1 /* SPDX-License-Identifier: GPL-2.0 2 * 3 * page_pool.c 4 * Author: Jesper Dangaard Brouer <netoptimizer@brouer.com> 5 * Copyright (C) 2016 Red Hat, Inc. 6 */ 7 8 #include <linux/types.h> 9 #include <linux/kernel.h> 10 #include <linux/slab.h> 11 #include <linux/device.h> 12 13 #include <net/page_pool.h> 14 #include <linux/dma-direction.h> 15 #include <linux/dma-mapping.h> 16 #include <linux/page-flags.h> 17 #include <linux/mm.h> /* for __put_page() */ 18 19 #include <trace/events/page_pool.h> 20 21 static int page_pool_init(struct page_pool *pool, 22 const struct page_pool_params *params) 23 { 24 unsigned int ring_qsize = 1024; /* Default */ 25 26 memcpy(&pool->p, params, sizeof(pool->p)); 27 28 /* Validate only known flags were used */ 29 if (pool->p.flags & ~(PP_FLAG_ALL)) 30 return -EINVAL; 31 32 if (pool->p.pool_size) 33 ring_qsize = pool->p.pool_size; 34 35 /* Sanity limit mem that can be pinned down */ 36 if (ring_qsize > 32768) 37 return -E2BIG; 38 39 /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL. 40 * DMA_BIDIRECTIONAL is for allowing page used for DMA sending, 41 * which is the XDP_TX use-case. 42 */ 43 if ((pool->p.dma_dir != DMA_FROM_DEVICE) && 44 (pool->p.dma_dir != DMA_BIDIRECTIONAL)) 45 return -EINVAL; 46 47 if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0) 48 return -ENOMEM; 49 50 atomic_set(&pool->pages_state_release_cnt, 0); 51 52 if (pool->p.flags & PP_FLAG_DMA_MAP) 53 get_device(pool->p.dev); 54 55 return 0; 56 } 57 58 struct page_pool *page_pool_create(const struct page_pool_params *params) 59 { 60 struct page_pool *pool; 61 int err = 0; 62 63 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid); 64 if (!pool) 65 return ERR_PTR(-ENOMEM); 66 67 err = page_pool_init(pool, params); 68 if (err < 0) { 69 pr_warn("%s() gave up with errno %d\n", __func__, err); 70 kfree(pool); 71 return ERR_PTR(err); 72 } 73 return pool; 74 } 75 EXPORT_SYMBOL(page_pool_create); 76 77 /* fast path */ 78 static struct page *__page_pool_get_cached(struct page_pool *pool) 79 { 80 struct ptr_ring *r = &pool->ring; 81 struct page *page; 82 83 /* Quicker fallback, avoid locks when ring is empty */ 84 if (__ptr_ring_empty(r)) 85 return NULL; 86 87 /* Test for safe-context, caller should provide this guarantee */ 88 if (likely(in_serving_softirq())) { 89 if (likely(pool->alloc.count)) { 90 /* Fast-path */ 91 page = pool->alloc.cache[--pool->alloc.count]; 92 return page; 93 } 94 /* Slower-path: Alloc array empty, time to refill 95 * 96 * Open-coded bulk ptr_ring consumer. 97 * 98 * Discussion: the ring consumer lock is not really 99 * needed due to the softirq/NAPI protection, but 100 * later need the ability to reclaim pages on the 101 * ring. Thus, keeping the locks. 102 */ 103 spin_lock(&r->consumer_lock); 104 while ((page = __ptr_ring_consume(r))) { 105 if (pool->alloc.count == PP_ALLOC_CACHE_REFILL) 106 break; 107 pool->alloc.cache[pool->alloc.count++] = page; 108 } 109 spin_unlock(&r->consumer_lock); 110 return page; 111 } 112 113 /* Slow-path: Get page from locked ring queue */ 114 page = ptr_ring_consume(&pool->ring); 115 return page; 116 } 117 118 /* slow path */ 119 noinline 120 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool, 121 gfp_t _gfp) 122 { 123 struct page *page; 124 gfp_t gfp = _gfp; 125 dma_addr_t dma; 126 127 /* We could always set __GFP_COMP, and avoid this branch, as 128 * prep_new_page() can handle order-0 with __GFP_COMP. 129 */ 130 if (pool->p.order) 131 gfp |= __GFP_COMP; 132 133 /* FUTURE development: 134 * 135 * Current slow-path essentially falls back to single page 136 * allocations, which doesn't improve performance. This code 137 * need bulk allocation support from the page allocator code. 138 */ 139 140 /* Cache was empty, do real allocation */ 141 page = alloc_pages_node(pool->p.nid, gfp, pool->p.order); 142 if (!page) 143 return NULL; 144 145 if (!