From 314b6580adc543cf21a1a84bce9c6a58a8dcb38c Mon Sep 17 00:00:00 2001 From: "Thomas Hellstrom (VMware)" Date: Tue, 24 Mar 2020 18:48:33 +0100 Subject: drm/ttm, drm/vmwgfx: Support huge TTM pagefaults MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Support huge (PMD-size and PUD-size) page-table entries by providing a huge_fault() callback. We still support private mappings and write-notify by splitting the huge page-table entries on write-access. Note that for huge page-faults to occur, either the kernel needs to be compiled with trans-huge-pages always enabled, or the kernel needs to be compiled with trans-huge-pages enabled using madvise, and the user-space app needs to call madvise() to enable trans-huge pages on a per-mapping basis. Furthermore huge page-faults will not succeed unless buffer objects and user-space addresses are aligned on huge page size boundaries. Cc: Andrew Morton Cc: Michal Hocko Cc: "Matthew Wilcox (Oracle)" Cc: "Kirill A. Shutemov" Cc: Ralph Campbell Cc: "Jérôme Glisse" Cc: "Christian König" Cc: Dan Williams Signed-off-by: Thomas Hellstrom (VMware) Reviewed-by: Roland Scheidegger Reviewed-by: Christian König --- drivers/gpu/drm/ttm/ttm_bo_vm.c | 161 +++++++++++++++++++++++++++++++++++++++- 1 file changed, 158 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/ttm') diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c index 389128b8c4dd..0af14835504c 100644 --- a/drivers/gpu/drm/ttm/ttm_bo_vm.c +++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c @@ -156,6 +156,89 @@ vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo, } EXPORT_SYMBOL(ttm_bo_vm_reserve); +#ifdef CONFIG_TRANSPARENT_HUGEPAGE +/** + * ttm_bo_vm_insert_huge - Insert a pfn for PUD or PMD faults + * @vmf: Fault data + * @bo: The buffer object + * @page_offset: Page offset from bo start + * @fault_page_size: The size of the fault in pages. + * @pgprot: The page protections. + * Does additional checking whether it's possible to insert a PUD or PMD + * pfn and performs the insertion. + * + * Return: VM_FAULT_NOPAGE on successful insertion, VM_FAULT_FALLBACK if + * a huge fault was not possible, or on insertion error. + */ +static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf, + struct ttm_buffer_object *bo, + pgoff_t page_offset, + pgoff_t fault_page_size, + pgprot_t pgprot) +{ + pgoff_t i; + vm_fault_t ret; + unsigned long pfn; + pfn_t pfnt; + struct ttm_tt *ttm = bo->ttm; + bool write = vmf->flags & FAULT_FLAG_WRITE; + + /* Fault should not cross bo boundary. */ + page_offset &= ~(fault_page_size - 1); + if (page_offset + fault_page_size > bo->num_pages) + goto out_fallback; + + if (bo->mem.bus.is_iomem) + pfn = ttm_bo_io_mem_pfn(bo, page_offset); + else + pfn = page_to_pfn(ttm->pages[page_offset]); + + /* pfn must be fault_page_size aligned. */ + if ((pfn & (fault_page_size - 1)) != 0) + goto out_fallback; + + /* Check that memory is contiguous. */ + if (!bo->mem.bus.is_iomem) { + for (i = 1; i < fault_page_size; ++i) { + if (page_to_pfn(ttm->pages[page_offset + i]) != pfn + i) + goto out_fallback; + } + } else if (bo->bdev->driver->io_mem_pfn) { + for (i = 1; i < fault_page_size; ++i) { + if (ttm_bo_io_mem_pfn(bo, page_offset + i) != pfn + i) + goto out_fallback; + } + } + + pfnt = __pfn_to_pfn_t(pfn, PFN_DEV); + if (fault_page_size == (HPAGE_PMD_SIZE >> PAGE_SHIFT)) + ret = vmf_insert_pfn_pmd_prot(vmf, pfnt, pgprot, write); +#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD + else if (fault_page_size == (HPAGE_PUD_SIZE >> PAGE_SHIFT)) + ret = vmf_insert_pfn_pud_prot(vmf, pfnt, pgprot, write); +#endif + else + WARN_ON_ONCE(ret = VM_FAULT_FALLBACK); + + if (ret != VM_FAULT_NOPAGE) + goto out_fallback; + + return VM_FAULT_NOPAGE; +out_fallback: + count_vm_event(THP_FAULT_FALLBACK); + return VM_FAULT_FALLBACK; +} +#else +static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf, + struct ttm_buffer_object *bo, + pgoff_t page_offset, + pgoff_t fault_page_size, + pgprot_t pgprot) +{ + return