1 /*
2 * Copyright © 2017 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 */
24
25 #include <linux/slab.h>
26
27 #include "i915_syncmap.h"
28
29 #include "i915_gem.h" /* GEM_BUG_ON() */
30 #include "i915_selftest.h"
31
32 #define SHIFT ilog2(KSYNCMAP)
33 #define MASK (KSYNCMAP - 1)
34
35 /*
36 * struct i915_syncmap is a layer of a radixtree that maps a u64 fence
37 * context id to the last u32 fence seqno waited upon from that context.
38 * Unlike lib/radixtree it uses a parent pointer that allows traversal back to
39 * the root. This allows us to access the whole tree via a single pointer
40 * to the most recently used layer. We expect fence contexts to be dense
41 * and most reuse to be on the same i915_gem_context but on neighbouring
42 * engines (i.e. on adjacent contexts) and reuse the same leaf, a very
43 * effective lookup cache. If the new lookup is not on the same leaf, we
44 * expect it to be on the neighbouring branch.
45 *
46 * A leaf holds an array of u32 seqno, and has height 0. The bitmap field
47 * allows us to store whether a particular seqno is valid (i.e. allows us
48 * to distinguish unset from 0).
49 *
50 * A branch holds an array of layer pointers, and has height > 0, and always
51 * has at least 2 layers (either branches or leaves) below it.
52 *
53 * For example,
54 * for x in
55 * 0 1 2 0x10 0x11 0x200 0x201
56 * 0x500000 0x500001 0x503000 0x503001
57 * 0xE<<60:
58 * i915_syncmap_set(&sync, x, lower_32_bits(x));
59 * will build a tree like:
60 * 0xXXXXXXXXXXXXXXXX
61 * 0-> 0x0000000000XXXXXX
62 * | 0-> 0x0000000000000XXX
63 * | | 0-> 0x00000000000000XX
64 * | | | 0-> 0x000000000000000X 0:0, 1:1, 2:2
65 * | | | 1-> 0x000000000000001X 0:10, 1:11
66 * | | 2-> 0x000000000000020X 0:200, 1:201
67 * | 5-> 0x000000000050XXXX
68 * | 0-> 0x000000000050000X 0:500000, 1:500001
69 * | 3-> 0x000000000050300X 0:503000, 1:503001
70 * e-> 0xe00000000000000X e:e
71 */
72
73 struct i915_syncmap {
74 u64 prefix;
75 unsigned int height;
76 unsigned int bitmap;
77 struct i915_syncmap *parent;
78 /*
79 * Following this header is an array of either seqno or child pointers:
80 * union {
81 * u32 seqno[KSYNCMAP];
82 * struct i915_syncmap *child[KSYNCMAP];
83 * };
84 */
85 };
86
87 /**
88 * i915_syncmap_init -- initialise the #i915_syncmap
89 * @root: pointer to the #i915_syncmap
90 */
i915_syncmap_init(struct i915_syncmap ** root)91 void i915_syncmap_init(struct i915_syncmap **root)
92 {
93 BUILD_BUG_ON_NOT_POWER_OF_2(KSYNCMAP);
94 BUILD_BUG_ON_NOT_POWER_OF_2(SHIFT);
95 BUILD_BUG_ON(KSYNCMAP > BITS_PER_TYPE((*root)->bitmap));
96 *root = NULL;
97 }
98
__sync_seqno(struct i915_syncmap * p)99 static inline u32 *__sync_seqno(struct i915_syncmap *p)
100 {
101 GEM_BUG_ON(p->height);
102 return (u32 *)(p + 1);
103 }
104
__sync_child(struct i915_syncmap * p)105 static inline struct i915_syncmap **__sync_child(struct i915_syncmap *p)
106 {
107 GEM_BUG_ON(!p->height);
108 return (struct i915_syncmap **)(p + 1);
109 }
110
111 static inline unsigned int
__sync_branch_idx(const struct i915_syncmap * p,u64 id)112 __sync_branch_idx(const struct i915_syncmap *p, u64 id)
113 {
114 return (id >> p->height) & MASK;
115 }
116
117 static inline unsigned int
__sync_leaf_idx(const struct i915_syncmap * p,u64 id)118 __sync_leaf_idx(const struct i915_syncmap *p, u64 id)
119 {
120 GEM_BUG_ON(p->height);
121 return id & MASK;
122 }
123
__sync_branch_prefix(const struct i915_syncmap * p,u64 id)124 static inline u64 __sync_branch_prefix(const struct i915_syncmap *p, u64 id)
125 {
126 return id >> p->height >> SHIFT;
127 }
128
__sync_leaf_prefix(const struct i915_syncmap * p,u64 id)129 static inline u64 __sync_leaf_prefix(const struct i915_syncmap *p, u64 id)
130 {
131 GEM_BUG_ON(p->height);
132 return id >> SHIFT;
133 }
134
seqno_later(u32 a,u32 b)135 static inline bool seqno_later(u32 a, u32 b)
136 {
137 return (s32)(a - b) >= 0;
138 }
139
140 /**
141 * i915_syncmap_is_later -- compare against the last know sync point
142 * @root: pointer to the #i915_syncmap
143 * @id: the context id (other timeline) we are synchronising to
144 * @seqno: the sequence number along the other timeline
145 *
146 * If we have already synchronised this @root timeline with another (@id) then
147 * we can omit any repeated or earlier synchronisation requests. If the two
148 * timelines are already coupled, we can also omit the dependency between the
149 * two as that is already known via the timeline.
150 *
151 * Returns true if the two timelines are already synchronised wrt to @seqno,
152 * false if not and the synchronisation must be emitted.
153 */
i915_syncmap_is_later(struct i915_syncmap ** root,u64 id,u32 seqno)154 bool i915_syncmap_is_later(struct i915_syncmap **root, u64 id, u32 seqno)
155 {
156 struct i915_syncmap *p;
157 unsigned int idx;
158
159 p = *root;
160 if (!p)
161 return false;
162
163 if (likely(__sync_leaf_prefix(p, id) == p->prefix))
164 goto found;
165
166 /* First climb the tree back to a parent branch */
167 do {
168 p = p->parent;
169 if (!p)
170 return false;
171
172 if (__sync_branch_prefix(p, id) == p->prefix)
173 break;
174 } while (1);
175
176 /* And then descend again until we find our leaf */
177 do {
178 if (!p->height)
179 break;
180
181 p = __sync_child(p)[__sync_branch_idx(p, id)];
182 if (!p)
183 return false;
184
185 if (__sync_branch_prefix(p, id) != p->prefix)
186 return false;
187 } while (1);
188
189 *root = p;
190 found:
191 idx = __sync_leaf_idx(p, id);
192 if (!(p->bitmap & BIT(idx)))
193 return false;
194
195 return seqno_later(__sync_seqno(p)[idx], seqno);
196 }
197
198 static struct i915_syncmap *
__sync_alloc_leaf(struct i915_syncmap * parent,u64 id)199 __sync_alloc_leaf(struct i915_syncmap *parent, u64 id)
200 {
201 struct i915_syncmap *p;
202
203 p = kmalloc(sizeof(*p) + KSYNCMAP * sizeof(u32), GFP_KERNEL);
204 if (unlikely(!p))
205 return NULL;
206
207 p->parent = parent;
208 p->height = 0;
209 p->bitmap = 0;
210 p->prefix = __sync_leaf_prefix(p, id);
211 return p;
212 }
213
__sync_set_seqno(struct i915_syncmap * p,u64 id,u32 seqno)214 static inline void __sync_set_seqno(struct i915_syncmap *p, u64 id, u32 seqno)
215 {
216 unsigned int idx = __sync_leaf_idx(p, id);
217
218 p->bitmap |= BIT(idx);
219 __sync_seqno(p)[idx] = seqno;
220 }
221
__sync_set_child(struct i915_syncmap * p,unsigned int idx,struct i915_syncmap * child)222 static inline void __sync_set_child(struct i915_syncmap *p,
223 unsigned int idx,
224 struct i915_syncmap *child)
225 {
226 p->bitmap |= BIT(idx);
227 __sync_child(p)[idx] = child;
228 }
229
__sync_set(struct i915_syncmap ** root,u64 id,u32 seqno)230 static noinline int __sync_set(struct i915_syncmap **root, u64 id, u32 seqno)
231 {
232 struct i915_syncmap *p = *root;
233 unsigned int idx;
234
235 if (!p) {
236 p = __sync_alloc_leaf(NULL, id);
237 if (unlikely(!p))
238 return -ENOMEM;
239
240 goto found;
241 }
242
243 /* Caller handled the likely cached case */
244 GEM_BUG_ON(__sync_leaf_prefix(p, id) == p->prefix);
245
246 /* Climb back up the tree until we find a common prefix */
247 do {
248 if (!p->parent)
249 break;
250
251 p = p->parent;
252
253 if (__sync_branch_prefix(p, id) == p->prefix)
254 break;
255 } while (1);
256
257 /*
258 * No shortcut, we have to descend the tree to find the right layer
259 * containing this fence.
260 *
261 * Each layer in the tree holds 16 (KSYNCMAP) pointers, either fences
262 * or lower layers. Leaf nodes (height = 0) contain the fences, all
263 * other nodes (height > 0) are internal layers that point to a lower
264 * node. Each internal layer has at least 2 descendents.
265 *
266 * Starting at the top, we check whether the current prefix matches. If
267 * it doesn't, we have gone past our target and need to insert a join
268 * into the tree, and a new leaf node for the target as a descendent
269 * of the join, as well as the original layer.
