Merge branch 'ttm-transhuge' of git://people.freedesktop.org/~thomash/linux into drm-next

Huge page-table entries for TTM

In order to reduce CPU usage [1] and in theory TLB misses this patchset enables
huge- and giant page-table entries for TTM and TTM-enabled graphics drivers.

Signed-off-by: Dave Airlie <[email protected]>
From: Thomas Hellstrom (VMware) <[email protected]>
Link: https://patchwork.freedesktop.org/patch/msgid/[email protected]

Author: airlied

drivers/gpu/drm/drm_file.c

@@ -48,6 +48,11 @@
 #include "drm_internal.h"
 #include "drm_legacy.h"
 
+#if defined(CONFIG_MMU) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
+#include <uapi/asm/mman.h>
+#include <drm/drm_vma_manager.h>
+#endif
+
 /* from BKL pushdown */
 DEFINE_MUTEX(drm_global_mutex);
 
@@ -872,3 +877,139 @@ struct file *mock_drm_getfile(struct drm_minor *minor, unsigned int flags)
 	return file;
 }
 EXPORT_SYMBOL_FOR_TESTS_ONLY(mock_drm_getfile);
+
+#ifdef CONFIG_MMU
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+/*
+ * drm_addr_inflate() attempts to construct an aligned area by inflating
+ * the area size and skipping the unaligned start of the area.
+ * adapted from shmem_get_unmapped_area()
+ */
+static unsigned long drm_addr_inflate(unsigned long addr,
+				      unsigned long len,
+				      unsigned long pgoff,
+				      unsigned long flags,
+				      unsigned long huge_size)
+{
+	unsigned long offset, inflated_len;
+	unsigned long inflated_addr;
+	unsigned long inflated_offset;
+
+	offset = (pgoff << PAGE_SHIFT) & (huge_size - 1);
+	if (offset && offset + len < 2 * huge_size)
+		return addr;
+	if ((addr & (huge_size - 1)) == offset)
+		return addr;
+
+	inflated_len = len + huge_size - PAGE_SIZE;
+	if (inflated_len > TASK_SIZE)
+		return addr;
+	if (inflated_len < len)
+		return addr;
+
+	inflated_addr = current->mm->get_unmapped_area(NULL, 0, inflated_len,
+						       0, flags);
+	if (IS_ERR_VALUE(inflated_addr))
+		return addr;
+	if (inflated_addr & ~PAGE_MASK)
+		return addr;
+
+	inflated_offset = inflated_addr & (huge_size - 1);
+	inflated_addr += offset - inflated_offset;
+	if (inflated_offset > offset)
+		inflated_addr += huge_size;
+
+	if (inflated_addr > TASK_SIZE - len)
+		return addr;
+
+	return inflated_addr;
+}
+
+/**
+ * drm_get_unmapped_area() - Get an unused user-space virtual memory area
+ * suitable for huge page table entries.
+ * @file: The struct file representing the address space being mmap()'d.
+ * @uaddr: Start address suggested by user-space.
+ * @len: Length of the area.
+ * @pgoff: The page offset into the address space.
+ * @flags: mmap flags
+ * @mgr: The address space manager used by the drm driver. This argument can
+ * probably be removed at some point when all drivers use the same
+ * address space manager.
+ *
+ * This function attempts to find an unused user-space virtual memory area
+ * that can accommodate the size we want to map, and that is properly
+ * aligned to facilitate huge page table entries matching actual
+ * huge pages or huge page aligned memory in buffer objects. Buffer objects
+ * are assumed to start at huge page boundary pfns (io memory) or be
+ * populated by huge pages aligned to the start of the buffer object
+ * (system- or coherent memory). Adapted from shmem_get_unmapped_area.
+ *
+ * Return: aligned user-space address.
+ */
+unsigned long drm_get_unmapped_area(struct file *file,
+				    unsigned long uaddr, unsigned long len,
+				    unsigned long pgoff, unsigned long flags,
+				    struct drm_vma_offset_manager *mgr)
+{
+	unsigned long addr;
+	unsigned long inflated_addr;
+	struct drm_vma_offset_node *node;
+
+	if (len > TASK_SIZE)
+		return -ENOMEM;
+
+	/*
+	 * @pgoff is the file page-offset the huge page boundaries of
+	 * which typically aligns to physical address huge page boundaries.
+	 * That's not true for DRM, however, where physical address huge
+	 * page boundaries instead are aligned with the offset from
+	 * buffer object start. So adjust @pgoff to be the offset from
+	 * buffer object start.
+	 */
+	drm_vma_offset_lock_lookup(mgr);
+	node = drm_vma_offset_lookup_locked(mgr, pgoff, 1);
+	if (node)
+		pgoff -= node->vm_node.start;
+	drm_vma_offset_unlock_lookup(mgr);
+
+	addr = current->mm->get_unmapped_area(file, uaddr, len, pgoff, flags);
+	if (IS_ERR_VALUE(addr))
+		return addr;
+	if (addr & ~PAGE_MASK)
+		return addr;
+	if (addr > TASK_SIZE - len)
+		return addr;
+
+	if (len < HPAGE_PMD_SIZE)
+		return addr;
+	if (flags & MAP_FIXED)
+		return addr;
+	/*
+	 * Our priority is to support MAP_SHARED mapped hugely;
+	 * and support MAP_PRIVATE mapped hugely too, until it is COWed.
+	 * But if caller specified an address hint, respect that as before.
+	 */
+	if (uaddr)
+		return addr;
+
+	inflated_addr = drm_addr_inflate(addr, len, pgoff, flags,
+					 HPAGE_PMD_SIZE);
+
+	if (IS_ENABLED(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) &&
+	    len >= HPAGE_PUD_SIZE)
+		inflated_addr = drm_addr_inflate(inflated_addr, len, pgoff,
+						 flags, HPAGE_PUD_SIZE);
+	return inflated_addr;
+}
+#else /* CONFIG_TRANSPARENT_HUGEPAGE */
+unsigned long drm_get_unmapped_area(struct file *file,
+				    unsigned long uaddr, unsigned long len,
+				    unsigned long pgoff, unsigned long flags,
+				    struct drm_vma_offset_manager *mgr)
+{
+	return current->mm->get_unmapped_area(file, uaddr, len, pgoff, flags);
+}
+#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
+EXPORT_SYMBOL_GPL(drm_get_unmapped_area);
+#endif /* CONFIG_MMU */

