seems that hmm_range_fault() deals with pages across multiple VMAs. Maybe we should make hmm_range_fault() bail out early when it encounters a VMA with a different permission than the existing ones?
— Best Regards, Yan Zi
I don't think so. The VMA might have read/write permission but the page table might have read-only permission in order to trigger a fault for copy-on-write. Or the PTE might be read-only or invalid to trigger faults for architectures that don't have hardware updated accessed bits and are using the minor faults to update LRU. The goal is that the MM core see the same faults whether the HMM device accesses memory or a CPU thread.
Nouveau currently only supports mapping PAGE_SIZE sized pages of system memory when shared virtual memory (SVM) is enabled. Use the new HMM_PFN_COMPOUND flag that hmm_range_fault() returns to support mapping system memory pages larger than PAGE_SIZE.
Signed-off-by: Ralph Campbell rcampbell@nvidia.com --- drivers/gpu/drm/nouveau/nouveau_svm.c | 47 ++++++++++++++----- .../gpu/drm/nouveau/nvkm/subdev/mmu/base.c | 4 ++ .../gpu/drm/nouveau/nvkm/subdev/mmu/priv.h | 2 + drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c | 8 ++-- .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c | 29 ++++++++---- 5 files changed, 66 insertions(+), 24 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_svm.c b/drivers/gpu/drm/nouveau/nouveau_svm.c index a87b9347d6ce..f9892e71a960 100644 --- a/drivers/gpu/drm/nouveau/nouveau_svm.c +++ b/drivers/gpu/drm/nouveau/nouveau_svm.c @@ -514,38 +514,51 @@ static const struct mmu_interval_notifier_ops nouveau_svm_mni_ops = { };
static void nouveau_hmm_convert_pfn(struct nouveau_drm *drm, - struct hmm_range *range, u64 *ioctl_addr) + struct hmm_range *range, + struct nouveau_pfnmap_args *args) { struct page *page;
/* - * The ioctl_addr prepared here is passed through nvif_object_ioctl() + * The address prepared here is passed through nvif_object_ioctl() * to an eventual DMA map in something like gp100_vmm_pgt_pfn() * * This is all just encoding the internal hmm representation into a * different nouveau internal representation. */ if (!(range->hmm_pfns[0] & HMM_PFN_VALID)) { - ioctl_addr[0] = 0; + args->p.phys[0] = 0; return; }
page = hmm_pfn_to_page(range->hmm_pfns[0]); + /* + * Only map compound pages to the GPU if the CPU is also mapping the + * page as a compound page. Otherwise, the PTE protections might not be + * consistent (i.e., CPU only maps part of a compound page). + */ + if (range->hmm_pfns[0] & HMM_PFN_COMPOUND) { + page = compound_head(page); + args->p.page = page_shift(page); + args->p.size = 1UL << args->p.page; + args->p.addr &= ~(args->p.size - 1); + } if (is_device_private_page(page)) - ioctl_addr[0] = nouveau_dmem_page_addr(page) | + args->p.phys[0] = nouveau_dmem_page_addr(page) | NVIF_VMM_PFNMAP_V0_V | NVIF_VMM_PFNMAP_V0_VRAM; else - ioctl_addr[0] = page_to_phys(page) | + args->p.phys[0] = page_to_phys(page) | NVIF_VMM_PFNMAP_V0_V | NVIF_VMM_PFNMAP_V0_HOST; if (range->hmm_pfns[0] & HMM_PFN_WRITE) - ioctl_addr[0] |= NVIF_VMM_PFNMAP_V0_W; + args->p.phys[0] |= NVIF_VMM_PFNMAP_V0_W; }
static int nouveau_range_fault(struct nouveau_svmm *svmm, - struct nouveau_drm *drm, void *data, u32 size, - u64 *ioctl_addr, unsigned long hmm_flags, + struct nouveau_drm *drm, + struct nouveau_pfnmap_args *args, u32 size, + unsigned long hmm_flags, struct svm_notifier *notifier) { unsigned long timeout = @@ -585,10 +598,10 @@ static int nouveau_range_fault(struct nouveau_svmm *svmm, break; }
