1 /* 2 * Copyright (c) 2006 Oracle. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 * 32 */ 33 #include <linux/kernel.h> 34 #include <linux/slab.h> 35 #include <linux/rculist.h> 36 #include <linux/llist.h> 37 38 #include "rds_single_path.h" 39 #include "ib_mr.h" 40 41 struct workqueue_struct *rds_ib_mr_wq; 42 43 static DEFINE_PER_CPU(unsigned long, clean_list_grace); 44 #define CLEAN_LIST_BUSY_BIT 0 45 46 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr) 47 { 48 struct rds_ib_device *rds_ibdev; 49 struct rds_ib_ipaddr *i_ipaddr; 50 51 rcu_read_lock(); 52 list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) { 53 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) { 54 if (i_ipaddr->ipaddr == ipaddr) { 55 atomic_inc(&rds_ibdev->refcount); 56 rcu_read_unlock(); 57 return rds_ibdev; 58 } 59 } 60 } 61 rcu_read_unlock(); 62 63 return NULL; 64 } 65 66 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr) 67 { 68 struct rds_ib_ipaddr *i_ipaddr; 69 70 i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL); 71 if (!i_ipaddr) 72 return -ENOMEM; 73 74 i_ipaddr->ipaddr = ipaddr; 75 76 spin_lock_irq(&rds_ibdev->spinlock); 77 list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list); 78 spin_unlock_irq(&rds_ibdev->spinlock); 79 80 return 0; 81 } 82 83 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr) 84 { 85 struct rds_ib_ipaddr *i_ipaddr; 86 struct rds_ib_ipaddr *to_free = NULL; 87 88 89 spin_lock_irq(&rds_ibdev->spinlock); 90 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) { 91 if (i_ipaddr->ipaddr == ipaddr) { 92 list_del_rcu(&i_ipaddr->list); 93 to_free = i_ipaddr; 94 break; 95 } 96 } 97 spin_unlock_irq(&rds_ibdev->spinlock); 98 99 if (to_free) 100 kfree_rcu(to_free, rcu); 101 } 102 103 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr) 104 { 105 struct rds_ib_device *rds_ibdev_old; 106 107 rds_ibdev_old = rds_ib_get_device(ipaddr); 108 if (!rds_ibdev_old) 109 return rds_ib_add_ipaddr(rds_ibdev, ipaddr); 110 111 if (rds_ibdev_old != rds_ibdev) { 112 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr); 113 rds_ib_dev_put(rds_ibdev_old); 114 return rds_ib_add_ipaddr(rds_ibdev, ipaddr); 115 } 116 rds_ib_dev_put(rds_ibdev_old); 117 118 return 0; 119 } 120 121 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn) 122 { 123 struct rds_ib_connection *ic = conn->c_transport_data; 124 125 /* conn was previously on the nodev_conns_list */ 126 spin_lock_irq(&ib_nodev_conns_lock); 127 BUG_ON(list_empty(&ib_nodev_conns)); 128 BUG_ON(list_empty(&ic->ib_node)); 129 list_del(&ic->ib_node); 130 131 spin_lock(&rds_ibdev->spinlock); 132 list_add_tail(&ic->ib_node, &rds_ibdev->conn_list); 133 spin_unlock(&rds_ibdev->spinlock); 134 spin_unlock_irq(&ib_nodev_conns_lock); 135 136 ic->rds_ibdev = rds_ibdev; 137 atomic_inc(&rds_ibdev->refcount); 138 } 139 140 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn) 141 { 142 struct rds_ib_connection *ic = conn->c_transport_data; 143 144 /* place conn on nodev_conns_list */ 145 spin_lock(&ib_nodev_conns_lock); 146 147 spin_lock_irq(&rds_ibdev->spinlock); 148 BUG_ON(list_empty(&ic->ib_node)); 149 list_del(&ic->ib_node); 150 spin_unlock_irq(&rds_ibdev->spinlock); 151 152 list_add_tail(&ic->ib_node, &ib_nodev_conns); 153 154 spin_unlock(&ib_nodev_conns_lock); 155 156 ic->rds_ibdev = NULL; 157 rds_ib_dev_put(rds_ibdev); 