diff options
author | Chris Wilson <chris@chris-wilson.co.uk> | 2010-08-07 06:01:24 -0400 |
---|---|---|
committer | Eric Anholt <eric@anholt.net> | 2010-08-09 14:24:32 -0400 |
commit | cd377ea93f34cbd6ec49c868b66a5a7ab184775c (patch) | |
tree | 2a9a09f5b41ca61150a924925269141b75112244 /drivers/gpu | |
parent | b47eb4a2b302f33adaed2a27d2b3bfc74fe35ac5 (diff) |
drm/i915: Implement fair lru eviction across both rings. (v2)
Based in a large part upon Daniel Vetter's implementation and adapted
for handling multiple rings in a single pass.
This should lead to better gtt usage and fixes the page-fault-of-doom
triggered. The fairness is provided by scanning through the GTT space
amalgamating space in rendering order. As soon as we have a contiguous
space in the GTT large enough for the new object (and its alignment),
evict any object which lies within that space. This should keep more
objects resident in the GTT.
Doing throughput testing on a PineView machine with cairo-perf-trace
indicates that there is very little difference with the new LRU scan,
perhaps a small improvement... Except oddly for the poppler trace.
Reference:
Bug 15911 - Intermittent X crash (freeze)
https://bugzilla.kernel.org/show_bug.cgi?id=15911
Bug 20152 - cannot view JPG in firefox when running UXA
https://bugs.freedesktop.org/show_bug.cgi?id=20152
Bug 24369 - Hang when scrolling firefox page with window in front
https://bugs.freedesktop.org/show_bug.cgi?id=24369
Bug 28478 - Intermittent graphics lockups due to overflow/loop
https://bugs.freedesktop.org/show_bug.cgi?id=28478
v2: Attempt to clarify the logic and order of eviction through the use
of comments and macros.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Signed-off-by: Eric Anholt <eric@anholt.net>
Diffstat (limited to 'drivers/gpu')
-rw-r--r-- | drivers/gpu/drm/i915/i915_drv.h | 2 | ||||
-rw-r--r-- | drivers/gpu/drm/i915/i915_gem_evict.c | 277 |
2 files changed, 146 insertions, 133 deletions
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 12c8f47f984b..6221f239fa5e 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h | |||
@@ -673,6 +673,8 @@ struct drm_i915_gem_object { | |||
673 | struct list_head list; | 673 | struct list_head list; |
674 | /** This object's place on GPU write list */ | 674 | /** This object's place on GPU write list */ |
675 | struct list_head gpu_write_list; | 675 | struct list_head gpu_write_list; |
676 | /** This object's place on eviction list */ | ||
677 | struct list_head evict_list; | ||
676 | 678 | ||
677 | /** | 679 | /** |
678 | * This is set if the object is on the active or flushing lists | 680 | * This is set if the object is on the active or flushing lists |
diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c index 479e450f931b..72cae3cccad8 100644 --- a/drivers/gpu/drm/i915/i915_gem_evict.c +++ b/drivers/gpu/drm/i915/i915_gem_evict.c | |||
@@ -31,167 +31,178 @@ | |||
31 | #include "i915_drv.h" | 31 | #include "i915_drv.h" |
32 | #include "i915_drm.h" | 32 | #include "i915_drm.h" |
33 | 33 | ||
34 | static inline int | 34 | static struct drm_i915_gem_object * |
35 | i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj_priv) | 35 | i915_gem_next_active_object(struct drm_device *dev, |
36 | { | 36 | struct list_head **render_iter, |
37 | return obj_priv->madv == I915_MADV_DONTNEED; | 37 | struct list_head **bsd_iter) |
38 | } | ||
39 | |||
40 | static int | ||
41 | i915_gem_scan_inactive_list_and_evict(struct drm_device *dev, int min_size, | ||
42 | unsigned alignment, int *found) | ||
43 | { | 38 | { |
44 | drm_i915_private_t *dev_priv = dev->dev_private; | 39 | drm_i915_private_t *dev_priv = dev->dev_private; |