- nouveau_hmm_convert_pfn(drm, &range, ioctl_addr); + nouveau_hmm_convert_pfn(drm, &range, args);
svmm->vmm->vmm.object.client->super = true; - ret = nvif_object_ioctl(&svmm->vmm->vmm.object, data, size, NULL); + ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, size, NULL); svmm->vmm->vmm.object.client->super = false; mutex_unlock(&svmm->mutex);
@@ -716,12 +729,13 @@ nouveau_svm_fault(struct nvif_notify *notify) args.i.p.addr, args.i.p.size, &nouveau_svm_mni_ops); if (!ret) { - ret = nouveau_range_fault(svmm, svm->drm, &args, - sizeof(args), args.phys, hmm_flags, ¬ifier); + ret = nouveau_range_fault(svmm, svm->drm, &args.i, + sizeof(args), hmm_flags, ¬ifier); mmu_interval_notifier_remove(¬ifier.notifier); } mmput(mm);
+ limit = args.i.p.addr + args.i.p.size; for (fn = fi; ++fn < buffer->fault_nr; ) { /* It's okay to skip over duplicate addresses from the * same SVMM as faults are ordered by access type such @@ -729,9 +743,16 @@ nouveau_svm_fault(struct nvif_notify *notify) * * ie. WRITE faults appear first, thus any handling of * pending READ faults will already be satisfied. + * But if a large page is mapped, make sure subsequent + * fault addresses have sufficient access permission. */ if (buffer->fault[fn]->svmm != svmm || - buffer->fault[fn]->addr >= limit) + buffer->fault[fn]->addr >= limit || + (buffer->fault[fi]->access == 0 /* READ. */ && + !(args.phys[0] & NVIF_VMM_PFNMAP_V0_V)) || + (buffer->fault[fi]->access != 0 /* READ. */ && + buffer->fault[fi]->access != 3 /* PREFETCH. */ && + !(args.phys[0] & NVIF_VMM_PFNMAP_V0_W))) break; }
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c index de91e9a26172..ecea365d72ad 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c @@ -94,6 +94,8 @@ nvkm_mmu_ptp_get(struct nvkm_mmu *mmu, u32 size, bool zero) } pt->ptp = ptp; pt->sub = true; + pt->ptei_shift = 3; + pt->page_shift = PAGE_SHIFT;
/* Sub-allocate from parent object, removing PTP from cache * if there's no more free slots left. @@ -203,6 +205,8 @@ nvkm_mmu_ptc_get(struct nvkm_mmu *mmu, u32 size, u32 align, bool zero) return NULL; pt->ptc = ptc; pt->sub = false; + pt->ptei_shift = 3; + pt->page_shift = PAGE_SHIFT;
ret = nvkm_memory_new(mmu->subdev.device, NVKM_MEM_TARGET_INST, size, align, zero, &pt->memory); diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h index 479b02344271..f2162bb35bea 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h @@ -55,6 +55,8 @@ struct nvkm_mmu_pt { struct nvkm_memory *memory; bool sub; u16 base; + u8 ptei_shift; + u8 page_shift; u64 addr; struct list_head head; }; diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c index 67b00dcef4b8..c7581f4f313e 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c @@ -562,6 +562,9 @@ nvkm_vmm_iter(struct nvkm_vmm *vmm, const struct nvkm_vmm_page *page, /* Handle PTE updates. */ if (!REF_PTES || REF_PTES(&it, pfn, ptei, ptes)) { struct nvkm_mmu_pt *pt = pgt->pt[type]; + + pt->page_shift = page->shift; + pt->ptei_shift = ilog2(desc->size); if (MAP_PTES || CLR_PTES) { if (MAP_PTES) MAP_PTES(vmm, pt, ptei, ptes, map); @@ -1204,7 +1207,6 @@ nvkm_vmm_pfn_unmap(struct nvkm_vmm *vmm, u64 addr, u64 size) /*TODO: * - Avoid PT readback (for dma_unmap