1 /*
2  * Copyright © 2008-2010 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  * Authors:
24  *    Eric Anholt <eric@anholt.net>
25  *    Chris Wilson <chris@chris-wilson.co.uuk>
26  *
27  */
28 
29 #include "gem/i915_gem_context.h"
30 #include "gt/intel_gt_requests.h"
31 
32 #include "i915_drv.h"
33 #include "i915_trace.h"
34 
35 I915_SELFTEST_DECLARE(static struct igt_evict_ctl {
36 	bool fail_if_busy:1;
37 } igt_evict_ctl;)
38 
39 static int ggtt_flush(struct intel_gt *gt)
40 {
41 	/*
42 	 * Not everything in the GGTT is tracked via vma (otherwise we
43 	 * could evict as required with minimal stalling) so we are forced
44 	 * to idle the GPU and explicitly retire outstanding requests in
45 	 * the hopes that we can then remove contexts and the like only
46 	 * bound by their active reference.
47 	 */
48 	return intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
49 }
50 
51 static bool
52 mark_free(struct drm_mm_scan *scan,
53 	  struct i915_vma *vma,
54 	  unsigned int flags,
55 	  struct list_head *unwind)
56 {
57 	if (i915_vma_is_pinned(vma))
58 		return false;
59 
60 	list_add(&vma->evict_link, unwind);
61 	return drm_mm_scan_add_block(scan, &vma->node);
62 }
63 
64 /**
65  * i915_gem_evict_something - Evict vmas to make room for binding a new one
66  * @vm: address space to evict from
67  * @min_size: size of the desired free space
68  * @alignment: alignment constraint of the desired free space
69  * @color: color for the desired space
70  * @start: start (inclusive) of the range from which to evict objects
71  * @end: end (exclusive) of the range from which to evict objects
72  * @flags: additional flags to control the eviction algorithm
73  *
74  * This function will try to evict vmas until a free space satisfying the
75  * requirements is found. Callers must check first whether any such hole exists
76  * already before calling this function.
77  *
78  * This function is used by the object/vma binding code.
79  *
80  * Since this function is only used to free up virtual address space it only
81  * ignores pinned vmas, and not object where the backing storage itself is
82  * pinned. Hence obj->pages_pin_count does not protect against eviction.
83  *
84  * To clarify: This is for freeing up virtual address space, not for freeing
85  * memory in e.g. the shrinker.
86  */
87 int
88 i915_gem_evict_something(struct i915_address_space *vm,
89 			 u64 min_size, u64 alignment,
90 			 unsigned long color,
91 			 u64 start, u64 end,
92 			 unsigned flags)
93 {
94 	struct drm_mm_scan scan;
95 	struct list_head eviction_list;
96 	struct i915_vma *vma, *next;
97 	struct drm_mm_node *node;
98 	enum drm_mm_insert_mode mode;
99 	struct i915_vma *active;
100 	int ret;
101 
102 	lockdep_assert_held(&vm->mutex);
103 	trace_i915_gem_evict(vm, min_size, alignment, flags);
104 
105 	/*
106 	 * The goal is to evict objects and amalgamate space in rough LRU order.
107 	 * Since both active and inactive objects reside on the same list,
108 	 * in a mix of creation and last scanned order, as we process the list
109 	 * we sort it into inactive/active, which keeps the active portion
110 	 * in a rough MRU order.
111 	 *
112 	 * The retirement sequence is thus:
113 	 *   1. Inactive objects (already retired, random order)
114 	 *   2. Active objects (will stall on unbinding, oldest scanned first)
115 	 */
116 	mode = DRM_MM_INSERT_BEST;
117 	if (flags & PIN_HIGH)
118 		mode = DRM_MM_INSERT_HIGH;
119 	if (flags & PIN_MAPPABLE)
120 		mode = DRM_MM_INSERT_LOW;
121 	drm_mm_scan_init_with_range(&scan, &vm->mm,
122 				    min_size, alignment, color,
123 				    start, end, mode);
124 
125 	intel_gt_retire_requests(vm->gt);
126 
127 search_again:
128 	active = NULL;
129 	INIT_LIST_HEAD(&eviction_list);
130 	list_for_each_entry_safe(vma, next, &vm->bound_list, vm_link) {
131 		/*
132 		 * We keep this list in a rough least-recently scanned order
133 		 * of active elements (inactive elements are cheap to reap).
134 		 * New entries are added to the end, and we move anything we
135 		 * scan to the end. The assumption is that the working set
136 		 * of applications is either steady state (and thanks to the
137 		 * userspace bo cache it almost always is) or volatile and
138 		 * frequently replaced after a frame, which are self-evicting!
139 		 * Given that assumption, the MRU order of the scan list is
140 		 * fairly static, and keeping it in least-recently scan order
141 		 * is suitable.
142 		 *
143 		 * To notice when we complete one full cycle, we record the
144 		 * first active element seen, before moving it to the tail.
145 		 */
146 		if (i915_vma_is_active(vma)) {
147 			if (vma == active) {
148 				if (flags & PIN_NONBLOCK)
149 					break;
150 
151 				active = ERR_PTR(-EAGAIN);
152 			}
153 
154 			if (active != ERR_PTR(-EAGAIN)) {
155 				if (!active)
156 					active = vma;
157 
158 				list_move_tail(&vma->vm_link, &vm->bound_list);
159 				continue;
160 			}
161 		}
162 
163 		if (mark_free(&scan, vma, flags, &eviction_list))
164 			goto found;
165 	}
166 
167 	/* Nothing found, clean up and bail out! */
168 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
169 		ret = drm_mm_scan_remove_block(&scan, &vma->node);
170 		BUG_ON(ret);
171 	}
172 
173 	/*
174 	 * Can we unpin some objects such as idle hw contents,
175 	 * or pending flips? But since only the GGTT has global entries
176 	 * such as scanouts, rinbuffers and contexts, we can skip the
177 	 * purge when inspecting per-process local address spaces.
178 	 */
179 	if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK)
180 		return -ENOSPC;
181 
182 	/*
183 	 * Not everything in the GGTT is tracked via VMA using
184 	 * i915_vma_move_to_active(), otherwise we could evict as required
185 	 * with minimal stalling. Instead we are forced to idle the GPU and
186 	 * explicitly retire outstanding requests which will then remove
187 	 * the pinning for active objects such as contexts and ring,
188 	 * enabling us to evict them on the next iteration.
189 	 *
190 	 * To ensure that all user contexts are evictable, we perform
191 	 * a switch to the perma-pinned kernel context. This all also gives
192 	 * us a termination condition, when the last retired context is
193 	 * the kernel's there is no more we can evict.
194 	 */
195 	if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy))
196 		return -EBUSY;
197 
198 	ret = ggtt_flush(vm->gt);
199 	if (ret)
200 		return ret;
201 
202 	cond_resched();
203 
204 	flags |= PIN_NONBLOCK;
205 	goto search_again;
206 
207 found:
208 	/* drm_mm doesn't allow any other other operations while
209 	 * scanning, therefore store to-be-evicted objects on a
210 	 * temporary list and take a reference for all before
211 	 * calling unbind (which may remove the active reference
212 	 * of any of our objects, thus corrupting the list).
213 	 */
214 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
215 		if (drm_mm_scan_remove_block(&scan, &vma->node))
216 			__i915_vma_pin(vma);
217 		else
218 			list_del(&vma->evict_link);
219 	}
220 
221 	/* Unbinding will emit any required flushes */
222 	ret = 0;
223 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
224 		__i915_vma_unpin(vma);
225 		if (ret == 0)
226 			ret = __i915_vma_unbind(vma);
227 	}
228 
229 	while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) {
230 		vma = container_of(node, struct i915_vma, node);
231 		ret = __i915_vma_unbind(vma);
232 	}
233 
234 	return ret;
235 }
236 
237 /**
238  * i915_gem_evict_for_vma - Evict vmas to make room for binding a new one
239  * @vm: address space to evict from
240  * @target: range (and color) to evict for
241  * @flags: additional flags to control the eviction algorithm
242  *
243  * This function will try to evict vmas that overlap the target node.
244  *
245  * To clarify: This is for freeing up virtual address space, not for freeing
246  * memory in e.g. the shrinker.
247  */
248 int i915_gem_evict_for_node(struct i915_address_space *vm,
249 			    struct drm_mm_node *target,
250 			    unsigned int flags)
251 {
252 	LIST_HEAD(eviction_list);
253 	struct drm_mm_node *node;
254 	u64 start = target->start;
255 	u64 end = start + target->size;
256 	struct i915_vma *vma, *next;
257 	int ret = 0;
258 
259 	lockdep_assert_held(&vm->mutex);
260 	GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
261 	GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
262 
263 	trace_i915_gem_evict_node(vm, target, flags);
264 
265 	/*
266 	 * Retire before we search the active list. Although we have
267 	 * reasonable accuracy in our retirement lists, we may have
268 	 * a stray pin (preventing eviction) that can only be resolved by
269 	 * retiring.