270 *
271 * The matching prefix means we are still following the right branch
272 * of the tree. If it has height 0, we have found our leaf and just
273 * need to replace the fence slot with ourselves. If the height is
274 * not zero, our slot contains the next layer in the tree (unless
275 * it is empty, in which case we can add ourselves as a new leaf).
276 * As descend the tree the prefix grows (and height decreases).
277 */
278 do {
279 struct i915_syncmap *next;
280
281 if (__sync_branch_prefix(p, id) != p->prefix) {
282 unsigned int above;
283
284 /* Insert a join above the current layer */
285 next = kzalloc(sizeof(*next) + KSYNCMAP * sizeof(next),
286 GFP_KERNEL);
287 if (unlikely(!next))
288 return -ENOMEM;
289
290 /* Compute the height at which these two diverge */
291 above = fls64(__sync_branch_prefix(p, id) ^ p->prefix);
292 above = round_up(above, SHIFT);
293 next->height = above + p->height;
294 next->prefix = __sync_branch_prefix(next, id);
295
296 /* Insert the join into the parent */
297 if (p->parent) {
298 idx = __sync_branch_idx(p->parent, id);
299 __sync_child(p->parent)[idx] = next;
300 GEM_BUG_ON(!(p->parent->bitmap & BIT(idx)));
301 }
302 next->parent = p->parent;
303
304 /* Compute the idx of the other branch, not our id! */
305 idx = p->prefix >> (above - SHIFT) & MASK;
306 __sync_set_child(next, idx, p);
307 p->parent = next;
308
309 /* Ascend to the join */
310 p = next;
311 } else {
312 if (!p->height)
313 break;
314 }
315
316 /* Descend into the next layer */
317 GEM_BUG_ON(!p->height);
318 idx = __sync_branch_idx(p, id);
319 next = __sync_child(p)[idx];
320 if (!next) {
321 next = __sync_alloc_leaf(p, id);
322 if (unlikely(!next))
323 return -ENOMEM;
324
325 __sync_set_child(p, idx, next);
326 p = next;
327 break;
328 }
329
330 p = next;
331 } while (1);
332
333 found:
334 GEM_BUG_ON(p->prefix != __sync_leaf_prefix(p, id));
335 __sync_set_seqno(p, id, seqno);
336 *root = p;
337 return 0;
338 }
339
340 /**
341 * i915_syncmap_set -- mark the most recent syncpoint between contexts
342 * @root: pointer to the #i915_syncmap
343 * @id: the context id (other timeline) we have synchronised to
344 * @seqno: the sequence number along the other timeline
345 *
346 * When we synchronise this @root timeline with another (@id), we also know
347 * that we have synchronized with all previous seqno along that timeline. If
348 * we then have a request to synchronise with the same seqno or older, we can
349 * omit it, see i915_syncmap_is_later()
350 *
351 * Returns 0 on success, or a negative error code.
352 */
i915_syncmap_set(struct i915_syncmap ** root,u64 id,u32 seqno)353 int i915_syncmap_set(struct i915_syncmap **root, u64 id, u32 seqno)
354 {
355 struct i915_syncmap *p = *root;
356
357 /*
358 * We expect to be called in sequence following is_later(id), which
359 * should have preloaded the root for us.
360 */
361 if (likely(p && __sync_leaf_prefix(p, id) == p->prefix)) {
362 __sync_set_seqno(p, id, seqno);
363 return 0;
364 }
365
366 return __sync_set(root, id, seqno);
367 }
368
__sync_free(struct i915_syncmap * p)369 static void __sync_free(struct i915_syncmap *p)
370 {
371 if (p->height) {
372 unsigned int i;
373
374 while ((i = ffs(p->bitmap))) {
375 p->bitmap &= ~0u << i;
376 __sync_free(__sync_child(p)[i - 1]);
377 }
378 }
379
380 kfree(p);
381 }
382
383 /**
384 * i915_syncmap_free -- free all memory associated with the syncmap
385 * @root: pointer to the #i915_syncmap
386 *
387 * Either when the timeline is to be freed and we no longer need the sync
388 * point tracking, or when the fences are all known to be signaled and the
389 * sync point tracking is redundant, we can free the #i915_syncmap to recover
390 * its allocations.
391 *
392 * Will reinitialise the @root pointer so that the #i915_syncmap is ready for
393 * reuse.
394 */
i915_syncmap_free(struct i915_syncmap ** root)395 void i915_syncmap_free(struct i915_syncmap **root)
396 {
397 struct i915_syncmap *p;
398
399 p = *root;
400 if (!p)
401 return;
402
403 while (p->parent)
404 p = p->parent;
405
406 __sync_free(p);
407 *root = NULL;
408 }
409
410 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
411 #include "selftests/i915_syncmap.c"
412 #endif
413