Intel-GFX Archive on lore.kernel.org
 help / color / mirror / Atom feed
From: Matthew Auld <matthew.auld@intel.com>
To: intel-gfx@lists.freedesktop.org
Subject: [PATCH 05/17] drm/i915: fallback to normal pages on vma insert failure
Date: Tue, 16 May 2017 09:29:36 +0100	[thread overview]
Message-ID: <20170516082948.28090-6-matthew.auld@intel.com> (raw)
In-Reply-To: <20170516082948.28090-1-matthew.auld@intel.com>

Part of the cost in choosing huge-gtt-pages is potentially using a
larger alignment and/or size. Therefore if our vma insert fails either
because of the insert/reserve or the pin-offset-fixed we should fallback
to normal pages and retry before giving up.

Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/i915_vma.c | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
index d2e8edd351cf..9d4ffd76184e 100644
--- a/drivers/gpu/drm/i915/i915_vma.c
+++ b/drivers/gpu/drm/i915/i915_vma.c
@@ -427,6 +427,8 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 {
 	struct drm_i915_private *dev_priv = vma->vm->i915;
 	struct drm_i915_gem_object *obj = vma->obj;
+	u64 requested_alignment;
+	u64 requested_size;
 	u64 start, end;
 	int ret;
 
@@ -471,6 +473,9 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 	if (ret)
 		return ret;
 
+	requested_alignment = alignment;
+	requested_size = size;
+
 	if (i915_vm_is_48bit(vma->vm) &&
 	    obj->gtt_page_size > I915_GTT_PAGE_SIZE) {
 		unsigned int page_alignment = obj->gtt_page_size;
@@ -488,6 +493,7 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 		GEM_BUG_ON(!IS_ALIGNED(vma->size, obj->gtt_page_size));
 	}
 
+retry_insert:
 	if (flags & PIN_OFFSET_FIXED) {
 		u64 offset = flags & PIN_OFFSET_MASK;
 		if (!IS_ALIGNED(offset, alignment) ||
@@ -522,6 +528,19 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 	return 0;
 
 err_unpin:
+	/* We try to use huge-gtt-pages whenever we can, but part of the cost
+	 * is that we may need adjust the alignment and possibly the size
+	 * before we insert into a vm, and so we should always fallback and
+	 * retry without huge-gtt-pages if we ever encounter a failure, before
+	 * giving up.
+	 */
+	if (alignment > requested_alignment || size > requested_size) {
+		obj->gtt_page_size = I915_GTT_PAGE_SIZE;
+		alignment = requested_alignment;
+		size = requested_size;
+		goto retry_insert;
+	}
+
 	i915_gem_object_unpin_pages(obj);
 	return ret;
 }
-- 
2.9.4

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

  parent reply	other threads:[~2017-05-16  8:29 UTC|newest]

Thread overview: 32+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2017-05-16  8:29 [PATCH 00/17] add support for huge-gtt-pages Matthew Auld
2017-05-16  8:29 ` [PATCH 01/17] drm/i915: introduce page_size_mask to dev_info Matthew Auld
2017-05-16  8:29 ` [PATCH 02/17] drm/i915: introduce gtt page size Matthew Auld
2017-05-16  8:41   ` Chris Wilson
2017-05-16  9:59   ` Chris Wilson
2017-05-23 12:42     ` Matthew Auld
2017-05-23 12:54       ` Chris Wilson
2017-05-23 13:57         ` Matthew Auld
2017-05-23 14:30           ` Chris Wilson
2017-05-16  8:29 ` [PATCH 03/17] drm/i915: align the vma start to the " Matthew Auld
2017-05-16  8:40   ` Chris Wilson
2017-05-16  8:29 ` [PATCH 04/17] drm/i915: align 64K objects to 2M Matthew Auld
2017-05-16  8:29 ` Matthew Auld [this message]
2017-05-16  8:39   ` [PATCH 05/17] drm/i915: fallback to normal pages on vma insert failure Chris Wilson
2017-05-16  8:29 ` [PATCH 06/17] mm/shmem: expose driver overridable huge option Matthew Auld
2017-05-16 10:02   ` Kirill A. Shutemov
2017-05-16  8:29 ` [PATCH 07/17] drm/i915: request THP for shmem backed objects Matthew Auld
2017-05-16  8:29 ` [PATCH 08/17] drm/i915: pass gtt page size to insert_entries Matthew Auld
2017-05-16  8:29 ` [PATCH 09/17] drm/i915: enable IPS bit for 64K pages Matthew Auld
2017-05-16  8:29 ` [PATCH 10/17] drm/i915: support inserting 64K pages into the 48b PPGTT Matthew Auld
2017-05-16  8:29 ` [PATCH 11/17] drm/i915: disable GTT cache for 2M/1G pages Matthew Auld
2017-05-16 10:04   ` Ville Syrjälä
2017-05-16 10:11     ` Chris Wilson
2017-05-16  8:29 ` [PATCH 12/17] drm/i915: support inserting 2M pages into the 48b PPGTT Matthew Auld
2017-05-16  8:29 ` [PATCH 13/17] drm/i915: support inserting 1G " Matthew Auld
2017-05-16  8:29 ` [PATCH 14/17] drm/i915/debugfs: include some gtt_page_size metrics Matthew Auld
2017-05-19 10:48   ` Chris Wilson
2017-05-19 10:51   ` Chris Wilson
2017-05-16  8:29 ` [PATCH 15/17] drm/i915: enable platform support for 64K pages Matthew Auld
2017-05-16  8:29 ` [PATCH 16/17] drm/i915: enable platform support for 2M pages Matthew Auld
2017-05-16  8:29 ` [PATCH 17/17] drm/i915: enable platform support for 1G pages Matthew Auld
2017-05-16  8:49 ` ✓ Fi.CI.BAT: success for add support for huge-gtt-pages Patchwork

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20170516082948.28090-6-matthew.auld@intel.com \
    --to=matthew.auld@intel.com \
    --cc=intel-gfx@lists.freedesktop.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox