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>
---
drivers/gpu/drm/i915/i915_drv.h | 3 +
drivers/gpu/drm/i915/i915_gem.c | 81 ++++++++++++++++-----------------------
2 files changed, 36 insertions(+), 48 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index f20ffe1..8ca5fbc 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -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 *io_mapping;
+
/**
* List of objects currently involved in rendering from the
* ringbuffer.
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 9255088..d38b052 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -177,14 +177,14 @@ 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, base;
char __user *user_data;
char __iomem *vaddr;
char *vaddr_atomic;
- int i, o, l;
+ int o, l;
int ret = 0;
- unsigned long pfn;
unsigned long unwritten;
user_data = (char __user *) (uintptr_t) args->data_ptr;
@@ -211,42 +211,38 @@ i915_gem_gtt_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
while (remain > 0) {
/* Operation in this page
*
- * i = page number
+ * base = page offset within aperture
* o = offset within page
* l = bytes to copy
*/
- i = offset >> PAGE_SHIFT;
+ base = (offset & ~(PAGE_SIZE-1));
o = offset & (PAGE_SIZE-1);
l = remain;
if ((o + l) > PAGE_SIZE)
l = PAGE_SIZE - o;
- pfn = (dev->agp->base >> PAGE_SHIFT) + i;
-
-#ifdef CONFIG_HIGHMEM
/* 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.
+ * io_mapping_map_atomic_wc maps card memory
+ * without taking IPIs.
*/
- 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
+ vaddr_atomic = io_mapping_map_atomic_wc(dev_priv->mm.io_mapping,
+ base);
unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + o,
user_data, l);
- kunmap_atomic(vaddr_atomic, KM_USER0);
-
+ io_mapping_unmap_atomic(vaddr_atomic);
+
+ /* If we get a fault while copying data, then (presumably) our
+ * source page isn't available. In this case, use the
+ * non-atomic __copy_from_user function
+ */
if (unwritten)
-#endif /* CONFIG_HIGHMEM */
{
- vaddr = ioremap_wc(pfn << PAGE_SHIFT, PAGE_SIZE);
+ vaddr = io_mapping_map_wc(dev_priv->mm.io_mapping, base);
#if WATCH_PWRITE
- DRM_INFO("pwrite slow i %d o %d l %d "
- "pfn %ld vaddr %p\n",
- i, o, l, pfn, vaddr);
+ DRM_INFO("pwrite slow base %ld o %d l %d "
+ "vaddr %p\n",
+ base, o, l, vaddr);
#endif
if (vaddr == NULL) {
ret = -EFAULT;
@@ -256,7 +252,7 @@ i915_gem_gtt_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
#if WATCH_PWRITE
DRM_INFO("unwritten %ld\n", unwritten);
#endif
- iounmap(vaddr);
+ io_mapping_unmap(vaddr);
if (unwritten) {
ret = -EFAULT;
goto fail;
@@ -1489,12 +1485,12 @@ 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);
@@ -1617,26 +1613,11 @@ 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.io_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
@@ -1645,6 +1626,7 @@ 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.
@@ -1660,9 +1642,6 @@ 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);
@@ -2504,6 +2483,10 @@ i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
if (ret != 0)
return ret;
+ dev_priv->mm.io_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));
@@ -2521,11 +2504,13 @@ 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.io_mapping);
return ret;
}
--
1.5.6.5
--