45 | struct drm_gem_object *obj; | 40 | struct drm_i915_gem_object *render_obj = NULL, *bsd_obj = NULL; |
46 | struct drm_i915_gem_object *obj_priv; | ||
47 | struct drm_gem_object *best = NULL; | ||
48 | struct drm_gem_object *first = NULL; | ||
49 | |||
50 | /* Try to find the smallest clean object */ | ||
51 | list_for_each_entry(obj_priv, &dev_priv->mm.inactive_list, list) { | ||
52 | struct drm_gem_object *obj = &obj_priv->base; | ||
53 | if (obj->size >= min_size) { | ||
54 | if ((!obj_priv->dirty || | ||
55 | i915_gem_object_is_purgeable(obj_priv)) && | ||
56 | (!best || obj->size < best->size)) { | ||
57 | best = obj; | ||
58 | if (best->size == min_size) | ||
59 | break; | ||
60 | } | ||
61 | if (!first) | ||
62 | first = obj; | ||
63 | } | ||
64 | } | ||
65 | 41 | ||
66 | obj = best ? best : first; | 42 | if (*render_iter != &dev_priv->render_ring.active_list) |
43 | render_obj = list_entry(*render_iter, | ||
44 | struct drm_i915_gem_object, | ||
45 | list); | ||
67 | 46 | ||
68 | if (!obj) { | 47 | if (HAS_BSD(dev)) { |
69 | *found = 0; | 48 | if (*bsd_iter != &dev_priv->bsd_ring.active_list) |
70 | return 0; | 49 | bsd_obj = list_entry(*bsd_iter, |
71 | } | 50 | struct drm_i915_gem_object, |
51 | list); | ||
72 | 52 | ||
73 | *found = 1; | 53 | if (render_obj == NULL) { |
54 | *bsd_iter = (*bsd_iter)->next; | ||
55 | return bsd_obj; | ||
56 | } | ||
74 | 57 | ||
75 | #if WATCH_LRU | 58 | if (bsd_obj == NULL) { |
76 | DRM_INFO("%s: evicting %p\n", __func__, obj); | 59 | *render_iter = (*render_iter)->next; |
77 | #endif | 60 | return render_obj; |
78 | obj_priv = to_intel_bo(obj); | 61 | } |
79 | BUG_ON(obj_priv->pin_count != 0); | ||
80 | BUG_ON(obj_priv->active); | ||
81 | 62 | ||
82 | /* Wait on the rendering and unbind the buffer. */ | 63 | /* XXX can we handle seqno wrapping? */ |
83 | return i915_gem_object_unbind(obj); | 64 | if (render_obj->last_rendering_seqno < bsd_obj->last_rendering_seqno) { |
65 | *render_iter = (*render_iter)->next; | ||
66 | return render_obj; | ||
67 | } else { | ||
68 | *bsd_iter = (*bsd_iter)->next; | ||
69 | return bsd_obj; | ||
70 | } | ||
71 | } else { | ||
72 | *render_iter = (*render_iter)->next; | ||
73 | return render_obj; | ||
74 | } | ||
84 | } | 75 | } |
85 | 76 | ||
86 | static void | 77 | static bool |
87 | i915_gem_flush_ring(struct drm_device *dev, | 78 | mark_free(struct drm_i915_gem_object *obj_priv, |
88 | uint32_t invalidate_domains, | 79 | struct list_head *unwind) |
89 | uint32_t flush_domains, | ||
90 | struct intel_ring_buffer *ring) | ||
91 | { | 80 | { |
92 | if (flush_domains & I915_GEM_DOMAIN_CPU) | 81 | list_add(&obj_priv->evict_list, unwind); |
93 | drm_agp_chipset_flush(dev); | 82 | return drm_mm_scan_add_block(obj_priv->gtt_space); |
94 | ring->flush(dev, ring, | ||
95 | invalidate_domains, | ||
96 | flush_domains); | ||
97 | } | 83 | } |
98 | 84 | ||
85 | #define i915_for_each_active_object(OBJ, R, B) \ | ||
86 | *(R) = dev_priv->render_ring.active_list.next; \ | ||
87 | *(B) = dev_priv->bsd_ring.active_list.next; \ | ||
88 | while (((OBJ) = i915_gem_next_active_object(dev, (R), (B))) != NULL) | ||
89 | |||
99 | int | 90 | int |
100 | i915_gem_evict_something(struct drm_device *dev, | 91 | i915_gem_evict_something(struct drm_device *dev, int min_size, unsigned alignment) |
101 | int min_size, unsigned alignment) | ||
102 | { | 92 | { |
103 | drm_i915_private_t *dev_priv = dev->dev_private; | 93 | drm_i915_private_t *dev_priv = dev->dev_private; |
104 | int ret, found; | 94 | struct list_head eviction_list, unwind_list; |
105 | 95 | struct drm_i915_gem_object *obj_priv, *tmp_obj_priv; | |
