xref: /linux/drivers/gpu/drm/i915/gt/intel_gt_buffer_pool.c (revision c098564d91c55d408ed31e8885b915a5e2006249)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2014-2018 Intel Corporation
4  */
5 
6 #include "gem/i915_gem_object.h"
7 
8 #include "i915_drv.h"
9 #include "intel_engine_pm.h"
10 #include "intel_gt_buffer_pool.h"
11 
12 static struct intel_gt *to_gt(struct intel_gt_buffer_pool *pool)
13 {
14 	return container_of(pool, struct intel_gt, buffer_pool);
15 }
16 
17 static struct list_head *
18 bucket_for_size(struct intel_gt_buffer_pool *pool, size_t sz)
19 {
20 	int n;
21 
22 	/*
23 	 * Compute a power-of-two bucket, but throw everything greater than
24 	 * 16KiB into the same bucket: i.e. the buckets hold objects of
25 	 * (1 page, 2 pages, 4 pages, 8+ pages).
26 	 */
27 	n = fls(sz >> PAGE_SHIFT) - 1;
28 	if (n >= ARRAY_SIZE(pool->cache_list))
29 		n = ARRAY_SIZE(pool->cache_list) - 1;
30 
31 	return &pool->cache_list[n];
32 }
33 
34 static void node_free(struct intel_gt_buffer_pool_node *node)
35 {
36 	i915_gem_object_put(node->obj);
37 	i915_active_fini(&node->active);
38 	kfree_rcu(node, rcu);
39 }
40 
41 static bool pool_free_older_than(struct intel_gt_buffer_pool *pool, long keep)
42 {
43 	struct intel_gt_buffer_pool_node *node, *stale = NULL;
44 	bool active = false;
45 	int n;
46 
47 	/* Free buffers that have not been used in the past second */
48 	for (n = 0; n < ARRAY_SIZE(pool->cache_list); n++) {
49 		struct list_head *list = &pool->cache_list[n];
50 
51 		if (list_empty(list))
52 			continue;
53 
54 		if (spin_trylock_irq(&pool->lock)) {
55 			struct list_head *pos;
56 
57 			/* Most recent at head; oldest at tail */
58 			list_for_each_prev(pos, list) {
59 				unsigned long age;
60 
61 				node = list_entry(pos, typeof(*node), link);
62 
63 				age = READ_ONCE(node->age);
64 				if (!age || jiffies - age < keep)
65 					break;
66 
67 				/* Check we are the first to claim this node */
68 				if (!xchg(&node->age, 0))
69 					break;
70 
71 				node->free = stale;
72 				stale = node;
73 			}
74 			if (!list_is_last(pos, list))
75 				__list_del_many(pos, list);
76 
77 			spin_unlock_irq(&pool->lock);
78 		}
79 
80 		active |= !list_empty(list);
81 	}
82 
83 	while ((node = stale)) {
84 		stale = stale->free;
85 		node_free(node);
86 	}
87 
88 	return active;
89 }
90 
91 static void pool_free_work(struct work_struct *wrk)
92 {
93 	struct intel_gt_buffer_pool *pool =
94 		container_of(wrk, typeof(*pool), work.work);
95 
96 	if (pool_free_older_than(pool, HZ))
97 		schedule_delayed_work(&pool->work,
98 				      round_jiffies_up_relative(HZ));
99 }
100 
101 __i915_active_call
102 static void pool_retire(struct i915_active *ref)
103 {
104 	struct intel_gt_buffer_pool_node *node =
105 		container_of(ref, typeof(*node), active);
106 	struct intel_gt_buffer_pool *pool = node->pool;
107 	struct list_head *list = bucket_for_size(pool, node->obj->base.size);
108 	unsigned long flags;
109 
110 	if (node->pinned) {
111 		i915_gem_object_unpin_pages(node->obj);
112 
113 		/* Return this object to the shrinker pool */
114 		i915_gem_object_make_purgeable(node->obj);
115 		node->pinned = false;
116 	}
117 
118 	GEM_BUG_ON(node->age);
119 	spin_lock_irqsave(&pool->lock, flags);
120 	list_add_rcu(&node->link, list);
121 	WRITE_ONCE(node->age, jiffies ?: 1); /* 0 reserved for active nodes */
122 	spin_unlock_irqrestore(&pool->lock, flags);
