| 1 | // SPDX-License-Identifier: MIT |
| 2 | /* |
| 3 | * Copyright © 2019 Intel Corporation |
| 4 | * Copyright © 2022 Maíra Canal <mairacanal@riseup.net> |
| 5 | */ |
| 6 | |
| 7 | #include <kunit/test.h> |
| 8 | |
| 9 | #include <linux/prime_numbers.h> |
| 10 | #include <linux/sched/signal.h> |
| 11 | #include <linux/sizes.h> |
| 12 | |
| 13 | #include <drm/drm_buddy.h> |
| 14 | |
| 15 | #include "../lib/drm_random.h" |
| 16 | |
| 17 | static unsigned int random_seed; |
| 18 | |
| 19 | static inline u64 get_size(int order, u64 chunk_size) |
| 20 | { |
| 21 | return (1 << order) * chunk_size; |
| 22 | } |
| 23 | |
| 24 | static void drm_test_buddy_fragmentation_performance(struct kunit *test) |
| 25 | { |
| 26 | struct drm_buddy_block *block, *tmp; |
| 27 | int num_blocks, i, ret, count = 0; |
| 28 | LIST_HEAD(allocated_blocks); |
| 29 | unsigned long elapsed_ms; |
| 30 | LIST_HEAD(reverse_list); |
| 31 | LIST_HEAD(test_blocks); |
| 32 | LIST_HEAD(clear_list); |
| 33 | LIST_HEAD(dirty_list); |
| 34 | LIST_HEAD(free_list); |
| 35 | struct drm_buddy mm; |
| 36 | u64 mm_size = SZ_4G; |
| 37 | ktime_t start, end; |
| 38 | |
| 39 | /* |
| 40 | * Allocation under severe fragmentation |
| 41 | * |
| 42 | * Create severe fragmentation by allocating the entire 4 GiB address space |
| 43 | * as tiny 8 KiB blocks but forcing a 64 KiB alignment. The resulting pattern |
| 44 | * leaves many scattered holes. Split the allocations into two groups and |
| 45 | * return them with different flags to block coalescing, then repeatedly |
| 46 | * allocate and free 64 KiB blocks while timing the loop. This stresses how |
| 47 | * quickly the allocator can satisfy larger, aligned requests from a pool of |
| 48 | * highly fragmented space. |
| 49 | */ |
| 50 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_init(&mm, mm_size, SZ_4K), |
| 51 | "buddy_init failed\n" ); |
| 52 | |
| 53 | num_blocks = mm_size / SZ_64K; |
| 54 | |
| 55 | start = ktime_get(); |
| 56 | /* Allocate with maximum fragmentation - 8K blocks with 64K alignment */ |
| 57 | for (i = 0; i < num_blocks; i++) |
| 58 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, SZ_8K, SZ_64K, |
| 59 | &allocated_blocks, 0), |
| 60 | "buddy_alloc hit an error size=%u\n" , SZ_8K); |
| 61 | |
| 62 | list_for_each_entry_safe(block, tmp, &allocated_blocks, link) { |
| 63 | if (count % 4 == 0 || count % 4 == 3) |
| 64 | list_move_tail(list: &block->link, head: &clear_list); |
| 65 | else |
| 66 | list_move_tail(list: &block->link, head: &dirty_list); |
| 67 | count++; |
| 68 | } |
| 69 | |
| 70 | /* Free with different flags to ensure no coalescing */ |
| 71 | drm_buddy_free_list(mm: &mm, objects: &clear_list, DRM_BUDDY_CLEARED); |
| 72 | drm_buddy_free_list(mm: &mm, objects: &dirty_list, flags: 0); |
| 73 | |
| 74 | for (i = 0; i < num_blocks; i++) |
| 75 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, SZ_64K, SZ_64K, |
| 76 | &test_blocks, 0), |
| 77 | "buddy_alloc hit an error size=%u\n" , SZ_64K); |
| 78 | drm_buddy_free_list(mm: &mm, objects: &test_blocks, flags: 0); |
| 79 | |
| 80 | end = ktime_get(); |
| 81 | elapsed_ms = ktime_to_ms(ktime_sub(end, start)); |
| 82 | |
| 83 | kunit_info(test, "Fragmented allocation took %lu ms\n" , elapsed_ms); |
| 84 | |
| 85 | drm_buddy_fini(mm: &mm); |
| 86 | |
| 87 | /* |
| 88 | * Reverse free order under fragmentation |
| 89 | * |
| 90 | * Construct a fragmented 4 GiB space by allocating every 8 KiB block with |
| 91 | * 64 KiB alignment, creating a dense scatter of small regions. Half of the |
| 92 | * blocks are selectively freed to form sparse gaps, while the remaining |
| 93 | * allocations are preserved, reordered in reverse, and released back with |
| 94 | * the cleared flag. This models a pathological reverse-ordered free pattern |
| 95 | * and measures how quickly the allocator can merge and reclaim space when |
| 96 | * deallocation occurs in the opposite order of allocation, exposing the |
| 97 | * cost difference between a linear freelist scan and an ordered tree lookup. |
| 98 | */ |
| 99 | ret = drm_buddy_init(mm: &mm, size: mm_size, SZ_4K); |
| 100 | KUNIT_ASSERT_EQ(test, ret, 0); |
| 101 | |
| 102 | start = ktime_get(); |
| 103 | /* Allocate maximum fragmentation */ |
| 104 | for (i = 0; i < num_blocks; i++) |
| 105 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, SZ_8K, SZ_64K, |
| 106 | &allocated_blocks, 0), |
| 107 | "buddy_alloc hit an error size=%u\n" , SZ_8K); |
| 108 | |
| 109 | list_for_each_entry_safe(block, tmp, &allocated_blocks, link) { |