drivers/gpu/drm/ttm/ttm_bo_vm.c

@@ -156,13 +156,97 @@ 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
  * @prot: The page protection to be used for this memory area.
  * @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,

drivers/gpu/drm/vmwgfx/Makefile

@@ -11,4 +11,5 @@ vmwgfx-y := vmwgfx_execbuf.o vmwgfx_gmr.o vmwgfx_kms.o vmwgfx_drv.o \
 	    vmwgfx_validation.o vmwgfx_page_dirty.o vmwgfx_streamoutput.o \
 	    ttm_object.o ttm_lock.o
 
+vmwgfx-$(CONFIG_TRANSPARENT_HUGEPAGE) += vmwgfx_thp.o
 obj-$(CONFIG_DRM_VMWGFX) := vmwgfx.o

drivers/gpu/drm/vmwgfx/vmwgfx_drv.c

@@ -1247,6 +1247,18 @@ static void vmw_remove(struct pci_dev *pdev)
 	pci_disable_device(pdev);
 }
 
+static unsigned long
+vmw_get_unmapped_area(struct file *file, unsigned long uaddr,
+		      unsigned long len, unsigned long pgoff,
+		      unsigned long flags)
+{
+	struct drm_file *file_priv = file->private_data;
+	struct vmw_private *dev_priv = vmw_priv(file_priv->minor->dev);
+
+	return drm_get_unmapped_area(file, uaddr, len, pgoff, flags,
+				     &dev_priv->vma_manager);
+}
+
 static int vmwgfx_pm_notifier(struct notifier_block *nb, unsigned long val,
 			      void *ptr)
 {
@@ -1418,6 +1430,7 @@ static const struct file_operations vmwgfx_driver_fops = {
 	.compat_ioctl = vmw_compat_ioctl,
 #endif
 	.llseek = noop_llseek,
+	.get_unmapped_area = vmw_get_unmapped_area,
 };
 
 static struct drm_driver driver = {