270 	 */
271 	intel_gt_retire_requests(vm->gt);
272 
273 	if (i915_vm_has_cache_coloring(vm)) {
274 		/* Expand search to cover neighbouring guard pages (or lack!) */
275 		if (start)
276 			start -= I915_GTT_PAGE_SIZE;
277 
278 		/* Always look at the page afterwards to avoid the end-of-GTT */
279 		end += I915_GTT_PAGE_SIZE;
280 	}
281 	GEM_BUG_ON(start >= end);
282 
283 	drm_mm_for_each_node_in_range(node, &vm->mm, start, end) {
284 		/* If we find any non-objects (!vma), we cannot evict them */
285 		if (node->color == I915_COLOR_UNEVICTABLE) {
286 			ret = -ENOSPC;
287 			break;
288 		}
289 
290 		GEM_BUG_ON(!drm_mm_node_allocated(node));
291 		vma = container_of(node, typeof(*vma), node);
292 
293 		/*
294 		 * If we are using coloring to insert guard pages between
295 		 * different cache domains within the address space, we have
296 		 * to check whether the objects on either side of our range
297 		 * abutt and conflict. If they are in conflict, then we evict
298 		 * those as well to make room for our guard pages.
299 		 */
300 		if (i915_vm_has_cache_coloring(vm)) {
301 			if (node->start + node->size == target->start) {
302 				if (node->color == target->color)
303 					continue;
304 			}
305 			if (node->start == target->start + target->size) {
306 				if (node->color == target->color)
307 					continue;
308 			}
309 		}
310 
311 		if (i915_vma_is_pinned(vma)) {
312 			ret = -ENOSPC;
313 			break;
314 		}
315 
316 		if (flags & PIN_NONBLOCK && i915_vma_is_active(vma)) {
317 			ret = -ENOSPC;
318 			break;
319 		}
320 
321 		/*
322 		 * Never show fear in the face of dragons!
323 		 *
324 		 * We cannot directly remove this node from within this
325 		 * iterator and as with i915_gem_evict_something() we employ
326 		 * the vma pin_count in order to prevent the action of
327 		 * unbinding one vma from freeing (by dropping its active
328 		 * reference) another in our eviction list.
329 		 */
330 		__i915_vma_pin(vma);
331 		list_add(&vma->evict_link, &eviction_list);
332 	}
333 
334 	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
335 		__i915_vma_unpin(vma);
336 		if (ret == 0)
337 			ret = __i915_vma_unbind(vma);
338 	}
339 
340 	return ret;
341 }
342 
343 /**
344  * i915_gem_evict_vm - Evict all idle vmas from a vm
345  * @vm: Address space to cleanse
346  *
347  * This function evicts all vmas from a vm.
348  *
349  * This is used by the execbuf code as a last-ditch effort to defragment the
350  * address space.
351  *
352  * To clarify: This is for freeing up virtual address space, not for freeing
353  * memory in e.g. the shrinker.
354  */
355 int i915_gem_evict_vm(struct i915_address_space *vm)
356 {
357 	int ret = 0;
358 
359 	lockdep_assert_held(&vm->mutex);
360 	trace_i915_gem_evict_vm(vm);
361 
362 	/* Switch back to the default context in order to unpin
363 	 * the existing context objects. However, such objects only
364 	 * pin themselves inside the global GTT and performing the
365 	 * switch otherwise is ineffective.
366 	 */
367 	if (i915_is_ggtt(vm)) {
368 		ret = ggtt_flush(vm->gt);
369 		if (ret)
370 			return ret;
371 	}
372 
373 	do {
374 		struct i915_vma *vma, *vn;
375 		LIST_HEAD(eviction_list);
376 
377 		list_for_each_entry(vma, &vm->bound_list, vm_link) {
378 			if (i915_vma_is_pinned(vma))
379 				continue;
380 
381 			__i915_vma_pin(vma);
382 			list_add(&vma->evict_link, &eviction_list);
383 		}
384 		if (list_empty(&eviction_list))
385 			break;
386 
387 		ret = 0;
388 		list_for_each_entry_safe(vma, vn, &eviction_list, evict_link) {
389 			__i915_vma_unpin(vma);
390 			if (ret == 0)
391 				ret = __i915_vma_unbind(vma);
392 			if (ret != -EINTR) /* "Get me out of here!" */
393 				ret = 0;
394 		}
395 	} while (ret == 0);
396 
397 	return ret;
398 }
399 
400 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
401 #include "selftests/i915_gem_evict.c"
402 #endif
403