| 110 | if (count % 2 == 0) |
| 111 | list_move_tail(list: &block->link, head: &free_list); |
| 112 | count++; |
| 113 | } |
| 114 | drm_buddy_free_list(mm: &mm, objects: &free_list, DRM_BUDDY_CLEARED); |
| 115 | |
| 116 | list_for_each_entry_safe_reverse(block, tmp, &allocated_blocks, link) |
| 117 | list_move(list: &block->link, head: &reverse_list); |
| 118 | drm_buddy_free_list(mm: &mm, objects: &reverse_list, DRM_BUDDY_CLEARED); |
| 119 | |
| 120 | end = ktime_get(); |
| 121 | elapsed_ms = ktime_to_ms(ktime_sub(end, start)); |
| 122 | |
| 123 | kunit_info(test, "Reverse-ordered free took %lu ms\n" , elapsed_ms); |
| 124 | |
| 125 | drm_buddy_fini(mm: &mm); |
| 126 | } |
| 127 | |
| 128 | static void drm_test_buddy_alloc_range_bias(struct kunit *test) |
| 129 | { |
| 130 | u32 mm_size, size, ps, bias_size, bias_start, bias_end, bias_rem; |
| 131 | DRM_RND_STATE(prng, random_seed); |
| 132 | unsigned int i, count, *order; |
| 133 | struct drm_buddy_block *block; |
| 134 | unsigned long flags; |
| 135 | struct drm_buddy mm; |
| 136 | LIST_HEAD(allocated); |
| 137 | |
| 138 | bias_size = SZ_1M; |
| 139 | ps = roundup_pow_of_two(prandom_u32_state(&prng) % bias_size); |
| 140 | ps = max(SZ_4K, ps); |
| 141 | mm_size = (SZ_8M-1) & ~(ps-1); /* Multiple roots */ |
| 142 | |
| 143 | kunit_info(test, "mm_size=%u, ps=%u\n" , mm_size, ps); |
| 144 | |
| 145 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_init(&mm, mm_size, ps), |
| 146 | "buddy_init failed\n" ); |
| 147 | |
| 148 | count = mm_size / bias_size; |
| 149 | order = drm_random_order(count, state: &prng); |
| 150 | KUNIT_EXPECT_TRUE(test, order); |
| 151 | |
| 152 | /* |
| 153 | * Idea is to split the address space into uniform bias ranges, and then |
| 154 | * in some random order allocate within each bias, using various |
| 155 | * patterns within. This should detect if allocations leak out from a |
| 156 | * given bias, for example. |
| 157 | */ |
| 158 | |
| 159 | for (i = 0; i < count; i++) { |
| 160 | LIST_HEAD(tmp); |
| 161 | u32 size; |
| 162 | |
| 163 | bias_start = order[i] * bias_size; |
| 164 | bias_end = bias_start + bias_size; |
| 165 | bias_rem = bias_size; |
| 166 | |
| 167 | /* internal round_up too big */ |
| 168 | KUNIT_ASSERT_TRUE_MSG(test, |
| 169 | drm_buddy_alloc_blocks(&mm, bias_start, |
| 170 | bias_end, bias_size + ps, bias_size, |
| 171 | &allocated, |
| 172 | DRM_BUDDY_RANGE_ALLOCATION), |
| 173 | "buddy_alloc failed with bias(%x-%x), size=%u, ps=%u\n" , |
| 174 | bias_start, bias_end, bias_size, bias_size); |
| 175 | |
| 176 | /* size too big */ |
| 177 | KUNIT_ASSERT_TRUE_MSG(test, |
| 178 | drm_buddy_alloc_blocks(&mm, bias_start, |
| 179 | bias_end, bias_size + ps, ps, |
| 180 | &allocated, |
| 181 | DRM_BUDDY_RANGE_ALLOCATION), |
| 182 | "buddy_alloc didn't fail with bias(%x-%x), size=%u, ps=%u\n" , |
| 183 | bias_start, bias_end, bias_size + ps, ps); |
| 184 | |
| 185 | /* bias range too small for size */ |
| 186 | KUNIT_ASSERT_TRUE_MSG(test, |
| 187 | drm_buddy_alloc_blocks(&mm, bias_start + ps, |
| 188 | bias_end, bias_size, ps, |
| 189 | &allocated, |
| 190 | DRM_BUDDY_RANGE_ALLOCATION), |
| 191 | "buddy_alloc didn't fail with bias(%x-%x), size=%u, ps=%u\n" , |
| 192 | bias_start + ps, bias_end, bias_size, ps); |
| 193 | |
| 194 | /* bias misaligned */ |
| 195 | KUNIT_ASSERT_TRUE_MSG(test, |
| 196 | drm_buddy_alloc_blocks(&mm, bias_start + ps, |
| 197 | bias_end - ps, |
| 198 | bias_size >> 1, bias_size >> 1, |
| 199 | &allocated, |
| 200 | DRM_BUDDY_RANGE_ALLOCATION), |
| 201 | "buddy_alloc h didn't fail with bias(%x-%x), size=%u, ps=%u\n" , |
| 202 | bias_start + ps, bias_end - ps, bias_size >> 1, bias_size >> 1); |
| 203 | |
| 204 | /* single big page */ |
| 205 | KUNIT_ASSERT_FALSE_MSG(test, |
| 206 | drm_buddy_alloc_blocks(&mm, bias_start, |
| 207 | bias_end, bias_size, bias_size, |
| 208 | &tmp, |
| 209 | DRM_BUDDY_RANGE_ALLOCATION), |
| 210 | "buddy_alloc i failed with bias(%x-%x), size=%u, ps=%u\n" , |
| 211 | bias_start, bias_end, bias_size, bias_size); |
| 212 | drm_buddy_free_list(mm: &mm, objects: &tmp, flags: 0); |
| 213 | |
| 214 | /* single page with internal round_up */ |
| 215 | KUNIT_ASSERT_FALSE_MSG(test, |
| 216 | drm_buddy_alloc_blocks(&mm, bias_start, |
| 217 | bias_end, ps, bias_size, |
| 218 | &tmp, |
| 219 | DRM_BUDDY_RANGE_ALLOCATION), |
| 220 | "buddy_alloc failed with bias(%x-%x), size=%u, ps=%u\n" , |
| 221 | bias_start, bias_end, ps, bias_size); |
| 222 | drm_buddy_free_list(mm: &mm, objects: &tmp, flags: 0); |
| 223 | |
| 224 | /* random size within */ |