(pool->p.flags & PP_FLAG_DMA_MAP)) 146 goto skip_dma_map; 147 148 /* Setup DMA mapping: use 'struct page' area for storing DMA-addr 149 * since dma_addr_t can be either 32 or 64 bits and does not always fit 150 * into page private data (i.e 32bit cpu with 64bit DMA caps) 151 * This mapping is kept for lifetime of page, until leaving pool. 152 */ 153 dma = dma_map_page_attrs(pool->p.dev, page, 0, 154 (PAGE_SIZE << pool->p.order), 155 pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC); 156 if (dma_mapping_error(pool->p.dev, dma)) { 157 put_page(page); 158 return NULL; 159 } 160 page->dma_addr = dma; 161 162 skip_dma_map: 163 /* Track how many pages are held 'in-flight' */ 164 pool->pages_state_hold_cnt++; 165 166 trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt); 167 168 /* When page just alloc'ed is should/must have refcnt 1. */ 169 return page; 170 } 171 172 /* For using page_pool replace: alloc_pages() API calls, but provide 173 * synchronization guarantee for allocation side. 174 */ 175 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp) 176 { 177 struct page *page; 178 179 /* Fast-path: Get a page from cache */ 180 page = __page_pool_get_cached(pool); 181 if (page) 182 return page; 183 184 /* Slow-path: cache empty, do real allocation */ 185 page = __page_pool_alloc_pages_slow(pool, gfp); 186 return page; 187 } 188 EXPORT_SYMBOL(page_pool_alloc_pages); 189 190 /* Calculate distance between two u32 values, valid if distance is below 2^(31) 191 * https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution 192 */ 193 #define _distance(a, b) (s32)((a) - (b)) 194 195 static s32 page_pool_inflight(struct page_pool *pool) 196 { 197 u32 release_cnt = atomic_read(&pool->pages_state_release_cnt); 198 u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt); 199 s32 distance; 200 201 distance = _distance(hold_cnt, release_cnt); 202 203 trace_page_pool_inflight(pool, distance, hold_cnt, release_cnt); 204 return distance; 205 } 206 207 static bool __page_pool_safe_to_destroy(struct page_pool *pool) 208 { 209 s32 inflight = page_pool_inflight(pool); 210 211 /* The distance should not be able to become negative */ 212 WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight); 213 214 return (inflight == 0); 215 } 216 217 /* Cleanup page_pool state from page */ 218 static void __page_pool_clean_page(struct page_pool *pool, 219 struct page *page) 220 { 221 dma_addr_t dma; 222 223 if (!(pool->p.flags & PP_FLAG_DMA_MAP)) 224 goto skip_dma_unmap; 225 226 dma = page->dma_addr; 227 /* DMA unmap */ 228 dma_unmap_page_attrs(pool->p.dev, dma, 229 PAGE_SIZE << pool->p.order, pool->p.dma_dir, 230 DMA_ATTR_SKIP_CPU_SYNC); 231 page->dma_addr = 0; 232 skip_dma_unmap: 233 atomic_inc(&pool->pages_state_release_cnt); 234 trace_page_pool_state_release(pool, page, 235 atomic_read(&pool->pages_state_release_cnt)); 236 } 237 238 /* unmap the page and clean our state */ 239 void page_pool_unmap_page(struct page_pool *pool, struct page *page) 240 { 241 /* When page is unmapped, this implies page will not be 242 * returned to page_pool. 243 */ 244 __page_pool_clean_page(pool, page); 245 } 246 EXPORT_SYMBOL(page_pool_unmap_page); 247 248 /* Return a page to the page allocator, cleaning up our state */ 249 static void __page_pool_return_page(struct page_pool *pool, struct page *page) 250 { 251 __page_pool_clean_page(pool, page); 252 253 put_page(page); 254 /* An optimization would be to call __free_pages(page, pool->p.order) 255 * knowing page is not part of page-cache (thus avoiding a 256 * __page_cache_release() call). 257 */ 258 } 259 260 static bool __page_pool_recycle_into_ring(struct page_pool *pool, 261 struct page *page) 262 { 263 int ret; 264 /* BH protection not needed if current is serving softirq */ 265 if (in_serving_softirq()) 266 ret = ptr_ring_produce(&pool->ring, page); 267 else 268 ret = ptr_ring_produce_bh(&pool->ring, page); 269 270 return (ret == 0) ? true : false; 271 } 272 273 /* Only allow direct recycling in special circumstances, into the 274 * alloc side cache. E.g. during RX-NAPI processing for XDP_DROP use-case. 275 * 276 * Caller must provide appropriate safe context. 