etc), this might end up being dealt * with inside HMM, which would be a lot nicer for us to deal with. - * - Multiple page sizes (particularly for huge page support). * - Support for systems without a 4KiB page size. */ int @@ -1220,8 +1222,8 @@ nvkm_vmm_pfn_map(struct nvkm_vmm *vmm, u8 shift, u64 addr, u64 size, u64 *pfn) /* Only support mapping where the page size of the incoming page * array matches a page size available for direct mapping. */ - while (page->shift && page->shift != shift && - page->desc->func->pfn == NULL) + while (page->shift && (page->shift != shift || + page->desc->func->pfn == NULL)) page++;
if (!page->shift || !IS_ALIGNED(addr, 1ULL << shift) || diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c index d86287565542..94507cb2cf75 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c @@ -39,12 +39,15 @@ gp100_vmm_pfn_unmap(struct nvkm_vmm *vmm,
nvkm_kmap(pt->memory); while (ptes--) { - u32 datalo = nvkm_ro32(pt->memory, pt->base + ptei * 8 + 0); - u32 datahi = nvkm_ro32(pt->memory, pt->base + ptei * 8 + 4); + u32 datalo = nvkm_ro32(pt->memory, + pt->base + (ptei << pt->ptei_shift) + 0); + u32 datahi = nvkm_ro32(pt->memory, + pt->base + (ptei << pt->ptei_shift) + 4); u64 data = (u64)datahi << 32 | datalo; if ((data & (3ULL << 1)) != 0) { addr = (data >> 8) << 12; - dma_unmap_page(dev, addr, PAGE_SIZE, DMA_BIDIRECTIONAL); + dma_unmap_page(dev, addr, 1UL << pt->page_shift, + DMA_BIDIRECTIONAL); } ptei++; } @@ -58,11 +61,14 @@ gp100_vmm_pfn_clear(struct nvkm_vmm *vmm, bool dma = false; nvkm_kmap(pt->memory); while (ptes--) { - u32 datalo = nvkm_ro32(pt->memory, pt->base + ptei * 8 + 0); - u32 datahi = nvkm_ro32(pt->memory, pt->base + ptei * 8 + 4); + u32 datalo = nvkm_ro32(pt->memory, + pt->base + (ptei << pt->ptei_shift) + 0); + u32 datahi = nvkm_ro32(pt->memory, + pt->base + (ptei << pt->ptei_shift) + 4); u64 data = (u64)datahi << 32 | datalo; if ((data & BIT_ULL(0)) && (data & (3ULL << 1)) != 0) { - VMM_WO064(pt, vmm, ptei * 8, data & ~BIT_ULL(0)); + VMM_WO064(pt, vmm, ptei << pt->ptei_shift, + data & ~BIT_ULL(0)); dma = true; } ptei++; @@ -87,7 +93,8 @@ gp100_vmm_pgt_pfn(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt, if (!(*map->pfn & NVKM_VMM_PFN_VRAM)) { addr = *map->pfn >> NVKM_VMM_PFN_ADDR_SHIFT; addr = dma_map_page(dev, pfn_to_page(addr), 0, - PAGE_SIZE, DMA_BIDIRECTIONAL); + 1UL << pt->page_shift, + DMA_BIDIRECTIONAL); if (!WARN_ON(dma_mapping_error(dev, addr))) { data |= addr >> 4; data |= 2ULL << 1; /* SYSTEM_COHERENT_MEMORY. */ @@ -99,7 +106,7 @@ gp100_vmm_pgt_pfn(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt, data |= BIT_ULL(0); /* VALID. */ }
- VMM_WO064(pt, vmm, ptei++ * 8, data); + VMM_WO064(pt, vmm, ptei++ << pt->ptei_shift, data); map->pfn++; } nvkm_done(pt->memory); @@ -264,6 +271,9 @@ gp100_vmm_desc_pd0 = { .sparse = gp100_vmm_pd0_sparse, .pde = gp100_vmm_pd0_pde, .mem = gp100_vmm_pd0_mem, + .pfn = gp100_vmm_pgt_pfn, + .pfn_clear = gp100_vmm_pfn_clear, + .pfn_unmap = gp100_vmm_pfn_unmap, };
static void @@ -286,6 +296,9 @@ gp100_vmm_desc_pd1 = { .unmap = gf100_vmm_pgt_unmap, .sparse = gp100_vmm_pgt_sparse, .pde = gp100_vmm_pd1_pde, + .pfn = gp100_vmm_pgt_pfn, + .pfn_clear = gp100_vmm_pfn_clear, + .pfn_unmap = gp100_vmm_pfn_unmap, };
const struct nvkm_vmm_desc
Add some sanity tests for hmm_range_fault() returning the HMM_PFN_COMPOUND flag.