| 225 | size = max(round_up(prandom_u32_state(&prng) % bias_rem, ps), ps); |
| 226 | if (size) |
| 227 | KUNIT_ASSERT_FALSE_MSG(test, |
| 228 | drm_buddy_alloc_blocks(&mm, bias_start, |
| 229 | bias_end, size, ps, |
| 230 | &tmp, |
| 231 | DRM_BUDDY_RANGE_ALLOCATION), |
| 232 | "buddy_alloc failed with bias(%x-%x), size=%u, ps=%u\n" , |
| 233 | bias_start, bias_end, size, ps); |
| 234 | |
| 235 | bias_rem -= size; |
| 236 | /* too big for current avail */ |
| 237 | KUNIT_ASSERT_TRUE_MSG(test, |
| 238 | drm_buddy_alloc_blocks(&mm, bias_start, |
| 239 | bias_end, bias_rem + ps, ps, |
| 240 | &allocated, |
| 241 | DRM_BUDDY_RANGE_ALLOCATION), |
| 242 | "buddy_alloc didn't fail with bias(%x-%x), size=%u, ps=%u\n" , |
| 243 | bias_start, bias_end, bias_rem + ps, ps); |
| 244 | |
| 245 | if (bias_rem) { |
| 246 | /* random fill of the remainder */ |
| 247 | size = max(round_up(prandom_u32_state(&prng) % bias_rem, ps), ps); |
| 248 | size = max(size, ps); |
| 249 | |
| 250 | KUNIT_ASSERT_FALSE_MSG(test, |
| 251 | drm_buddy_alloc_blocks(&mm, bias_start, |
| 252 | bias_end, size, ps, |
| 253 | &allocated, |
| 254 | DRM_BUDDY_RANGE_ALLOCATION), |
| 255 | "buddy_alloc failed with bias(%x-%x), size=%u, ps=%u\n" , |
| 256 | bias_start, bias_end, size, ps); |
| 257 | /* |
| 258 | * Intentionally allow some space to be left |
| 259 | * unallocated, and ideally not always on the bias |
| 260 | * boundaries. |
| 261 | */ |
| 262 | drm_buddy_free_list(mm: &mm, objects: &tmp, flags: 0); |
| 263 | } else { |
| 264 | list_splice_tail(list: &tmp, head: &allocated); |
| 265 | } |
| 266 | } |
| 267 | |
| 268 | kfree(objp: order); |
| 269 | drm_buddy_free_list(mm: &mm, objects: &allocated, flags: 0); |
| 270 | drm_buddy_fini(mm: &mm); |
| 271 | |
| 272 | /* |
| 273 | * Something more free-form. Idea is to pick a random starting bias |
| 274 | * range within the address space and then start filling it up. Also |
| 275 | * randomly grow the bias range in both directions as we go along. This |
| 276 | * should give us bias start/end which is not always uniform like above, |
| 277 | * and in some cases will require the allocator to jump over already |
| 278 | * allocated nodes in the middle of the address space. |
| 279 | */ |
| 280 | |
| 281 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_init(&mm, mm_size, ps), |
| 282 | "buddy_init failed\n" ); |
| 283 | |
| 284 | bias_start = round_up(prandom_u32_state(&prng) % (mm_size - ps), ps); |
| 285 | bias_end = round_up(bias_start + prandom_u32_state(&prng) % (mm_size - bias_start), ps); |
| 286 | bias_end = max(bias_end, bias_start + ps); |
| 287 | bias_rem = bias_end - bias_start; |
| 288 | |
| 289 | do { |
| 290 | u32 size = max(round_up(prandom_u32_state(&prng) % bias_rem, ps), ps); |
| 291 | |
| 292 | KUNIT_ASSERT_FALSE_MSG(test, |
| 293 | drm_buddy_alloc_blocks(&mm, bias_start, |
| 294 | bias_end, size, ps, |
| 295 | &allocated, |
| 296 | DRM_BUDDY_RANGE_ALLOCATION), |
| 297 | "buddy_alloc failed with bias(%x-%x), size=%u, ps=%u\n" , |
| 298 | bias_start, bias_end, size, ps); |
| 299 | bias_rem -= size; |
| 300 | |
| 301 | /* |
| 302 | * Try to randomly grow the bias range in both directions, or |
| 303 | * only one, or perhaps don't grow at all. |
| 304 | */ |
| 305 | do { |
| 306 | u32 old_bias_start = bias_start; |
| 307 | u32 old_bias_end = bias_end; |
| 308 | |
| 309 | if (bias_start) |
| 310 | bias_start -= round_up(prandom_u32_state(&prng) % bias_start, ps); |
| 311 | if (bias_end != mm_size) |
| 312 | bias_end += round_up(prandom_u32_state(&prng) % (mm_size - bias_end), ps); |
| 313 | |
| 314 | bias_rem += old_bias_start - bias_start; |
| 315 | bias_rem += bias_end - old_bias_end; |
| 316 | } while (!bias_rem && (bias_start || bias_end != mm_size)); |
| 317 | } while (bias_rem); |
| 318 | |
| 319 | KUNIT_ASSERT_EQ(test, bias_start, 0); |
| 320 | KUNIT_ASSERT_EQ(test, bias_end, mm_size); |
| 321 | KUNIT_ASSERT_TRUE_MSG(test, |
| 322 | drm_buddy_alloc_blocks(&mm, bias_start, bias_end, |
| 323 | ps, ps, |
| 324 | &allocated, |
| 325 | DRM_BUDDY_RANGE_ALLOCATION), |
| 326 | "buddy_alloc passed with bias(%x-%x), size=%u\n" , |
| 327 | bias_start, bias_end, ps); |
| 328 | |
| 329 | drm_buddy_free_list(mm: &mm, objects: &allocated, flags: 0); |
| 330 | drm_buddy_fini(mm: &mm); |
| 331 | |
| 332 | /* |
| 333 | * Allocate cleared blocks in the bias range when the DRM buddy's clear avail is |
| 334 | * zero. This will validate the bias range allocation in scenarios like system boot |
| 335 | * when no cleared blocks are available and exercise the fallback path too. The resulting |
