1/* 2 * linux/fs/nfs/pagelist.c 3 * 4 * A set of helper functions for managing NFS read and write requests. 5 * The main purpose of these routines is to provide support for the 6 * coalescing of several requests into a single RPC call. 7 * 8 * Copyright 2000, 2001 (c) Trond Myklebust <trond.myklebust@fys.uio.no> 9 * 10 */ 11 12#include <linux/slab.h> 13#include <linux/file.h> 14#include <linux/sched.h> 15#include <linux/sunrpc/clnt.h> 16#include <linux/nfs.h> 17#include <linux/nfs3.h> 18#include <linux/nfs4.h> 19#include <linux/nfs_page.h> 20#include <linux/nfs_fs.h> 21#include <linux/nfs_mount.h> 22#include <linux/export.h> 23 24#include "internal.h" 25#include "pnfs.h" 26 27#define NFSDBG_FACILITY NFSDBG_PAGECACHE 28 29static struct kmem_cache *nfs_page_cachep; 30static const struct rpc_call_ops nfs_pgio_common_ops; 31 32static bool nfs_pgarray_set(struct nfs_page_array *p, unsigned int pagecount) 33{ 34 p->npages = pagecount; 35 if (pagecount <= ARRAY_SIZE(p->page_array)) 36 p->pagevec = p->page_array; 37 else { 38 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_KERNEL); 39 if (!p->pagevec) 40 p->npages = 0; 41 } 42 return p->pagevec != NULL; 43} 44 45struct nfs_pgio_mirror * 46nfs_pgio_current_mirror(struct nfs_pageio_descriptor *desc) 47{ 48 return nfs_pgio_has_mirroring(desc) ? 49 &desc->pg_mirrors[desc->pg_mirror_idx] : 50 &desc->pg_mirrors[0]; 51} 52EXPORT_SYMBOL_GPL(nfs_pgio_current_mirror); 53 54void nfs_pgheader_init(struct nfs_pageio_descriptor *desc, 55 struct nfs_pgio_header *hdr, 56 void (*release)(struct nfs_pgio_header *hdr)) 57{ 58 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc); 59 60 61 hdr->req = nfs_list_entry(mirror->pg_list.next); 62 hdr->inode = desc->pg_inode; 63 hdr->cred = hdr->req->wb_context->cred; 64 hdr->io_start = req_offset(hdr->req); 65 hdr->good_bytes = mirror->pg_count; 66 hdr->dreq = desc->pg_dreq; 67 hdr->layout_private = desc->pg_layout_private; 68 hdr->release = release; 69 hdr->completion_ops = desc->pg_completion_ops; 70 if (hdr->completion_ops->init_hdr) 71 hdr->completion_ops->init_hdr(hdr); 72 73 hdr->pgio_mirror_idx = desc->pg_mirror_idx; 74} 75EXPORT_SYMBOL_GPL(nfs_pgheader_init); 76 77void nfs_set_pgio_error(struct nfs_pgio_header *hdr, int error, loff_t pos) 78{ 79 spin_lock(&hdr->lock); 80 if (!test_and_set_bit(NFS_IOHDR_ERROR, &hdr->flags) 81 || pos < hdr->io_start + hdr->good_bytes) { 82 clear_bit(NFS_IOHDR_EOF, &hdr->flags); 83 hdr->good_bytes = pos - hdr->io_start; 84 hdr->error = error; 85 } 86 spin_unlock(&hdr->lock); 87} 88 89static inline struct nfs_page * 90nfs_page_alloc(void) 91{ 92 struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_NOIO); 93 if (p) 94 INIT_LIST_HEAD(&p->wb_list); 95 return p; 96} 97 98static inline void 99nfs_page_free(struct nfs_page *p) 100{ 101 kmem_cache_free(nfs_page_cachep, p); 102} 103 104static void 105nfs_iocounter_inc(struct nfs_io_counter *c) 106{ 107 atomic_inc(&c->io_count); 108} 109 110static void 111nfs_iocounter_dec(struct nfs_io_counter *c) 112{ 113 if (atomic_dec_and_test(&c->io_count)) { 114 clear_bit(NFS_IO_INPROGRESS, &c->flags); 115 smp_mb__after_atomic(); 116 wake_up_bit(&c->flags, NFS_IO_INPROGRESS); 117 } 118} 119 120static int 121__nfs_iocounter_wait(struct nfs_io_counter *c) 122{ 123 