158 } 159 160 void rds_ib_destroy_nodev_conns(void) 161 { 162 struct rds_ib_connection *ic, *_ic; 163 LIST_HEAD(tmp_list); 164 165 /* avoid calling conn_destroy with irqs off */ 166 spin_lock_irq(&ib_nodev_conns_lock); 167 list_splice(&ib_nodev_conns, &tmp_list); 168 spin_unlock_irq(&ib_nodev_conns_lock); 169 170 list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node) 171 rds_conn_destroy(ic->conn); 172 } 173 174 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo) 175 { 176 struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool; 177 178 iinfo->rdma_mr_max = pool_1m->max_items; 179 iinfo->rdma_mr_size = pool_1m->fmr_attr.max_pages; 180 } 181 182 struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool) 183 { 184 struct rds_ib_mr *ibmr = NULL; 185 struct llist_node *ret; 186 unsigned long *flag; 187 188 preempt_disable(); 189 flag = this_cpu_ptr(&clean_list_grace); 190 set_bit(CLEAN_LIST_BUSY_BIT, flag); 191 ret = llist_del_first(&pool->clean_list); 192 if (ret) { 193 ibmr = llist_entry(ret, struct rds_ib_mr, llnode); 194 if (pool->pool_type == RDS_IB_MR_8K_POOL) 195 rds_ib_stats_inc(s_ib_rdma_mr_8k_reused); 196 else 197 rds_ib_stats_inc(s_ib_rdma_mr_1m_reused); 198 } 199 200 clear_bit(CLEAN_LIST_BUSY_BIT, flag); 201 preempt_enable(); 202 return ibmr; 203 } 204 205 static inline void wait_clean_list_grace(void) 206 { 207 int cpu; 208 unsigned long *flag; 209 210 for_each_online_cpu(cpu) { 211 flag = &per_cpu(clean_list_grace, cpu); 212 while (test_bit(CLEAN_LIST_BUSY_BIT, flag)) 213 cpu_relax(); 214 } 215 } 216 217 void rds_ib_sync_mr(void *trans_private, int direction) 218 { 219 struct rds_ib_mr *ibmr = trans_private; 220 struct rds_ib_device *rds_ibdev = ibmr->device; 221 222 switch (direction) { 223 case DMA_FROM_DEVICE: 224 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg, 225 ibmr->sg_dma_len, DMA_BIDIRECTIONAL); 226 break; 227 case DMA_TO_DEVICE: 228 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg, 229 ibmr->sg_dma_len, DMA_BIDIRECTIONAL); 230 break; 231 } 232 } 233 234 void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr) 235 { 236 struct rds_ib_device *rds_ibdev = ibmr->device; 237 238 if (ibmr->sg_dma_len) { 239 ib_dma_unmap_sg(rds_ibdev->dev, 240 ibmr->sg, ibmr->sg_len, 241 DMA_BIDIRECTIONAL); 242 ibmr->sg_dma_len = 0; 243 } 244 245 /* Release the s/g list */ 246 if (ibmr->sg_len) { 247 unsigned int i; 248 249 for (i = 0; i < ibmr->sg_len; ++i) { 250 struct page *page = sg_page(&ibmr->sg[i]); 251 252 /* FIXME we need a way to tell a r/w MR 253 * from a r/o MR */ 254 WARN_ON(!page->mapping && irqs_disabled()); 255 set_page_dirty(page); 256 put_page(page); 257 } 258 kfree(ibmr->sg); 259 260 ibmr->sg = NULL; 261 ibmr->sg_len = 0; 262 } 263 } 264 265 void rds_ib_teardown_mr(struct rds_ib_mr *ibmr) 266 { 267 unsigned int pinned = ibmr->sg_len; 268 269 __rds_ib_teardown_mr(ibmr); 270 if (pinned) { 271 struct rds_ib_mr_pool *pool = ibmr->pool; 272 273 atomic_sub(pinned, &pool->free_pinned); 274 } 275 } 276 277 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all) 278 { 279 unsigned int item_count; 280 281 item_count = atomic_read(&pool->item_count); 282 if (free_all) 283 return item_count; 284 285 return 0; 286 } 287 288 /* 289 * given an llist of mrs, put them all into the list_head for more processing 290 */ 291 static