Signed-off-by: Ralph Campbell rcampbell@nvidia.com --- lib/test_hmm.c | 2 + lib/test_hmm_uapi.h | 2 + tools/testing/selftests/vm/hmm-tests.c | 76 ++++++++++++++++++++++++++ 3 files changed, 80 insertions(+)
diff --git a/lib/test_hmm.c b/lib/test_hmm.c index 00bca6116f93..8448921e6d33 100644 --- a/lib/test_hmm.c +++ b/lib/test_hmm.c @@ -757,6 +757,8 @@ static void dmirror_mkentry(struct dmirror *dmirror, struct hmm_range *range, *perm |= HMM_DMIRROR_PROT_WRITE; else *perm |= HMM_DMIRROR_PROT_READ; + if (entry & HMM_PFN_COMPOUND) + *perm |= HMM_DMIRROR_PROT_COMPOUND; }
static bool dmirror_snapshot_invalidate(struct mmu_interval_notifier *mni, diff --git a/lib/test_hmm_uapi.h b/lib/test_hmm_uapi.h index 67b3b2e6ff5d..21cf4da6f020 100644 --- a/lib/test_hmm_uapi.h +++ b/lib/test_hmm_uapi.h @@ -40,6 +40,7 @@ struct hmm_dmirror_cmd { * HMM_DMIRROR_PROT_NONE: unpopulated PTE or PTE with no access * HMM_DMIRROR_PROT_READ: read-only PTE * HMM_DMIRROR_PROT_WRITE: read/write PTE + * HMM_DMIRROR_PROT_COMPOUND: compound page is fully mapped by same permissions * HMM_DMIRROR_PROT_ZERO: special read-only zero page * HMM_DMIRROR_PROT_DEV_PRIVATE_LOCAL: Migrated device private page on the * device the ioctl() is made @@ -51,6 +52,7 @@ enum { HMM_DMIRROR_PROT_NONE = 0x00, HMM_DMIRROR_PROT_READ = 0x01, HMM_DMIRROR_PROT_WRITE = 0x02, + HMM_DMIRROR_PROT_COMPOUND = 0x04, HMM_DMIRROR_PROT_ZERO = 0x10, HMM_DMIRROR_PROT_DEV_PRIVATE_LOCAL = 0x20, HMM_DMIRROR_PROT_DEV_PRIVATE_REMOTE = 0x30, diff --git a/tools/testing/selftests/vm/hmm-tests.c b/tools/testing/selftests/vm/hmm-tests.c index 79db22604019..e9576d9802c7 100644 --- a/tools/testing/selftests/vm/hmm-tests.c +++ b/tools/testing/selftests/vm/hmm-tests.c @@ -1291,6 +1291,82 @@ TEST_F(hmm2, snapshot) hmm_buffer_free(buffer); }
+/* + * Test the hmm_range_fault() HMM_PFN_COMPOUND flag for large pages that + * should be mapped by a large page table entry. + */ +TEST_F(hmm, compound) +{ + struct hmm_buffer *buffer; + unsigned long npages; + unsigned long size; + int *ptr; + unsigned char *m; + int ret; + long pagesizes[4]; + int n, idx; + unsigned long i; + + /* Skip test if we can't allocate a hugetlbfs page. */ + + n = gethugepagesizes(pagesizes, 4); + if (n <= 0) + return; + for (idx = 0; --n > 0; ) { + if (pagesizes[n] < pagesizes[idx]) + idx = n; + } + size = ALIGN(TWOMEG, pagesizes[idx]); + npages = size >> self->page_shift; + + buffer = malloc(sizeof(*buffer)); + ASSERT_NE(buffer, NULL); + + buffer->ptr = get_hugepage_region(size, GHR_STRICT); + if (buffer->ptr == NULL) { + free(buffer); + return; + } + + buffer->size = size; + buffer->mirror = malloc(npages); + ASSERT_NE(buffer->mirror, NULL); + + /* Initialize the pages the device will snapshot in buffer->ptr. */ + for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i) + ptr[i] = i; + + /* Simulate a device snapshotting CPU pagetables. */ + ret = hmm_dmirror_cmd(self->fd, HMM_DMIRROR_SNAPSHOT, buffer, npages); + ASSERT_EQ(ret, 0); + ASSERT_EQ(buffer->cpages, npages); + + /* Check what the device saw. */ + m = buffer->mirror; + for (i = 0; i < npages; ++i) + ASSERT_EQ(m[i], HMM_DMIRROR_PROT_WRITE | + HMM_DMIRROR_PROT_COMPOUND); + + /* Make the region read-only. */ + ret = mprotect(buffer->ptr, size, PROT_READ); + ASSERT_EQ(ret, 0); + + /* Simulate a device snapshotting CPU pagetables. */ + ret = hmm_dmirror_cmd(self->fd, HMM_DMIRROR_SNAPSHOT, buffer, npages); + ASSERT_EQ(ret, 0); + ASSERT_EQ(buffer->cpages, npages); + + /* Check what the device saw. */ + m = buffer->mirror; + for (i = 0; i < npages; ++i) + ASSERT_EQ(m[i], HMM_DMIRROR_PROT_READ | + HMM_DMIRROR_PROT_COMPOUND); + + free_hugepage_region(buffer->ptr); + buffer->ptr = NULL; + hmm_buffer_free(buffer); +} + /* * Test two devices reading the same memory (double mapped). */