| 336 | * blocks should always be dirty. |
| 337 | */ |
| 338 | |
| 339 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_init(&mm, mm_size, ps), |
| 340 | "buddy_init failed\n" ); |
| 341 | |
| 342 | bias_start = round_up(prandom_u32_state(&prng) % (mm_size - ps), ps); |
| 343 | bias_end = round_up(bias_start + prandom_u32_state(&prng) % (mm_size - bias_start), ps); |
| 344 | bias_end = max(bias_end, bias_start + ps); |
| 345 | bias_rem = bias_end - bias_start; |
| 346 | |
| 347 | flags = DRM_BUDDY_CLEAR_ALLOCATION | DRM_BUDDY_RANGE_ALLOCATION; |
| 348 | size = max(round_up(prandom_u32_state(&prng) % bias_rem, ps), ps); |
| 349 | |
| 350 | KUNIT_ASSERT_FALSE_MSG(test, |
| 351 | drm_buddy_alloc_blocks(&mm, bias_start, |
| 352 | bias_end, size, ps, |
| 353 | &allocated, |
| 354 | flags), |
| 355 | "buddy_alloc failed with bias(%x-%x), size=%u, ps=%u\n" , |
| 356 | bias_start, bias_end, size, ps); |
| 357 | |
| 358 | list_for_each_entry(block, &allocated, link) |
| 359 | KUNIT_EXPECT_EQ(test, drm_buddy_block_is_clear(block), false); |
| 360 | |
| 361 | drm_buddy_free_list(mm: &mm, objects: &allocated, flags: 0); |
| 362 | drm_buddy_fini(mm: &mm); |
| 363 | } |
| 364 | |
| 365 | static void drm_test_buddy_alloc_clear(struct kunit *test) |
| 366 | { |
| 367 | unsigned long n_pages, total, i = 0; |
| 368 | const unsigned long ps = SZ_4K; |
| 369 | struct drm_buddy_block *block; |
| 370 | const int max_order = 12; |
| 371 | LIST_HEAD(allocated); |
| 372 | struct drm_buddy mm; |
| 373 | unsigned int order; |
| 374 | u32 mm_size, size; |
| 375 | LIST_HEAD(dirty); |
| 376 | LIST_HEAD(clean); |
| 377 | |
| 378 | mm_size = SZ_4K << max_order; |
| 379 | KUNIT_EXPECT_FALSE(test, drm_buddy_init(&mm, mm_size, ps)); |
| 380 | |
| 381 | KUNIT_EXPECT_EQ(test, mm.max_order, max_order); |
| 382 | |
| 383 | /* |
| 384 | * Idea is to allocate and free some random portion of the address space, |
| 385 | * returning those pages as non-dirty and randomly alternate between |
| 386 | * requesting dirty and non-dirty pages (not going over the limit |
| 387 | * we freed as non-dirty), putting that into two separate lists. |
| 388 | * Loop over both lists at the end checking that the dirty list |
| 389 | * is indeed all dirty pages and vice versa. Free it all again, |
| 390 | * keeping the dirty/clear status. |
| 391 | */ |
| 392 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 393 | 5 * ps, ps, &allocated, |
| 394 | DRM_BUDDY_TOPDOWN_ALLOCATION), |
| 395 | "buddy_alloc hit an error size=%lu\n" , 5 * ps); |
| 396 | drm_buddy_free_list(mm: &mm, objects: &allocated, DRM_BUDDY_CLEARED); |
| 397 | |
| 398 | n_pages = 10; |
| 399 | do { |
| 400 | unsigned long flags; |
| 401 | struct list_head *list; |
| 402 | int slot = i % 2; |
| 403 | |
| 404 | if (slot == 0) { |
| 405 | list = &dirty; |
| 406 | flags = 0; |
| 407 | } else { |
| 408 | list = &clean; |
| 409 | flags = DRM_BUDDY_CLEAR_ALLOCATION; |
| 410 | } |
| 411 | |
| 412 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 413 | ps, ps, list, |
| 414 | flags), |
| 415 | "buddy_alloc hit an error size=%lu\n" , ps); |
| 416 | } while (++i < n_pages); |
| 417 | |
| 418 | list_for_each_entry(block, &clean, link) |
| 419 | KUNIT_EXPECT_EQ(test, drm_buddy_block_is_clear(block), true); |
| 420 | |
| 421 | list_for_each_entry(block, &dirty, link) |
| 422 | KUNIT_EXPECT_EQ(test, drm_buddy_block_is_clear(block), false); |
| 423 | |
| 424 | drm_buddy_free_list(mm: &mm, objects: &clean, DRM_BUDDY_CLEARED); |
| 425 | |
| 426 | /* |
| 427 | * Trying to go over the clear limit for some allocation. |
| 428 | * The allocation should never fail with reasonable page-size. |
| 429 | */ |
| 430 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 431 | 10 * ps, ps, &clean, |
| 432 | DRM_BUDDY_CLEAR_ALLOCATION), |
| 433 | "buddy_alloc hit an error size=%lu\n" , 10 * ps); |
| 434 | |
| 435 | drm_buddy_free_list(mm: &mm, objects: &clean, DRM_BUDDY_CLEARED); |
| 436 | drm_buddy_free_list(mm: &mm, objects: &dirty, flags: 0); |
| 437 | drm_buddy_fini(mm: &mm); |
| 438 | |
| 439 | KUNIT_EXPECT_FALSE(test, drm_buddy_init(&mm, mm_size, ps)); |
| 440 | |
| 441 | /* |
| 442 | * Create a new mm. Intentionally fragment the address space by creating |
| 443 | * two alternating lists. Free both lists, one as dirty the other as clean. |
| 444 | * Try to allocate double the previous size with matching min_page_size. The |
| 445 | * allocation should never fail as it calls the force_merge. Also check that |