unsigned int llist_append_to_list(struct llist_head *llist, 292 struct list_head *list) 293 { 294 struct rds_ib_mr *ibmr; 295 struct llist_node *node; 296 struct llist_node *next; 297 unsigned int count = 0; 298 299 node = llist_del_all(llist); 300 while (node) { 301 next = node->next; 302 ibmr = llist_entry(node, struct rds_ib_mr, llnode); 303 list_add_tail(&ibmr->unmap_list, list); 304 node = next; 305 count++; 306 } 307 return count; 308 } 309 310 /* 311 * this takes a list head of mrs and turns it into linked llist nodes 312 * of clusters. Each cluster has linked llist nodes of 313 * MR_CLUSTER_SIZE mrs that are ready for reuse. 314 */ 315 static void list_to_llist_nodes(struct rds_ib_mr_pool *pool, 316 struct list_head *list, 317 struct llist_node **nodes_head, 318 struct llist_node **nodes_tail) 319 { 320 struct rds_ib_mr *ibmr; 321 struct llist_node *cur = NULL; 322 struct llist_node **next = nodes_head; 323 324 list_for_each_entry(ibmr, list, unmap_list) { 325 cur = &ibmr->llnode; 326 *next = cur; 327 next = &cur->next; 328 } 329 *next = NULL; 330 *nodes_tail = cur; 331 } 332 333 /* 334 * Flush our pool of MRs. 335 * At a minimum, all currently unused MRs are unmapped. 336 * If the number of MRs allocated exceeds the limit, we also try 337 * to free as many MRs as needed to get back to this limit. 338 */ 339 int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, 340 int free_all, struct rds_ib_mr **ibmr_ret) 341 { 342 struct rds_ib_mr *ibmr; 343 struct llist_node *clean_nodes; 344 struct llist_node *clean_tail; 345 LIST_HEAD(unmap_list); 346 unsigned long unpinned = 0; 347 unsigned int nfreed = 0, dirty_to_clean = 0, free_goal; 348 349 if (pool->pool_type == RDS_IB_MR_8K_POOL) 350 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush); 351 else 352 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush); 353 354 if (ibmr_ret) { 355 DEFINE_WAIT(wait); 356 while (!mutex_trylock(&pool->flush_lock)) { 357 ibmr = rds_ib_reuse_mr(pool); 358 if (ibmr) { 359 *ibmr_ret = ibmr; 360 finish_wait(&pool->flush_wait, &wait); 361 goto out_nolock; 362 } 363 364 prepare_to_wait(&pool->flush_wait, &wait, 365 TASK_UNINTERRUPTIBLE); 366 if (llist_empty(&pool->clean_list)) 367 schedule(); 368 369 ibmr = rds_ib_reuse_mr(pool); 370 if (ibmr) { 371 *ibmr_ret = ibmr; 372 finish_wait(&pool->flush_wait, &wait); 373 goto out_nolock; 374 } 375 } 376 finish_wait(&pool->flush_wait, &wait); 377 } else 378 mutex_lock(&pool->flush_lock); 379 380 if (ibmr_ret) { 381 ibmr = rds_ib_reuse_mr(pool); 382 if (ibmr) { 383 *ibmr_ret = ibmr; 384 goto out; 385 } 386 } 387 388 /* Get the list of all MRs to be dropped. Ordering matters - 389 * we want to put drop_list ahead of free_list. 390 */ 391 dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list); 392 dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list); 393 if (free_all) 394 llist_append_to_list(&pool->clean_list, &unmap_list); 395 396 free_goal = rds_ib_flush_goal(pool, free_all); 397 398 if (list_empty(&unmap_list)) 399 goto out; 400 401 if (pool->use_fastreg) 402 rds_ib_unreg_frmr(&unmap_list, &nfreed, &unpinned, free_goal); 403 else 404 rds_ib_unreg_fmr(&unmap_list, &nfreed, &unpinned, free_goal); 405 406 if (!list_empty(&unmap_list)) { 407 /* we have to make sure that none of the things we're about 408 * to put on the clean list would race with other cpus trying 409 * to pull items off. The llist would explode if we managed to 410 * remove something from the clean list and then add it back again 411 * while another CPU was spinning on that same item in llist_del_first. 