i915: use io-mapping interfaces instead of a variety of mapping kludges
authorKeith Packard <keithp@keithp.com>
Fri, 31 Oct 2008 02:38:48 +0000 (19:38 -0700)
committerIngo Molnar <mingo@elte.hu>
Fri, 31 Oct 2008 09:12:40 +0000 (10:12 +0100)
Impact: optimize/clean-up the IO mapping implementation of the i915 DRM driver

Switch the i915 device aperture mapping to the io-mapping interface, taking
advantage of the cleaner API to extend it across all of the mapping uses,
including both pwrite and relocation updates.

This dramatically improves performance on 64-bit kernels which were using
the same slow path as 32-bit non-HIGHMEM kernels prior to this patch.

Signed-off-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem.c

index f20ffe1..126b2f1 100644 (file)
@@ -31,6 +31,7 @@
 #define _I915_DRV_H_
 
 #include "i915_reg.h"
+#include <linux/io-mapping.h>
 
 /* General customization:
  */
@@ -246,6 +247,8 @@ typedef struct drm_i915_private {
        struct {
                struct drm_mm gtt_space;
 
+               struct io_mapping *gtt_mapping;
+
                /**
                 * List of objects currently involved in rendering from the
                 * ringbuffer.
index 17ae330..61183b9 100644 (file)
@@ -171,35 +171,50 @@ i915_gem_pread_ioctl(struct drm_device *dev, void *data,
        return 0;
 }
 
-/*
- * Try to write quickly with an atomic kmap. Return true on success.
- *
- * If this fails (which includes a partial write), we'll redo the whole
- * thing with the slow version.
- *
- * This is a workaround for the low performance of iounmap (approximate
- * 10% cpu cost on normal 3D workloads).  kmap_atomic on HIGHMEM kernels
- * happens to let us map card memory without taking IPIs.  When the vmap
- * rework lands we should be able to dump this hack.
+/* This is the fast write path which cannot handle
+ * page faults in the source data
  */
-static inline int fast_user_write(unsigned long pfn, char __user *user_data,
-                                 int l, int o)
+
+static inline int
+fast_user_write(struct io_mapping *mapping,
+               loff_t page_base, int page_offset,
+               char __user *user_data,
+               int length)
 {
-#ifdef CONFIG_HIGHMEM
-       unsigned long unwritten;
        char *vaddr_atomic;
+       unsigned long unwritten;
 
-       vaddr_atomic = kmap_atomic_pfn(pfn, KM_USER0);
-#if WATCH_PWRITE
-       DRM_INFO("pwrite i %d o %d l %d pfn %ld vaddr %p\n",
-                i, o, l, pfn, vaddr_atomic);
-#endif
-       unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + o, user_data, l);
-       kunmap_atomic(vaddr_atomic, KM_USER0);
-       return !unwritten;
-#else
+       vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
+       unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
+                                                     user_data, length);
+       io_mapping_unmap_atomic(vaddr_atomic);
+       if (unwritten)
+               return -EFAULT;
+       return 0;
+}
+
+/* Here's the write path which can sleep for
+ * page faults
+ */
+
+static inline int
+slow_user_write(struct io_mapping *mapping,
+               loff_t page_base, int page_offset,
+               char __user *user_data,
+               int length)
+{
+       char __iomem *vaddr;
+       unsigned long unwritten;
+
+       vaddr = io_mapping_map_wc(mapping, page_base);
+       if (vaddr == NULL)
+               return -EFAULT;
+       unwritten = __copy_from_user(vaddr + page_offset,
+                                    user_data, length);
+       io_mapping_unmap(vaddr);
+       if (unwritten)
+               return -EFAULT;
        return 0;
-#endif
 }
 
 static int
@@ -208,10 +223,12 @@ i915_gem_gtt_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
                    struct drm_file *file_priv)
 {
        struct drm_i915_gem_object *obj_priv = obj->driver_private;
+       drm_i915_private_t *dev_priv = dev->dev_private;
        ssize_t remain;
-       loff_t offset;
+       loff_t offset, page_base;
        char __user *user_data;
-       int ret = 0;
+       int page_offset, page_length;
+       int ret;
 
        user_data = (char __user *) (uintptr_t) args->data_ptr;
        remain = args->size;
@@ -235,57 +252,37 @@ i915_gem_gtt_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
        obj_priv->dirty = 1;
 