VM_FAULT_FALLBACK; +} +#endif + /** * ttm_bo_vm_fault_reserved - TTM fault helper * @vmf: The struct vm_fault given as argument to the fault callback @@ -163,6 +246,7 @@ EXPORT_SYMBOL(ttm_bo_vm_reserve); * @num_prefault: Maximum number of prefault pages. The caller may want to * specify this based on madvice settings and the size of the GPU object * backed by the memory. + * @fault_page_size: The size of the fault in pages. * * This function inserts one or more page table entries pointing to the * memory backing the buffer object, and then returns a return code @@ -176,7 +260,8 @@ EXPORT_SYMBOL(ttm_bo_vm_reserve); */ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, pgprot_t prot, - pgoff_t num_prefault) + pgoff_t num_prefault, + pgoff_t fault_page_size) { struct vm_area_struct *vma = vmf->vma; struct ttm_buffer_object *bo = vma->vm_private_data; @@ -268,6 +353,13 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, prot = pgprot_decrypted(prot); } + /* We don't prefault on huge faults. Yet. */ + if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && fault_page_size != 1) { + ret = ttm_bo_vm_insert_huge(vmf, bo, page_offset, + fault_page_size, prot); + goto out_io_unlock; + } + /* * Speculatively prefault a number of pages. Only error on * first page. @@ -334,7 +426,7 @@ vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf) return ret; prot = vma->vm_page_prot; - ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT); + ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT, 1); if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) return ret; @@ -344,6 +436,66 @@ vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf) } EXPORT_SYMBOL(ttm_bo_vm_fault); +#ifdef CONFIG_TRANSPARENT_HUGEPAGE +/** + * ttm_pgprot_is_wrprotecting - Is a page protection value write-protecting? + * @prot: The page protection value + * + * Return: true if @prot is write-protecting. false otherwise. + */ +static bool ttm_pgprot_is_wrprotecting(pgprot_t prot) +{ + /* + * This is meant to say "pgprot_wrprotect(prot) == prot" in a generic + * way. Unfortunately there is no generic pgprot_wrprotect. + */ + return pte_val(pte_wrprotect(__pte(pgprot_val(prot)))) == + pgprot_val(prot); +} + +static vm_fault_t ttm_bo_vm_huge_fault(struct vm_fault *vmf, + enum page_entry_size pe_size) +{ + struct vm_area_struct *vma = vmf->vma; + pgprot_t prot; + struct ttm_buffer_object *bo = vma->vm_private_data; + vm_fault_t ret; + pgoff_t fault_page_size = 0; + bool write = vmf->flags & FAULT_FLAG_WRITE; + + switch (pe_size) { + case PE_SIZE_PMD: + fault_page_size = HPAGE_PMD_SIZE >> PAGE_SHIFT; + break; +#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD + case PE_SIZE_PUD: + fault_page_size = HPAGE_PUD_SIZE >> PAGE_SHIFT; + break; +#endif + default: + WARN_ON_ONCE(1); + return VM_FAULT_FALLBACK; + } + + /* Fallback on write dirty-tracking or COW */ + if (write && ttm_pgprot_is_wrprotecting(vma->vm_page_prot)) + return VM_FAULT_FALLBACK; + + ret = ttm_bo_vm_reserve(bo, vmf); + if (ret) + return ret; + + prot = vm_get_page_prot(vma->vm_flags); + ret = ttm_bo_vm_fault_reserved(vmf, prot, 1, fault_page_size); + if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) + return ret; + + dma_resv_unlock(bo->base.resv); + + return ret; +} +#endif + void ttm_bo_vm_open(struct vm_area_struct *vma) { struct ttm_buffer_object *bo = vma->vm_private_data; @@ -445,7 +597,10 @@ static const struct vm_operations_struct ttm_bo_vm_ops = { .fault = ttm_bo_vm_fault, .open = ttm_bo_vm_open, .close = ttm_bo_vm_close, - .access = ttm_bo_vm_access + .access = ttm_bo_vm_access, +#ifdef CONFIG_TRANSPARENT_HUGEPAGE + .huge_fault = ttm_bo_vm_huge_fault, +#endif }; static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev, -- cgit From 4064b982706375025628094e51d11cf1a958a5d3 Mon Sep 17 00:00:00 2001 From: Peter Xu Date: Wed, 1 Apr 2020 21:08:45 -0700 Subject: mm: allow VM_FAULT_RETRY for multiple times The idea comes from a discussion between Linus and Andrea [1]. Before this patch we only allow a page fault to retry once. We achieved this by clearing the FAULT_FLAG_ALLOW_RETRY flag when doing handle_mm_fault() the second time. This was majorly used to avoid unexpected starvation of the system by looping over forever to handle the page fault on a single page. However that should hardly happen, and after all for each code path to return a VM_FAULT_RETRY we'll first wait for a condition (during which time we should possibly yield the cpu) to happen before VM_FAULT_RETRY is really returned. This patch removes the restriction by keeping the FAULT_FLAG_ALLOW_RETRY flag when we receive VM_FAULT_RETRY. It means that the page fault handler now can retry the page fault for multiple times if necessary without the need to generate another page fault event. Meanwhile we still keep the FAULT_FLAG_TRIED flag so page fault handler can still identify whether a page fault is the first attempt or not. Then we'll have these combinations of fault flags (only considering ALLOW_RETRY flag and TRIED flag): - ALLOW_RETRY and !TRIED: this means the page fault allows to retry, and this is the first try - ALLOW_RETRY and TRIED: this means the page fault allows to retry, and this is not the first try - !ALLOW_RETRY and !TRIED: this means the page fault does not allow to retry at all - !ALLOW_RETRY and TRIED: this is forbidden and should never be used In existing code we have multiple places that has taken special care of the first condition above by checking against (fault_flags & FAULT_FLAG_ALLOW_RETRY). This patch introduces a simple helper to detect the first retry of a page fault by checking against both (fault_flags & FAULT_FLAG_ALLOW_RETRY) and !(fault_flag & FAULT_FLAG_TRIED) because now even the 2nd try will have the ALLOW_RETRY set, then use that helper in all existing special paths. One example is in __lock_page_or_retry(), now we'll drop the mmap_sem only in the first attempt of page fault and we'll keep it in follow up retries, so old locking behavior will be retained. This will be a nice enhancement for current code [2] at the same time a supporting material for the future userfaultfd-writeprotect work, since in that work there will always be an explicit userfault writeprotect retry for protected pages, and if that cannot resolve the page fault (e.g., when userfaultfd-writeprotect is used in conjunction with swapped pages) then we'll possibly need a 3rd retry of the page fault. It might also benefit other potential users who will have similar requirement like userfault write-protection. GUP code is not touched yet and will be covered in follow up patch. Please read the thread below for more information. [1] https://lore.kernel.org/lkml/20171102193644.GB22686@redhat.com/ [2] https://lore.kernel.org/lkml/20181230154648.GB9832@redhat.com/ Suggested-by: Linus Torvalds Suggested-by: Andrea Arcangeli Signed-off-by: Peter Xu Signed-off-by: Andrew Morton Tested-by: Brian Geffon Cc: Bobby Powers Cc: David Hildenbrand Cc: Denis Plotnikov Cc: "Dr . David Alan Gilbert" Cc: Hugh Dickins Cc: Jerome Glisse Cc: Johannes Weiner Cc: "Kirill A . Shutemov" Cc: Martin Cracauer Cc: Marty McFadden Cc: Matthew Wilcox Cc: Maya Gokhale Cc: Mel Gorman Cc: Mike Kravetz Cc: Mike Rapoport Cc: Pavel Emelyanov Link: http://lkml.kernel.org/r/20200220160246.9790-1-peterx@redhat.com Signed-off-by: Linus Torvalds --- drivers/gpu/drm/ttm/ttm_bo_vm.c | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/ttm') diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c index 389128b8c4dd..cb8829ca6c7f 100644 --- a/drivers/gpu/drm/ttm/ttm_bo_vm.c +++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c @@ -59,9 +59,10 @@ static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo, /* * If possible, avoid waiting for GPU with mmap_sem - * held. + * held. We only do this if the fault allows retry and this + * is the first attempt. */ - if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) { + if (fault_flag_allow_retry_first(vmf->flags)) { ret = VM_FAULT_RETRY; if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) goto out_unlock; @@ -135,7 +136,12 @@ vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo, * for the buffer to become unreserved. */ if (unlikely(!dma_resv_trylock(bo->base.resv))) { - if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) { + /* + * If the fault allows retry and this is the first + * fault attempt, we try to release the mmap_sem + * before waiting + */ + if (fault_flag_allow_retry_first(vmf->flags)) { if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) { ttm_bo_get(bo); up_read(&vmf->vma->vm_mm->mmap_sem); -- cgit From c445c1668dca9cb908f75e000ab20351bce7c416 Mon Sep 17 00:00:00 2001 From: "Thomas Hellstrom (VMware)" Date: Thu, 9 Apr 2020 18:49:25 +0200 Subject: drm/ttm: Temporarily disable the huge_fault() callback MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit With amdgpu and CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS=y, there are errors like: BUG: non-zero pgtables_bytes on freeing mm and: BUG: Bad rss-counter state with TTM transparent huge-pages. Until we've figured out what other TTM drivers do differently compared to vmwgfx, disable the huge_fault() callback, eliminating transhuge page-table entries. Cc: Christian König Signed-off-by: Thomas Hellstrom (VMware) Reported-by: Alex Xu (Hello71) Tested-by: Alex Xu (Hello71) Acked-by: Christian König Signed-off-by: Dave Airlie Link: https://patchwork.freedesktop.org/patch/msgid/20200409164925.11912-1-thomas_os@shipmail.org --- drivers/gpu/drm/ttm/ttm_bo_vm.c | 63 ----------------------------------------- 1 file changed, 63 deletions(-) (limited to 'drivers/gpu/drm/ttm') diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c index 0af14835504c..0871c7d6cb8c 100644 --- a/drivers/gpu/drm/ttm/ttm_bo_vm.c +++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c @@ -436,66 +436,6 @@ vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf) } EXPORT_SYMBOL(ttm_bo_vm_fault); -#ifdef CONFIG_TRANSPARENT_HUGEPAGE -/** - * ttm_pgprot_is_wrprotecting - Is a page protection value write-protecting? - * @prot: The page protection value - * - * Return: true if @prot is write-protecting. false otherwise. - */ -static bool ttm_pgprot_is_wrprotecting(pgprot_t prot) -{ - /* - * This is meant to say "pgprot_wrprotect(prot) == prot" in a generic - * way. Unfortunately there is no generic pgprot_wrprotect. - */ - return pte_val(pte_wrprotect(__pte(pgprot_val(prot)))) == - pgprot_val(prot); -} - -static vm_fault_t ttm_bo_vm_huge_fault(struct vm_fault *vmf, - enum page_entry_size pe_size) -{ - struct vm_area_struct *vma = vmf->vma; - pgprot_t prot; - struct ttm_buffer_object *bo = vma->vm_private_data; - vm_fault_t ret; - pgoff_t fault_page_size = 0; - bool write = vmf->flags & FAULT_FLAG_WRITE; - - switch (pe_size) { - case PE_SIZE_PMD: - fault_page_size = HPAGE_PMD_SIZE >> PAGE_SHIFT; - break; -#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD - case PE_SIZE_PUD: - fault_page_size = HPAGE_PUD_SIZE >> PAGE_SHIFT; - break; -#endif - default: - WARN_ON_ONCE(1); - return VM_FAULT_FALLBACK; - } - - /* Fallback on write dirty-tracking or COW */ - if (write && ttm_pgprot_is_wrprotecting(vma->vm_page_prot)) - return VM_FAULT_FALLBACK; - - ret = ttm_bo_vm_reserve(bo, vmf); - if (ret) - return ret; - - prot = vm_get_page_prot(vma->vm_flags); - ret = ttm_bo_vm_fault_reserved(vmf, prot, 1, fault_page_size); - if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) - return ret; - - dma_resv_unlock(bo->base.resv); - - return ret; -} -#endif - void ttm_bo_vm_open(struct vm_area_struct *vma) { struct ttm_buffer_object *bo = vma->vm_private_data; @@ -598,9 +538,6 @@ static const struct vm_operations_struct ttm_bo_vm_ops = { .open = ttm_bo_vm_open, .close = ttm_bo_vm_close, .access = ttm_bo_vm_access, -#ifdef CONFIG_TRANSPARENT_HUGEPAGE - .huge_fault = ttm_bo_vm_huge_fault, -#endif }; static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev, -- cgit