412 * 413 * This is pretty unlikely, but just in case wait for an llist grace period 414 * here before adding anything back into the clean list. 415 */ 416 wait_clean_list_grace(); 417 418 list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail); 419 if (ibmr_ret) 420 *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode); 421 422 /* more than one entry in llist nodes */ 423 if (clean_nodes->next) 424 llist_add_batch(clean_nodes->next, clean_tail, &pool->clean_list); 425 426 } 427 428 atomic_sub(unpinned, &pool->free_pinned); 429 atomic_sub(dirty_to_clean, &pool->dirty_count); 430 atomic_sub(nfreed, &pool->item_count); 431 432 out: 433 mutex_unlock(&pool->flush_lock); 434 if (waitqueue_active(&pool->flush_wait)) 435 wake_up(&pool->flush_wait); 436 out_nolock: 437 return 0; 438 } 439 440 struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool) 441 { 442 struct rds_ib_mr *ibmr = NULL; 443 int iter = 0; 444 445 if (atomic_read(&pool->dirty_count) >= pool->max_items_soft / 10) 446 queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10); 447 448 while (1) { 449 ibmr = rds_ib_reuse_mr(pool); 450 if (ibmr) 451 return ibmr; 452 453 if (atomic_inc_return(&pool->item_count) <= pool->max_items) 454 break; 455 456 atomic_dec(&pool->item_count); 457 458 if (++iter > 2) { 459 if (pool->pool_type == RDS_IB_MR_8K_POOL) 460 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted); 461 else 462 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted); 463 return ERR_PTR(-EAGAIN); 464 } 465 466 /* We do have some empty MRs. Flush them out. */ 467 if (pool->pool_type == RDS_IB_MR_8K_POOL) 468 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait); 469 else 470 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait); 471 472 rds_ib_flush_mr_pool(pool, 0, &ibmr); 473 if (ibmr) 474 return ibmr; 475 } 476 477 return ibmr; 478 } 479 480 static void rds_ib_mr_pool_flush_worker(struct work_struct *work) 481 { 482 struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work); 483 484 rds_ib_flush_mr_pool(pool, 0, NULL); 485 } 486 487 void rds_ib_free_mr(void *trans_private, int invalidate) 488 { 489 struct rds_ib_mr *ibmr = trans_private; 490 struct rds_ib_mr_pool *pool = ibmr->pool; 491 struct rds_ib_device *rds_ibdev = ibmr->device; 492 493 rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len); 494 495 /* Return it to the pool's free list */ 496 if (rds_ibdev->use_fastreg) 497 rds_ib_free_frmr_list(ibmr); 498 else 499 rds_ib_free_fmr_list(ibmr); 500 501 atomic_add(ibmr->sg_len, &pool->free_pinned); 502 atomic_inc(&pool->dirty_count); 503 504 /* If we've pinned too many pages, request a flush */ 505 if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned || 506 atomic_read(&pool->dirty_count) >= pool->max_items / 5) 507 queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10); 508 509 if (invalidate) { 510 if (likely(!in_interrupt())) { 511 rds_ib_flush_mr_pool(pool, 0, NULL); 512 } else { 513 /* We get here if the user created a MR marked 514 * as use_once and invalidate at the same time. 515 */ 516 queue_delayed_work(rds_ib_mr_wq, 