On Fri, May 08, 2020 at 12:20:03PM -0700, Ralph Campbell wrote:
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. This series adds a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available. Nouveau and the HMM tests are updated to use the new flag.
This explanation doesn't make any sense. It doesn't matter how somebody else has it mapped; if it's a PMD-sized page, you can map it with a 2MB mapping.
On 5/8/20 12:59 PM, Matthew Wilcox wrote:
On Fri, May 08, 2020 at 12:20:03PM -0700, Ralph Campbell wrote:
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. This series adds a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available. Nouveau and the HMM tests are updated to use the new flag.
This explanation doesn't make any sense. It doesn't matter how somebody else has it mapped; if it's a PMD-sized page, you can map it with a 2MB mapping.
Sure, the I/O will work OK, but is it safe? Copy on write isn't an issue? splitting a PMD in one process due to mprotect of a shared page will cause other process' page tables to be split the same way? Recall that these are system memory pages that could be THPs, shmem, hugetlbfs, mmap shared file pages, etc.
On Fri, May 08, 2020 at 01:17:55PM -0700, Ralph Campbell wrote:
On 5/8/20 12:59 PM, Matthew Wilcox wrote:
On Fri, May 08, 2020 at 12:20:03PM -0700, Ralph Campbell wrote:
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. This series adds a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available. Nouveau and the HMM tests are updated to use the new flag.
This explanation doesn't make any sense. It doesn't matter how somebody else has it mapped; if it's a PMD-sized page, you can map it with a 2MB mapping.
Sure, the I/O will work OK, but is it safe? Copy on write isn't an issue? splitting a PMD in one process due to mprotect of a shared page will cause other process' page tables to be split the same way?
Are you saying that if you call this function on an address range of a process which has done COW of a single page in the middle of a THP, you want to return with this flag clear, but if the THP is still intact, you want to set this flag?
Recall that these are system memory pages that could be THPs, shmem, hugetlbfs, mmap shared file pages, etc.
On 5/8/20 8:17 PM, Matthew Wilcox wrote:
On Fri, May 08, 2020 at 01:17:55PM -0700, Ralph Campbell wrote:
On 5/8/20 12:59 PM, Matthew Wilcox wrote:
On Fri, May 08, 2020 at 12:20:03PM -0700, Ralph Campbell wrote:
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. This series adds a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available. Nouveau and the HMM tests are updated to use the new flag.
This explanation doesn't make any sense. It doesn't matter how somebody else has it mapped; if it's a PMD-sized page, you can map it with a 2MB mapping.
Sure, the I/O will work OK, but is it safe? Copy on write isn't an issue? splitting a PMD in one process due to mprotect of a shared page will cause other process' page tables to be split the same way?
Are you saying that if you call this function on an address range of a process which has done COW of a single page in the middle of a THP, you want to return with this flag clear, but if the THP is still intact, you want to set this flag?
Correct. I want the GPU to see the same faults that the CPU would see when trying to access the same addresses. All faults, whether from CPU or GPU, end up calling handle_mm_fault() to handle the fault and update the GPU/CPU page tables.
Recall that these are system memory pages that could be THPs, shmem, hugetlbfs, mmap shared file pages, etc.
On Fri, May 08, 2020 at 12:20:03PM -0700, Ralph Campbell wrote:
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. This series adds a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available. Nouveau and the HMM tests are updated to use the new flag.
Note that this series depends on a patch queued in Ben Skeggs' nouveau tree ("nouveau/hmm: map pages after migration") and the patches queued in Jason's HMM tree. There is also a patch outstanding ("nouveau/hmm: fix nouveau_dmem_chunk allocations") that is independent of the above and could be applied before or after.
Did Christoph and Matt's remarks get addressed here?
I think ODP could use something like this, currently it checks every page to get back to the huge page size and this flag would optimze that
Jason
On 5/25/20 6:41 AM, Jason Gunthorpe wrote:
On Fri, May 08, 2020 at 12:20:03PM -0700, Ralph Campbell wrote:
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. This series adds a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available. Nouveau and the HMM tests are updated to use the new flag.
Note that this series depends on a patch queued in Ben Skeggs' nouveau tree ("nouveau/hmm: map pages after migration") and the patches queued in Jason's HMM tree. There is also a patch outstanding ("nouveau/hmm: fix nouveau_dmem_chunk allocations") that is independent of the above and could be applied before or after.
Did Christoph and Matt's remarks get addressed here?
Both questioned the need to add the HMM_PFN_COMPOUND flag to the hmm_range_fault() output array saying that the PFN can be used to get the struct page pointer and the page can be examined to determine the page size. My response is that while is true, it is also important that the device only access the same parts of a large page that the process/cpu has access to. There are places where a large page is mapped with smaller page table entries when a page is shared by multiple processes. After I explained this, I haven't seen any further comments from Christoph and Matt. I'm still looking for reviews, acks, or suggested changes.
I think ODP could use something like this, currently it checks every page to get back to the huge page size and this flag would optimze that
Jason
On Tue, May 26, 2020 at 10:32:48AM -0700, Ralph Campbell wrote:
On 5/25/20 6:41 AM, Jason Gunthorpe wrote:
On Fri, May 08, 2020 at 12:20:03PM -0700, Ralph Campbell wrote:
hmm_range_fault() returns an array of page frame numbers and flags for how the pages are mapped in the requested process' page tables. The PFN can be used to get the struct page with hmm_pfn_to_page() and the page size order can be determined with compound_order(page) but if the page is larger than order 0 (PAGE_SIZE), there is no indication that the page is mapped using a larger page size. To be fully general, hmm_range_fault() would need to return the mapping size to handle cases like a 1GB compound page being mapped with 2MB PMD entries. However, the most common case is the mapping size the same as the underlying compound page size. This series adds a new output flag to indicate this so that callers know it is safe to use a large device page table mapping if one is available. Nouveau and the HMM tests are updated to use the new flag.
Note that this series depends on a patch queued in Ben Skeggs' nouveau tree ("nouveau/hmm: map pages after migration") and the patches queued in Jason's HMM tree. There is also a patch outstanding ("nouveau/hmm: fix nouveau_dmem_chunk allocations") that is independent of the above and could be applied before or after.
Did Christoph and Matt's remarks get addressed here?
Both questioned the need to add the HMM_PFN_COMPOUND flag to the hmm_range_fault() output array saying that the PFN can be used to get the struct page pointer and the page can be examined to determine the page size. My response is that while is true, it is also important that the device only access the same parts of a large page that the process/cpu has access to. There are places where a large page is mapped with smaller page table entries when a page is shared by multiple processes. After I explained this, I haven't seen any further comments from Christoph and Matt. I'm still looking for reviews, acks, or suggested changes.
Okay, well, we reached the merge window, so since there may be some conflicts repost again in three weeks.
It would be more compelling if there was some performance data if it is much of a win vs the 'compute large page' algorithm something like ODP uses.
Jason
linux-kselftest-mirror@lists.linaro.org
-
Christoph Hellwig -
Jason Gunthorpe -
Matthew Wilcox -
Ralph Campbell -
Zi Yan