123 
124 	schedule_delayed_work(&pool->work,
125 			      round_jiffies_up_relative(HZ));
126 }
127 
128 void intel_gt_buffer_pool_mark_used(struct intel_gt_buffer_pool_node *node)
129 {
130 	assert_object_held(node->obj);
131 
132 	if (node->pinned)
133 		return;
134 
135 	__i915_gem_object_pin_pages(node->obj);
136 	/* Hide this pinned object from the shrinker until retired */
137 	i915_gem_object_make_unshrinkable(node->obj);
138 	node->pinned = true;
139 }
140 
141 static struct intel_gt_buffer_pool_node *
142 node_create(struct intel_gt_buffer_pool *pool, size_t sz,
143 	    enum i915_map_type type)
144 {
145 	struct intel_gt *gt = to_gt(pool);
146 	struct intel_gt_buffer_pool_node *node;
147 	struct drm_i915_gem_object *obj;
148 
149 	node = kmalloc(sizeof(*node),
150 		       GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
151 	if (!node)
152 		return ERR_PTR(-ENOMEM);
153 
154 	node->age = 0;
155 	node->pool = pool;
156 	node->pinned = false;
157 	i915_active_init(&node->active, NULL, pool_retire);
158 
159 	obj = i915_gem_object_create_internal(gt->i915, sz);
160 	if (IS_ERR(obj)) {
161 		i915_active_fini(&node->active);
162 		kfree(node);
163 		return ERR_CAST(obj);
164 	}
165 
166 	i915_gem_object_set_readonly(obj);
167 
168 	node->type = type;
169 	node->obj = obj;
170 	return node;
171 }
172 
173 struct intel_gt_buffer_pool_node *
174 intel_gt_get_buffer_pool(struct intel_gt *gt, size_t size,
175 			 enum i915_map_type type)
176 {
177 	struct intel_gt_buffer_pool *pool = &gt->buffer_pool;
178 	struct intel_gt_buffer_pool_node *node;
179 	struct list_head *list;
180 	int ret;
181 
182 	size = PAGE_ALIGN(size);
183 	list = bucket_for_size(pool, size);
184 
185 	rcu_read_lock();
186 	list_for_each_entry_rcu(node, list, link) {
187 		unsigned long age;
188 
189 		if (node->obj->base.size < size)
190 			continue;
191 
192 		if (node->type != type)
193 			continue;
194 
195 		age = READ_ONCE(node->age);
196 		if (!age)
197 			continue;
198 
199 		if (cmpxchg(&node->age, age, 0) == age) {
200 			spin_lock_irq(&pool->lock);
201 			list_del_rcu(&node->link);
202 			spin_unlock_irq(&pool->lock);
203 			break;
204 		}
205 	}
206 	rcu_read_unlock();
207 
208 	if (&node->link == list) {
209 		node = node_create(pool, size, type);
210 		if (IS_ERR(node))
211 			return node;
212 	}
213 
214 	ret = i915_active_acquire(&node->active);
215 	if (ret) {
216 		node_free(node);
217 		return ERR_PTR(ret);
218 	}
219 
220 	return node;
221 }
222 
223 void intel_gt_init_buffer_pool(struct intel_gt *gt)
224 {
225 	struct intel_gt_buffer_pool *pool = &gt->buffer_pool;
226 	int n;
227 
228 	spin_lock_init(&pool->lock);
229 	for (n = 0; n < ARRAY_SIZE(pool->cache_list); n++)
230 		INIT_LIST_HEAD(&pool->cache_list[n]);
231 	INIT_DELAYED_WORK(&pool->work, pool_free_work);
232 }
233 
234 void intel_gt_flush_buffer_pool(struct intel_gt *gt)
235 {
236 	struct intel_gt_buffer_pool *pool = &gt->buffer_pool;
237 
238 	do {
239 		while (pool_free_older_than(pool, 0))
240 			;
241 	} while (cancel_delayed_work_sync(&pool->work));
242 }
243 
244 void intel_gt_fini_buffer_pool(struct intel_gt *gt)
245 {
246 	struct intel_gt_buffer_pool *pool = &gt->buffer_pool;
247 	int n;
248 
249 	intel_gt_flush_buffer_pool(gt);
250 
251 	for (n = 0; n < ARRAY_SIZE(pool->cache_list); n++)
252 		GEM_BUG_ON(!list_empty(&pool->cache_list[n]));
253 }
254