| 446 | * the page is always dirty after force_merge. Free the page as dirty, then |
| 447 | * repeat the whole thing, increment the order until we hit the max_order. |
| 448 | */ |
| 449 | |
| 450 | i = 0; |
| 451 | n_pages = mm_size / ps; |
| 452 | do { |
| 453 | struct list_head *list; |
| 454 | int slot = i % 2; |
| 455 | |
| 456 | if (slot == 0) |
| 457 | list = &dirty; |
| 458 | else |
| 459 | list = &clean; |
| 460 | |
| 461 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 462 | ps, ps, list, 0), |
| 463 | "buddy_alloc hit an error size=%lu\n" , ps); |
| 464 | } while (++i < n_pages); |
| 465 | |
| 466 | drm_buddy_free_list(mm: &mm, objects: &clean, DRM_BUDDY_CLEARED); |
| 467 | drm_buddy_free_list(mm: &mm, objects: &dirty, flags: 0); |
| 468 | |
| 469 | order = 1; |
| 470 | do { |
| 471 | size = SZ_4K << order; |
| 472 | |
| 473 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 474 | size, size, &allocated, |
| 475 | DRM_BUDDY_CLEAR_ALLOCATION), |
| 476 | "buddy_alloc hit an error size=%u\n" , size); |
| 477 | total = 0; |
| 478 | list_for_each_entry(block, &allocated, link) { |
| 479 | if (size != mm_size) |
| 480 | KUNIT_EXPECT_EQ(test, drm_buddy_block_is_clear(block), false); |
| 481 | total += drm_buddy_block_size(mm: &mm, block); |
| 482 | } |
| 483 | KUNIT_EXPECT_EQ(test, total, size); |
| 484 | |
| 485 | drm_buddy_free_list(mm: &mm, objects: &allocated, flags: 0); |
| 486 | } while (++order <= max_order); |
| 487 | |
| 488 | drm_buddy_fini(mm: &mm); |
| 489 | |
| 490 | /* |
| 491 | * Create a new mm with a non power-of-two size. Allocate a random size from each |
| 492 | * root, free as cleared and then call fini. This will ensure the multi-root |
| 493 | * force merge during fini. |
| 494 | */ |
| 495 | mm_size = (SZ_4K << max_order) + (SZ_4K << (max_order - 2)); |
| 496 | |
| 497 | KUNIT_EXPECT_FALSE(test, drm_buddy_init(&mm, mm_size, ps)); |
| 498 | KUNIT_EXPECT_EQ(test, mm.max_order, max_order); |
| 499 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, SZ_4K << max_order, |
| 500 | 4 * ps, ps, &allocated, |
| 501 | DRM_BUDDY_RANGE_ALLOCATION), |
| 502 | "buddy_alloc hit an error size=%lu\n" , 4 * ps); |
| 503 | drm_buddy_free_list(mm: &mm, objects: &allocated, DRM_BUDDY_CLEARED); |
| 504 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, SZ_4K << max_order, |
| 505 | 2 * ps, ps, &allocated, |
| 506 | DRM_BUDDY_CLEAR_ALLOCATION), |
| 507 | "buddy_alloc hit an error size=%lu\n" , 2 * ps); |
| 508 | drm_buddy_free_list(mm: &mm, objects: &allocated, DRM_BUDDY_CLEARED); |
| 509 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, SZ_4K << max_order, mm_size, |
| 510 | ps, ps, &allocated, |
| 511 | DRM_BUDDY_RANGE_ALLOCATION), |
| 512 | "buddy_alloc hit an error size=%lu\n" , ps); |
| 513 | drm_buddy_free_list(mm: &mm, objects: &allocated, DRM_BUDDY_CLEARED); |
| 514 | drm_buddy_fini(mm: &mm); |
| 515 | } |
| 516 | |
| 517 | static void drm_test_buddy_alloc_contiguous(struct kunit *test) |
| 518 | { |
| 519 | const unsigned long ps = SZ_4K, mm_size = 16 * 3 * SZ_4K; |
| 520 | unsigned long i, n_pages, total; |
| 521 | struct drm_buddy_block *block; |
| 522 | struct drm_buddy mm; |
| 523 | LIST_HEAD(left); |
| 524 | LIST_HEAD(middle); |
| 525 | LIST_HEAD(right); |
| 526 | LIST_HEAD(allocated); |
| 527 | |
| 528 | KUNIT_EXPECT_FALSE(test, drm_buddy_init(&mm, mm_size, ps)); |
| 529 | |
| 530 | /* |
| 531 | * Idea is to fragment the address space by alternating block |
| 532 | * allocations between three different lists; one for left, middle and |
| 533 | * right. We can then free a list to simulate fragmentation. In |
| 534 | * particular we want to exercise the DRM_BUDDY_CONTIGUOUS_ALLOCATION, |
| 535 | * including the try_harder path. |
| 536 | */ |
| 537 | |
| 538 | i = 0; |
| 539 | n_pages = mm_size / ps; |
| 540 | do { |
| 541 | struct list_head *list; |
| 542 | int slot = i % 3; |
| 543 | |
| 544 | if (slot == 0) |
| 545 | list = &left; |
| 546 | else if (slot == 1) |
| 547 | list = &middle; |
| 548 | else |
| 549 | list = &right; |
| 550 | KUNIT_ASSERT_FALSE_MSG(test, |
| 551 | drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 552 | ps, ps, list, 0), |
| 553 | "buddy_alloc hit an error size=%lu\n" , |
| 554 | ps); |
| 555 | } while (++i < n_pages); |
| 556 | |
| 557 | KUNIT_ASSERT_TRUE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 558 | 3 * ps, ps, &allocated, |
| 559 | DRM_BUDDY_CONTIGUOUS_ALLOCATION), |
| 560 | "buddy_alloc didn't error size=%lu\n" , 3 * ps); |
| 561 | |
| 562 | drm_buddy_free_list(mm: &mm, objects: &middle, flags: 0); |