277 */ 278 static bool __page_pool_recycle_direct(struct page *page, 279 struct page_pool *pool) 280 { 281 if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) 282 return false; 283 284 /* Caller MUST have verified/know (page_ref_count(page) == 1) */ 285 pool->alloc.cache[pool->alloc.count++] = page; 286 return true; 287 } 288 289 void __page_pool_put_page(struct page_pool *pool, 290 struct page *page, bool allow_direct) 291 { 292 /* This allocator is optimized for the XDP mode that uses 293 * one-frame-per-page, but have fallbacks that act like the 294 * regular page allocator APIs. 295 * 296 * refcnt == 1 means page_pool owns page, and can recycle it. 297 */ 298 if (likely(page_ref_count(page) == 1)) { 299 /* Read barrier done in page_ref_count / READ_ONCE */ 300 301 if (allow_direct && in_serving_softirq()) 302 if (__page_pool_recycle_direct(page, pool)) 303 return; 304 305 if (!__page_pool_recycle_into_ring(pool, page)) { 306 /* Cache full, fallback to free pages */ 307 __page_pool_return_page(pool, page); 308 } 309 return; 310 } 311 /* Fallback/non-XDP mode: API user have elevated refcnt. 312 * 313 * Many drivers split up the page into fragments, and some 314 * want to keep doing this to save memory and do refcnt based 315 * recycling. Support this use case too, to ease drivers 316 * switching between XDP/non-XDP. 317 * 318 * In-case page_pool maintains the DMA mapping, API user must 319 * call page_pool_put_page once. In this elevated refcnt 320 * case, the DMA is unmapped/released, as driver is likely 321 * doing refcnt based recycle tricks, meaning another process 322 * will be invoking put_page. 323 */ 324 __page_pool_clean_page(pool, page); 325 put_page(page); 326 } 327 EXPORT_SYMBOL(__page_pool_put_page); 328 329 static void __page_pool_empty_ring(struct page_pool *pool) 330 { 331 struct page *page; 332 333 /* Empty recycle ring */ 334 while ((page = ptr_ring_consume_bh(&pool->ring))) { 335 /* Verify the refcnt invariant of cached pages */ 336 if (!(page_ref_count(page) == 1)) 337 pr_crit("%s() page_pool refcnt %d violation\n", 338 __func__, page_ref_count(page)); 339 340 __page_pool_return_page(pool, page); 341 } 342 } 343 344 static void __warn_in_flight(struct page_pool *pool) 345 { 346 u32 release_cnt = atomic_read(&pool->pages_state_release_cnt); 347 u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt); 348 s32 distance; 349 350 distance = _distance(hold_cnt, release_cnt); 351 352 /* Drivers should fix this, but only problematic when DMA is used */ 353 WARN(1, "Still in-flight pages:%d hold:%u released:%u", 354 distance, hold_cnt, release_cnt); 355 } 356 357 void __page_pool_free(struct page_pool *pool) 358 { 359 WARN(pool->alloc.count, "API usage violation"); 360 WARN(!ptr_ring_empty(&pool->ring), "ptr_ring is not empty"); 361 362 /* Can happen due to forced shutdown */ 363 if (!__page_pool_safe_to_destroy(pool)) 364 __warn_in_flight(pool); 365 366 ptr_ring_cleanup(&pool->ring, NULL); 367 368 if (pool->p.flags & PP_FLAG_DMA_MAP) 369 put_device(pool->p.dev); 370 371 kfree(pool); 372 } 373 EXPORT_SYMBOL(__page_pool_free); 374 375 /* Request to shutdown: release pages cached by page_pool, and check 376 * for in-flight pages 377 */ 378 bool __page_pool_request_shutdown(struct page_pool *pool) 379 { 380 struct page *page; 381 382 /* Empty alloc cache, assume caller made sure this is 383 * no-longer in use, and page_pool_alloc_pages() cannot be 384 * call concurrently. 385 */ 386 while (pool->alloc.count) { 387 page = pool->alloc.cache[--pool->alloc.count]; 388 __page_pool_return_page(pool, page); 389 } 390 391 /* No more consumers should exist, but producers could still 392 * be in-flight. 393 */ 394 __page_pool_empty_ring(pool); 395 396 return __page_pool_safe_to_destroy(pool); 397 } 398 EXPORT_SYMBOL(__page_pool_request_shutdown); 399