106 | struct intel_ring_buffer *render_ring = &dev_priv->render_ring; | 96 | struct list_head *render_iter, *bsd_iter; |
107 | struct intel_ring_buffer *bsd_ring = &dev_priv->bsd_ring; | 97 | int ret = 0; |
108 | for (;;) { | ||
109 | i915_gem_retire_requests(dev); | ||
110 | |||
111 | /* If there's an inactive buffer available now, grab it | ||
112 | * and be done. | ||
113 | */ | ||
114 | ret = i915_gem_scan_inactive_list_and_evict(dev, min_size, | ||
115 | alignment, | ||
116 | &found); | ||
117 | if (found) | ||
118 | return ret; | ||
119 | 98 | ||
120 | /* If we didn't get anything, but the ring is still processing | 99 | i915_gem_retire_requests(dev); |
121 | * things, wait for the next to finish and hopefully leave us | ||
122 | * a buffer to evict. | ||
123 | */ | ||
124 | if (!list_empty(&render_ring->request_list)) { | ||
125 | struct drm_i915_gem_request *request; | ||
126 | 100 | ||
127 | request = list_first_entry(&render_ring->request_list, | 101 | /* Re-check for free space after retiring requests */ |
128 | struct drm_i915_gem_request, | 102 | if (drm_mm_search_free(&dev_priv->mm.gtt_space, |
129 | list); | 103 | min_size, alignment, 0)) |
104 | return 0; | ||
130 | 105 | ||
131 | ret = i915_do_wait_request(dev, request->seqno, true, request->ring); | 106 | /* |
132 | if (ret) | 107 | * The goal is to evict objects and amalgamate space in LRU order. |
133 | return ret; | 108 | * The oldest idle objects reside on the inactive list, which is in |
109 | * retirement order. The next objects to retire are those on the (per | ||
110 | * ring) active list that do not have an outstanding flush. Once the | ||
111 | * hardware reports completion (the seqno is updated after the | ||
112 | * batchbuffer has been finished) the clean buffer objects would | ||
113 | * be retired to the inactive list. Any dirty objects would be added | ||
114 | * to the tail of the flushing list. So after processing the clean | ||
115 | * active objects we need to emit a MI_FLUSH to retire the flushing | ||
116 | * list, hence the retirement order of the flushing list is in | ||
117 | * advance of the dirty objects on the active lists. | ||
118 | * | ||
119 | * The retirement sequence is thus: | ||
120 | * 1. Inactive objects (already retired) | ||
121 | * 2. Clean active objects | ||
122 | * 3. Flushing list | ||
123 | * 4. Dirty active objects. | ||
124 | * | ||
125 | * On each list, the oldest objects lie at the HEAD with the freshest | ||
126 | * object on the TAIL. | ||
127 | */ | ||
128 | |||
129 | INIT_LIST_HEAD(&unwind_list); | ||
130 | drm_mm_init_scan(&dev_priv->mm.gtt_space, min_size, alignment); | ||
131 | |||
132 | /* First see if there is a large enough contiguous idle region... */ | ||
133 | list_for_each_entry(obj_priv, &dev_priv->mm.inactive_list, list) { | ||
134 | if (mark_free(obj_priv, &unwind_list)) | ||
135 | goto found; | ||
136 | } | ||
134 | 137 | ||
138 | /* Now merge in the soon-to-be-expired objects... */ | ||
139 | i915_for_each_active_object(obj_priv, &render_iter, &bsd_iter) { | ||
140 | /* Does the object require an outstanding flush? */ | ||
141 | if (obj_priv->base.write_domain || obj_priv->pin_count) | ||
135 | continue; | 142 | continue; |
136 | } | ||
137 | 143 | ||
138 | if (HAS_BSD(dev) && !list_empty(&bsd_ring->request_list)) { | 144 | if (mark_free(obj_priv, &unwind_list)) |
139 | struct drm_i915_gem_request *request; | 145 | goto found; |
140 | 146 | } | |
141 | request = list_first_entry(&bsd_ring->request_list, | ||
142 | struct drm_i915_gem_request, | ||
143 | list); | ||
144 | 147 | ||
145 | ret = i915_do_wait_request(dev, request->seqno, true, request->ring); | 148 | /* Finally add anything with a pending flush (in order of retirement) */ |