| 563 | KUNIT_ASSERT_TRUE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 564 | 3 * ps, ps, &allocated, |
| 565 | DRM_BUDDY_CONTIGUOUS_ALLOCATION), |
| 566 | "buddy_alloc didn't error size=%lu\n" , 3 * ps); |
| 567 | KUNIT_ASSERT_TRUE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 568 | 2 * ps, ps, &allocated, |
| 569 | DRM_BUDDY_CONTIGUOUS_ALLOCATION), |
| 570 | "buddy_alloc didn't error size=%lu\n" , 2 * ps); |
| 571 | |
| 572 | drm_buddy_free_list(mm: &mm, objects: &right, flags: 0); |
| 573 | KUNIT_ASSERT_TRUE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 574 | 3 * ps, ps, &allocated, |
| 575 | DRM_BUDDY_CONTIGUOUS_ALLOCATION), |
| 576 | "buddy_alloc didn't error size=%lu\n" , 3 * ps); |
| 577 | /* |
| 578 | * At this point we should have enough contiguous space for 2 blocks, |
| 579 | * however they are never buddies (since we freed middle and right) so |
| 580 | * will require the try_harder logic to find them. |
| 581 | */ |
| 582 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 583 | 2 * ps, ps, &allocated, |
| 584 | DRM_BUDDY_CONTIGUOUS_ALLOCATION), |
| 585 | "buddy_alloc hit an error size=%lu\n" , 2 * ps); |
| 586 | |
| 587 | drm_buddy_free_list(mm: &mm, objects: &left, flags: 0); |
| 588 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, 0, mm_size, |
| 589 | 3 * ps, ps, &allocated, |
| 590 | DRM_BUDDY_CONTIGUOUS_ALLOCATION), |
| 591 | "buddy_alloc hit an error size=%lu\n" , 3 * ps); |
| 592 | |
| 593 | total = 0; |
| 594 | list_for_each_entry(block, &allocated, link) |
| 595 | total += drm_buddy_block_size(mm: &mm, block); |
| 596 | |
| 597 | KUNIT_ASSERT_EQ(test, total, ps * 2 + ps * 3); |
| 598 | |
| 599 | drm_buddy_free_list(mm: &mm, objects: &allocated, flags: 0); |
| 600 | drm_buddy_fini(mm: &mm); |
| 601 | } |
| 602 | |
| 603 | static void drm_test_buddy_alloc_pathological(struct kunit *test) |
| 604 | { |
| 605 | u64 mm_size, size, start = 0; |
| 606 | struct drm_buddy_block *block; |
| 607 | const int max_order = 3; |
| 608 | unsigned long flags = 0; |
| 609 | int order, top; |
| 610 | struct drm_buddy mm; |
| 611 | LIST_HEAD(blocks); |
| 612 | LIST_HEAD(holes); |
| 613 | LIST_HEAD(tmp); |
| 614 | |
| 615 | /* |
| 616 | * Create a pot-sized mm, then allocate one of each possible |
| 617 | * order within. This should leave the mm with exactly one |
| 618 | * page left. Free the largest block, then whittle down again. |
| 619 | * Eventually we will have a fully 50% fragmented mm. |
| 620 | */ |
| 621 | |
| 622 | mm_size = SZ_4K << max_order; |
| 623 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_init(&mm, mm_size, SZ_4K), |
| 624 | "buddy_init failed\n" ); |
| 625 | |
| 626 | KUNIT_EXPECT_EQ(test, mm.max_order, max_order); |
| 627 | |
| 628 | for (top = max_order; top; top--) { |
| 629 | /* Make room by freeing the largest allocated block */ |
| 630 | block = list_first_entry_or_null(&blocks, typeof(*block), link); |
| 631 | if (block) { |
| 632 | list_del(entry: &block->link); |
| 633 | drm_buddy_free_block(mm: &mm, block); |
| 634 | } |
| 635 | |
| 636 | for (order = top; order--;) { |
| 637 | size = get_size(order, chunk_size: mm.chunk_size); |
| 638 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, start, |
| 639 | mm_size, size, size, |
| 640 | &tmp, flags), |
| 641 | "buddy_alloc hit -ENOMEM with order=%d, top=%d\n" , |
| 642 | order, top); |
| 643 | |
| 644 | block = list_first_entry_or_null(&tmp, struct drm_buddy_block, link); |
| 645 | KUNIT_ASSERT_TRUE_MSG(test, block, "alloc_blocks has no blocks\n" ); |
| 646 | |
| 647 | list_move_tail(list: &block->link, head: &blocks); |
| 648 | } |
| 649 | |
| 650 | /* There should be one final page for this sub-allocation */ |
| 651 | size = get_size(order: 0, chunk_size: mm.chunk_size); |
| 652 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, start, mm_size, |
| 653 | size, size, &tmp, flags), |
| 654 | "buddy_alloc hit -ENOMEM for hole\n" ); |
| 655 | |
| 656 | block = list_first_entry_or_null(&tmp, struct drm_buddy_block, link); |
| 657 | KUNIT_ASSERT_TRUE_MSG(test, block, "alloc_blocks has no blocks\n" ); |
| 658 | |
| 659 | list_move_tail(list: &block->link, head: &holes); |
| 660 | |
| 661 | size = get_size(order: top, chunk_size: mm.chunk_size); |
| 662 | KUNIT_ASSERT_TRUE_MSG(test, drm_buddy_alloc_blocks(&mm, start, mm_size, |
| 663 | size, size, &tmp, flags), |
| 664 | "buddy_alloc unexpectedly succeeded at top-order %d/%d, it should be full!" , |
| 665 | top, max_order); |
| 666 | } |
| 667 | |
| 668 | drm_buddy_free_list(mm: &mm, objects: &holes, flags: 0); |