        while (remain > 0) {
-               unsigned long pfn;
-               int i, o, l;
-
                /* Operation in this page
                 *
-                * i = page number
-                * o = offset within page
-                * l = bytes to copy
+                * page_base = page offset within aperture
+                * page_offset = offset within page
+                * page_length = bytes to copy for this page
                 */
-               i = offset >> PAGE_SHIFT;
-               o = offset & (PAGE_SIZE-1);
-               l = remain;
-               if ((o + l) > PAGE_SIZE)
-                       l = PAGE_SIZE - o;
-
-               pfn = (dev->agp->base >> PAGE_SHIFT) + i;
-
-               if (!fast_user_write(pfn, user_data, l, o)) {
-                       unsigned long unwritten;
-                       char __iomem *vaddr;
-
-                       vaddr = ioremap_wc(pfn << PAGE_SHIFT, PAGE_SIZE);
-#if WATCH_PWRITE
-                       DRM_INFO("pwrite slow i %d o %d l %d "
-                                "pfn %ld vaddr %p\n",
-                                i, o, l, pfn, vaddr);
-#endif
-                       if (vaddr == NULL) {
-                               ret = -EFAULT;
-                               goto fail;
-                       }
-                       unwritten = __copy_from_user(vaddr + o, user_data, l);
-#if WATCH_PWRITE
-                       DRM_INFO("unwritten %ld\n", unwritten);
-#endif
-                       iounmap(vaddr);
-                       if (unwritten) {
-                               ret = -EFAULT;
+               page_base = (offset & ~(PAGE_SIZE-1));
+               page_offset = offset & (PAGE_SIZE-1);
+               page_length = remain;
+               if ((page_offset + remain) > PAGE_SIZE)
+                       page_length = PAGE_SIZE - page_offset;
+
+               ret = fast_user_write (dev_priv->mm.gtt_mapping, page_base,
+                                      page_offset, user_data, page_length);
+
+               /* If we get a fault while copying data, then (presumably) our
+                * source page isn't available. In this case, use the
+                * non-atomic function
+                */
+               if (ret) {
+                       ret = slow_user_write (dev_priv->mm.gtt_mapping,
+                                              page_base, page_offset,
+                                              user_data, page_length);
+                       if (ret)
                                goto fail;
-                       }
                }
 
-               remain -= l;
-               user_data += l;
-               offset += l;
+               remain -= page_length;
+               user_data += page_length;
+               offset += page_length;
        }
-#if WATCH_PWRITE && 1
-       i915_gem_clflush_object(obj);
-       i915_gem_dump_object(obj, args->offset + args->size, __func__, ~0);
-       i915_gem_clflush_object(obj);
-#endif
 
 fail:
        i915_gem_object_unpin(obj);
@@ -1503,12 +1500,12 @@ i915_gem_object_pin_and_relocate(struct drm_gem_object *obj,
                                 struct drm_i915_gem_exec_object *entry)
 {
        struct drm_device *dev = obj->dev;
+       drm_i915_private_t *dev_priv = dev->dev_private;
        struct drm_i915_gem_relocation_entry reloc;
        struct drm_i915_gem_relocation_entry __user *relocs;
        struct drm_i915_gem_object *obj_priv = obj->driver_private;
        int i, ret;
-       uint32_t last_reloc_offset = -1;
-       void __iomem *reloc_page = NULL;
+       void __iomem *reloc_page;
 
        /* Choose the GTT offset for our buffer and put it there. */
        ret = i915_gem_object_pin(obj, (uint32_t) entry->alignment);
@@ -1631,26 +1628,11 @@ i915_gem_object_pin_and_relocate(struct drm_gem_object *obj,
                 * perform.
                 */
                reloc_offset = obj_priv->gtt_offset + reloc.offset;
-               if (reloc_page == NULL ||
-                   (last_reloc_offset & ~(PAGE_SIZE - 1)) !=
-                   (reloc_offset & ~(PAGE_SIZE - 1))) {
-                       if (reloc_page != NULL)
-                               iounmap(reloc_page);
-
-                       reloc_page = ioremap_wc(dev->agp->base +
-                                               (reloc_offset &
-                                                ~(PAGE_SIZE - 1)),
-                                               PAGE_SIZE);
-                       last_reloc_offset = reloc_offset;
-                       if (reloc_page == NULL) {
-                               drm_gem_object_unreference(target_obj);
-                               i915_gem_object_unpin(obj);
-                               return -ENOMEM;
-                       }
-               }
-
+               reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
+                                                     (reloc_offset &
+                                                      ~(PAGE_SIZE - 1)));
                reloc_entry = (uint32_t __iomem *)(reloc_page +
-                                          (reloc_offset & (PAGE_SIZE - 1)));
+                                                  (reloc_offset & (PAGE_SIZE - 1)));
                reloc_val = target_obj_priv->gtt_offset + reloc.delta;
 
 #if WATCH_BUF
@@ -1659,6 +1641,7 @@ i915_gem_object_pin_and_relocate(struct drm_gem_object *obj,
                          readl(reloc_entry), reloc_val);
 #endif
                writel(reloc_val, reloc_entry);
+               io_mapping_unmap_atomic(reloc_page);
 
                /* Write the updated presumed offset for this entry back out
                 * to the user.
@@ -1674,9 +1657,6 @@ i915_gem_object_pin_and_relocate(struct drm_gem_object *obj,
                drm_gem_object_unreference(target_obj);
        }
 
-       if (reloc_page != NULL)
-               iounmap(reloc_page);
-
 #if WATCH_BUF
        if (0)
                i915_gem_dump_object(obj, 128, __func__, ~0);
@@ -2518,6 +2498,10 @@ i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
        if (ret != 0)
                return ret;
 
+       dev_priv->mm.gtt_mapping = io_mapping_create_wc(dev->agp->base,
+                                                       dev->agp->agp_info.aper_size
+                                                       * 1024 * 1024);
+
        mutex_lock(&dev->struct_mutex);
        BUG_ON(!list_empty(&dev_priv->mm.active_list));
        BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
@@ -2535,11 +2519,13 @@ int
 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
                       struct drm_file *file_priv)
 {
+       drm_i915_private_t *dev_priv = dev->dev_private;
        int ret;
 
        ret = i915_gem_idle(dev);
        drm_irq_uninstall(dev);
 
+       io_mapping_free(dev_priv->mm.gtt_mapping);
        return ret;
 }