146 | if (ret) | 149 | list_for_each_entry(obj_priv, &dev_priv->mm.flushing_list, list) { |
147 | return ret; | 150 | if (obj_priv->pin_count) |
151 | continue; | ||
148 | 152 | ||
153 | if (mark_free(obj_priv, &unwind_list)) | ||
154 | goto found; | ||
155 | } | ||
156 | i915_for_each_active_object(obj_priv, &render_iter, &bsd_iter) { | ||
157 | if (! obj_priv->base.write_domain || obj_priv->pin_count) | ||
149 | continue; | 158 | continue; |
150 | } | ||
151 | 159 | ||
152 | /* If we didn't have anything on the request list but there | 160 | if (mark_free(obj_priv, &unwind_list)) |
153 | * are buffers awaiting a flush, emit one and try again. | 161 | goto found; |
154 | * When we wait on it, those buffers waiting for that flush | 162 | } |
155 | * will get moved to inactive. | 163 | |
156 | */ | 164 | /* Nothing found, clean up and bail out! */ |
157 | if (!list_empty(&dev_priv->mm.flushing_list)) { | 165 | list_for_each_entry(obj_priv, &unwind_list, evict_list) { |
158 | struct drm_gem_object *obj = NULL; | 166 | ret = drm_mm_scan_remove_block(obj_priv->gtt_space); |
159 | struct drm_i915_gem_object *obj_priv; | 167 | BUG_ON(ret); |
160 | 168 | } | |
161 | /* Find an object that we can immediately reuse */ | 169 | |
162 | list_for_each_entry(obj_priv, &dev_priv->mm.flushing_list, list) { | 170 | /* We expect the caller to unpin, evict all and try again, or give up. |
163 | obj = &obj_priv->base; | 171 | * So calling i915_gem_evict_everything() is unnecessary. |
164 | if (obj->size >= min_size) | 172 | */ |
165 | break; | 173 | return -ENOSPC; |
166 | 174 | ||
167 | obj = NULL; | 175 | found: |
168 | } | 176 | INIT_LIST_HEAD(&eviction_list); |
169 | 177 | list_for_each_entry_safe(obj_priv, tmp_obj_priv, | |
170 | if (obj != NULL) { | 178 | &unwind_list, evict_list) { |
171 | uint32_t seqno; | 179 | if (drm_mm_scan_remove_block(obj_priv->gtt_space)) { |
172 | 180 | /* drm_mm doesn't allow any other other operations while | |
173 | i915_gem_flush_ring(dev, | 181 | * scanning, therefore store to be evicted objects on a |
174 | obj->write_domain, | 182 | * temporary list. */ |
175 | obj->write_domain, | 183 | list_move(&obj_priv->evict_list, &eviction_list); |
176 | obj_priv->ring); | ||
177 | seqno = i915_add_request(dev, NULL, | ||
178 | obj->write_domain, | ||
179 | obj_priv->ring); | ||
180 | if (seqno == 0) | ||
181 | return -ENOMEM; | ||
182 | continue; | ||
183 | } | ||
184 | } | 184 | } |
185 | } | ||
185 | 186 | ||
186 | /* If we didn't do any of the above, there's no single buffer | 187 | /* Unbinding will emit any required flushes */ |
187 | * large enough to swap out for the new one, so just evict | 188 | list_for_each_entry_safe(obj_priv, tmp_obj_priv, |
188 | * everything and start again. (This should be rare.) | 189 | &eviction_list, evict_list) { |
189 | */ | 190 | #if WATCH_LRU |
190 | if (!list_empty(&dev_priv->mm.inactive_list)) | 191 | DRM_INFO("%s: evicting %p\n", __func__, obj); |
191 | return i915_gem_evict_inactive(dev); | 192 | #endif |
192 | else | 193 | ret = i915_gem_object_unbind(&obj_priv->base); |
193 | return i915_gem_evict_everything(dev); | 194 | if (ret) |
195 | return ret; | ||
194 | } | 196 | } |
197 | |||
198 | /* The just created free hole should be on the top of the free stack | ||
199 | * maintained by drm_mm, so this BUG_ON actually executes in O(1). | ||
200 | * Furthermore all accessed data has just recently been used, so it | ||
201 | * should be really fast, too. */ | ||
202 | BUG_ON(!drm_mm_search_free(&dev_priv->mm.gtt_space, min_size, | ||
203 | alignment, 0)); | ||
204 | |||
205 | return 0; | ||
195 | } | 206 | } |
196 | 207 | ||
197 | int | 208 | int |