| 669 | |
| 670 | /* Nothing larger than blocks of chunk_size now available */ |
| 671 | for (order = 1; order <= max_order; order++) { |
| 672 | size = get_size(order, chunk_size: mm.chunk_size); |
| 673 | KUNIT_ASSERT_TRUE_MSG(test, drm_buddy_alloc_blocks(&mm, start, mm_size, |
| 674 | size, size, &tmp, flags), |
| 675 | "buddy_alloc unexpectedly succeeded at order %d, it should be full!" , |
| 676 | order); |
| 677 | } |
| 678 | |
| 679 | list_splice_tail(list: &holes, head: &blocks); |
| 680 | drm_buddy_free_list(mm: &mm, objects: &blocks, flags: 0); |
| 681 | drm_buddy_fini(mm: &mm); |
| 682 | } |
| 683 | |
| 684 | static void drm_test_buddy_alloc_pessimistic(struct kunit *test) |
| 685 | { |
| 686 | u64 mm_size, size, start = 0; |
| 687 | struct drm_buddy_block *block, *bn; |
| 688 | const unsigned int max_order = 16; |
| 689 | unsigned long flags = 0; |
| 690 | struct drm_buddy mm; |
| 691 | unsigned int order; |
| 692 | LIST_HEAD(blocks); |
| 693 | LIST_HEAD(tmp); |
| 694 | |
| 695 | /* |
| 696 | * Create a pot-sized mm, then allocate one of each possible |
| 697 | * order within. This should leave the mm with exactly one |
| 698 | * page left. |
| 699 | */ |
| 700 | |
| 701 | mm_size = SZ_4K << max_order; |
| 702 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_init(&mm, mm_size, SZ_4K), |
| 703 | "buddy_init failed\n" ); |
| 704 | |
| 705 | KUNIT_EXPECT_EQ(test, mm.max_order, max_order); |
| 706 | |
| 707 | for (order = 0; order < max_order; order++) { |
| 708 | size = get_size(order, chunk_size: mm.chunk_size); |
| 709 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, start, mm_size, |
| 710 | size, size, &tmp, flags), |
| 711 | "buddy_alloc hit -ENOMEM with order=%d\n" , |
| 712 | order); |
| 713 | |
| 714 | block = list_first_entry_or_null(&tmp, struct drm_buddy_block, link); |
| 715 | KUNIT_ASSERT_TRUE_MSG(test, block, "alloc_blocks has no blocks\n" ); |
| 716 | |
| 717 | list_move_tail(list: &block->link, head: &blocks); |
| 718 | } |
| 719 | |
| 720 | /* And now the last remaining block available */ |
| 721 | size = get_size(order: 0, chunk_size: mm.chunk_size); |
| 722 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, start, mm_size, |
| 723 | size, size, &tmp, flags), |
| 724 | "buddy_alloc hit -ENOMEM on final alloc\n" ); |
| 725 | |
| 726 | block = list_first_entry_or_null(&tmp, struct drm_buddy_block, link); |
| 727 | KUNIT_ASSERT_TRUE_MSG(test, block, "alloc_blocks has no blocks\n" ); |
| 728 | |
| 729 | list_move_tail(list: &block->link, head: &blocks); |
| 730 | |
| 731 | /* Should be completely full! */ |
| 732 | for (order = max_order; order--;) { |
| 733 | size = get_size(order, chunk_size: mm.chunk_size); |
| 734 | KUNIT_ASSERT_TRUE_MSG(test, drm_buddy_alloc_blocks(&mm, start, mm_size, |
| 735 | size, size, &tmp, flags), |
| 736 | "buddy_alloc unexpectedly succeeded, it should be full!" ); |
| 737 | } |
| 738 | |
| 739 | block = list_last_entry(&blocks, typeof(*block), link); |
| 740 | list_del(entry: &block->link); |
| 741 | drm_buddy_free_block(mm: &mm, block); |
| 742 | |
| 743 | /* As we free in increasing size, we make available larger blocks */ |
| 744 | order = 1; |
| 745 | list_for_each_entry_safe(block, bn, &blocks, link) { |
| 746 | list_del(entry: &block->link); |
| 747 | drm_buddy_free_block(mm: &mm, block); |
| 748 | |
| 749 | size = get_size(order, chunk_size: mm.chunk_size); |
| 750 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, start, mm_size, |
| 751 | size, size, &tmp, flags), |
| 752 | "buddy_alloc hit -ENOMEM with order=%d\n" , |
| 753 | order); |
| 754 | |
| 755 | block = list_first_entry_or_null(&tmp, struct drm_buddy_block, link); |
| 756 | KUNIT_ASSERT_TRUE_MSG(test, block, "alloc_blocks has no blocks\n" ); |
| 757 | |
| 758 | list_del(entry: &block->link); |
| 759 | drm_buddy_free_block(mm: &mm, block); |
| 760 | order++; |
| 761 | } |
| 762 | |
| 763 | /* To confirm, now the whole mm should be available */ |
| 764 | size = get_size(order: max_order, chunk_size: mm.chunk_size); |
| 765 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, start, mm_size, |
| 766 | size, size, &tmp, flags), |
| 767 | "buddy_alloc (realloc) hit -ENOMEM with order=%d\n" , |
| 768 | max_order); |
| 769 | |
| 770 | block = list_first_entry_or_null(&tmp, struct drm_buddy_block, link); |
| 771 | KUNIT_ASSERT_TRUE_MSG(test, block, "alloc_blocks has no blocks\n" ); |
| 772 | |
| 773 | list_del(entry: &block->link); |
| 774 | drm_buddy_free_block(mm: &mm, block); |
| 775 | drm_buddy_free_list(mm: &mm, objects: &blocks, flags: 0); |