517 &pool->flush_worker, 10); 518 } 519 } 520 521 rds_ib_dev_put(rds_ibdev); 522 } 523 524 void rds_ib_flush_mrs(void) 525 { 526 struct rds_ib_device *rds_ibdev; 527 528 down_read(&rds_ib_devices_lock); 529 list_for_each_entry(rds_ibdev, &rds_ib_devices, list) { 530 if (rds_ibdev->mr_8k_pool) 531 rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL); 532 533 if (rds_ibdev->mr_1m_pool) 534 rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL); 535 } 536 up_read(&rds_ib_devices_lock); 537 } 538 539 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents, 540 struct rds_sock *rs, u32 *key_ret) 541 { 542 struct rds_ib_device *rds_ibdev; 543 struct rds_ib_mr *ibmr = NULL; 544 struct rds_ib_connection *ic = rs->rs_conn->c_transport_data; 545 int ret; 546 547 rds_ibdev = rds_ib_get_device(rs->rs_bound_addr); 548 if (!rds_ibdev) { 549 ret = -ENODEV; 550 goto out; 551 } 552 553 if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) { 554 ret = -ENODEV; 555 goto out; 556 } 557 558 if (rds_ibdev->use_fastreg) 559 ibmr = rds_ib_reg_frmr(rds_ibdev, ic, sg, nents, key_ret); 560 else 561 ibmr = rds_ib_reg_fmr(rds_ibdev, sg, nents, key_ret); 562 if (ibmr) 563 rds_ibdev = NULL; 564 565 out: 566 if (!ibmr) 567 pr_warn("RDS/IB: rds_ib_get_mr failed (errno=%d)\n", ret); 568 569 if (rds_ibdev) 570 rds_ib_dev_put(rds_ibdev); 571 572 return ibmr; 573 } 574 575 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool) 576 { 577 cancel_delayed_work_sync(&pool->flush_worker); 578 rds_ib_flush_mr_pool(pool, 1, NULL); 579 WARN_ON(atomic_read(&pool->item_count)); 580 WARN_ON(atomic_read(&pool->free_pinned)); 581 kfree(pool); 582 } 583 584 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev, 585 int pool_type) 586 { 587 struct rds_ib_mr_pool *pool; 588 589 pool = kzalloc(sizeof(*pool), GFP_KERNEL); 590 if (!pool) 591 return ERR_PTR(-ENOMEM); 592 593 pool->pool_type = pool_type; 594 init_llist_head(&pool->free_list); 595 init_llist_head(&pool->drop_list); 596 init_llist_head(&pool->clean_list); 597 mutex_init(&pool->flush_lock); 598 init_waitqueue_head(&pool->flush_wait); 599 INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker); 600 601 if (pool_type == RDS_IB_MR_1M_POOL) { 602 /* +1 allows for unaligned MRs */ 603 pool->fmr_attr.max_pages = RDS_MR_1M_MSG_SIZE + 1; 604 pool->max_items = RDS_MR_1M_POOL_SIZE; 605 } else { 606 /* pool_type == RDS_IB_MR_8K_POOL */ 607 pool->fmr_attr.max_pages = RDS_MR_8K_MSG_SIZE + 1; 608 pool->max_items = RDS_MR_8K_POOL_SIZE; 609 } 610 611 pool->max_free_pinned = pool->max_items * pool->fmr_attr.max_pages / 4; 612 pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps; 613 pool->fmr_attr.page_shift = PAGE_SHIFT; 614 pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4; 615 pool->use_fastreg = rds_ibdev->use_fastreg; 616 617 return pool; 618 } 619 620 int rds_ib_mr_init(void) 621 { 622 rds_ib_mr_wq = alloc_workqueue("rds_mr_flushd", WQ_MEM_RECLAIM, 0); 623 if (!rds_ib_mr_wq) 624 return -ENOMEM; 625 return 0; 626 } 627 628 /* By the time this is called all the IB devices should have been torn down and 629 * had their pools freed. As each pool is freed its work struct is waited on, 630 * so the pool flushing work queue should be idle by the time we get here. 631 */ 632 void rds_ib_mr_exit(void) 633 { 634 destroy_workqueue(rds_ib_mr_wq); 635 } 636