| 776 | drm_buddy_fini(mm: &mm); |
| 777 | } |
| 778 | |
| 779 | static void drm_test_buddy_alloc_optimistic(struct kunit *test) |
| 780 | { |
| 781 | u64 mm_size, size, start = 0; |
| 782 | struct drm_buddy_block *block; |
| 783 | unsigned long flags = 0; |
| 784 | const int max_order = 16; |
| 785 | struct drm_buddy mm; |
| 786 | LIST_HEAD(blocks); |
| 787 | LIST_HEAD(tmp); |
| 788 | int order; |
| 789 | |
| 790 | /* |
| 791 | * Create a mm with one block of each order available, and |
| 792 | * try to allocate them all. |
| 793 | */ |
| 794 | |
| 795 | mm_size = SZ_4K * ((1 << (max_order + 1)) - 1); |
| 796 | |
| 797 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_init(&mm, mm_size, SZ_4K), |
| 798 | "buddy_init failed\n" ); |
| 799 | |
| 800 | KUNIT_EXPECT_EQ(test, mm.max_order, max_order); |
| 801 | |
| 802 | for (order = 0; order <= max_order; order++) { |
| 803 | size = get_size(order, chunk_size: mm.chunk_size); |
| 804 | KUNIT_ASSERT_FALSE_MSG(test, drm_buddy_alloc_blocks(&mm, start, mm_size, |
| 805 | size, size, &tmp, flags), |
| 806 | "buddy_alloc hit -ENOMEM with order=%d\n" , |
| 807 | order); |
| 808 | |
| 809 | block = list_first_entry_or_null(&tmp, struct drm_buddy_block, link); |
| 810 | KUNIT_ASSERT_TRUE_MSG(test, block, "alloc_blocks has no blocks\n" ); |
| 811 | |
| 812 | list_move_tail(list: &block->link, head: &blocks); |
| 813 | } |
| 814 | |
| 815 | /* Should be completely full! */ |
| 816 | size = get_size(order: 0, chunk_size: mm.chunk_size); |
| 817 | KUNIT_ASSERT_TRUE_MSG(test, drm_buddy_alloc_blocks(&mm, start, mm_size, |
| 818 | size, size, &tmp, flags), |
| 819 | "buddy_alloc unexpectedly succeeded, it should be full!" ); |
| 820 | |
| 821 | drm_buddy_free_list(mm: &mm, objects: &blocks, flags: 0); |
| 822 | drm_buddy_fini(mm: &mm); |
| 823 | } |
| 824 | |
| 825 | static void drm_test_buddy_alloc_limit(struct kunit *test) |
| 826 | { |
| 827 | u64 size = U64_MAX, start = 0; |
| 828 | struct drm_buddy_block *block; |
| 829 | unsigned long flags = 0; |
| 830 | LIST_HEAD(allocated); |
| 831 | struct drm_buddy mm; |
| 832 | |
| 833 | KUNIT_EXPECT_FALSE(test, drm_buddy_init(&mm, size, SZ_4K)); |
| 834 | |
| 835 | KUNIT_EXPECT_EQ_MSG(test, mm.max_order, DRM_BUDDY_MAX_ORDER, |
| 836 | "mm.max_order(%d) != %d\n" , mm.max_order, |
| 837 | DRM_BUDDY_MAX_ORDER); |
| 838 | |
| 839 | size = mm.chunk_size << mm.max_order; |
| 840 | KUNIT_EXPECT_FALSE(test, drm_buddy_alloc_blocks(&mm, start, size, size, |
| 841 | mm.chunk_size, &allocated, flags)); |
| 842 | |
| 843 | block = list_first_entry_or_null(&allocated, struct drm_buddy_block, link); |
| 844 | KUNIT_EXPECT_TRUE(test, block); |
| 845 | |
| 846 | KUNIT_EXPECT_EQ_MSG(test, drm_buddy_block_order(block), mm.max_order, |
| 847 | "block order(%d) != %d\n" , |
| 848 | drm_buddy_block_order(block), mm.max_order); |
| 849 | |
| 850 | KUNIT_EXPECT_EQ_MSG(test, drm_buddy_block_size(&mm, block), |
| 851 | BIT_ULL(mm.max_order) * mm.chunk_size, |
| 852 | "block size(%llu) != %llu\n" , |
| 853 | drm_buddy_block_size(&mm, block), |
| 854 | BIT_ULL(mm.max_order) * mm.chunk_size); |
| 855 | |
| 856 | drm_buddy_free_list(mm: &mm, objects: &allocated, flags: 0); |
| 857 | drm_buddy_fini(mm: &mm); |
| 858 | } |
| 859 | |
| 860 | static int drm_buddy_suite_init(struct kunit_suite *suite) |
| 861 | { |
| 862 | while (!random_seed) |
| 863 | random_seed = get_random_u32(); |
| 864 | |
| 865 | kunit_info(suite, "Testing DRM buddy manager, with random_seed=0x%x\n" , |
| 866 | random_seed); |
| 867 | |
| 868 | return 0; |
| 869 | } |
| 870 | |
| 871 | static struct kunit_case drm_buddy_tests[] = { |
| 872 | KUNIT_CASE(drm_test_buddy_alloc_limit), |
| 873 | KUNIT_CASE(drm_test_buddy_alloc_optimistic), |
| 874 | KUNIT_CASE(drm_test_buddy_alloc_pessimistic), |
| 875 | KUNIT_CASE(drm_test_buddy_alloc_pathological), |
| 876 | KUNIT_CASE(drm_test_buddy_alloc_contiguous), |
| 877 | KUNIT_CASE(drm_test_buddy_alloc_clear), |
| 878 | KUNIT_CASE(drm_test_buddy_alloc_range_bias), |
| 879 | KUNIT_CASE(drm_test_buddy_fragmentation_performance), |
| 880 | {} |
| 881 | }; |
| 882 | |
| 883 | static struct kunit_suite drm_buddy_test_suite = { |
| 884 | .name = "drm_buddy" , |
| 885 | .suite_init = drm_buddy_suite_init, |
| 886 | .test_cases = drm_buddy_tests, |
| 887 | }; |
| 888 | |
| 889 | kunit_test_suite(drm_buddy_test_suite); |
| 890 | |
| 891 | MODULE_AUTHOR("Intel Corporation" ); |
| 892 | MODULE_DESCRIPTION("Kunit test for drm_buddy functions" ); |
| 893 | MODULE_LICENSE("GPL" ); |
| 894 | |