wait_queue_head_t *wq = bit_waitqueue(&c->flags, NFS_IO_INPROGRESS); 124 DEFINE_WAIT_BIT(q, &c->flags, NFS_IO_INPROGRESS); 125 int ret = 0; 126 127 do { 128 prepare_to_wait(wq, &q.wait, TASK_KILLABLE); 129 set_bit(NFS_IO_INPROGRESS, &c->flags); 130 if (atomic_read(&c->io_count) == 0) 131 break; 132 ret = nfs_wait_bit_killable(&q.key); 133 } while (atomic_read(&c->io_count) != 0 && !ret); 134 finish_wait(wq, &q.wait); 135 return ret; 136} 137 138/** 139 * nfs_iocounter_wait - wait for i/o to complete 140 * @c: nfs_io_counter to use 141 * 142 * returns -ERESTARTSYS if interrupted by a fatal signal. 143 * Otherwise returns 0 once the io_count hits 0. 144 */ 145int 146nfs_iocounter_wait(struct nfs_io_counter *c) 147{ 148 if (atomic_read(&c->io_count) == 0) 149 return 0; 150 return __nfs_iocounter_wait(c); 151} 152 153/* 154 * nfs_page_group_lock - lock the head of the page group 155 * @req - request in group that is to be locked 156 * @nonblock - if true don't block waiting for lock 157 * 158 * this lock must be held if modifying the page group list 159 * 160 * return 0 on success, < 0 on error: -EDELAY if nonblocking or the 161 * result from wait_on_bit_lock 162 * 163 * NOTE: calling with nonblock=false should always have set the 164 * lock bit (see fs/buffer.c and other uses of wait_on_bit_lock 165 * with TASK_UNINTERRUPTIBLE), so there is no need to check the result. 166 */ 167int 168nfs_page_group_lock(struct nfs_page *req, bool nonblock) 169{ 170 struct nfs_page *head = req->wb_head; 171 172 WARN_ON_ONCE(head != head->wb_head); 173 174 if (!test_and_set_bit(PG_HEADLOCK, &head->wb_flags)) 175 return 0; 176 177 if (!nonblock) 178 return wait_on_bit_lock(&head->wb_flags, PG_HEADLOCK, 179 TASK_UNINTERRUPTIBLE); 180 181 return -EAGAIN; 182} 183 184/* 185 * nfs_page_group_lock_wait - wait for the lock to clear, but don't grab it 186 * @req - a request in the group 187 * 188 * This is a blocking call to wait for the group lock to be cleared. 189 */ 190void 191nfs_page_group_lock_wait(struct nfs_page *req) 192{ 193 struct nfs_page *head = req->wb_head; 194 195 WARN_ON_ONCE(head != head->wb_head); 196 197 wait_on_bit(&head->wb_flags, PG_HEADLOCK, 198 TASK_UNINTERRUPTIBLE); 199} 200 201/* 202 * nfs_page_group_unlock - unlock the head of the page group 203 * @req - request in group that is to be unlocked 204 */ 205void 206nfs_page_group_unlock(struct nfs_page *req) 207{ 208 struct nfs_page *head = req->wb_head; 209 210 WARN_ON_ONCE(head != head->wb_head); 211 212 smp_mb__before_atomic(); 213 clear_bit(PG_HEADLOCK, &head->wb_flags); 214 smp_mb__after_atomic(); 215 wake_up_bit(&head->wb_flags, PG_HEADLOCK); 216} 217 218/* 219 * nfs_page_group_sync_on_bit_locked 220 * 221 * must be called with page group lock held 222 */ 223static bool 224nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit) 225{ 226 struct nfs_page *head = req->wb_head; 227 struct nfs_page *tmp; 228 229 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_flags)); 230 WARN_ON_ONCE(test_and_set_bit(bit, &req->wb_flags)); 231 232 tmp = req->wb_this_page; 233 while (tmp != req) { 234 if (!test_bit(bit, &tmp->wb_flags)) 235 return false; 236 tmp = tmp->wb_this_page; 237 } 238 239 /* true! reset all bits */ 240 tmp = req; 241 do { 242 clear_bit(bit, &tmp->wb_flags); 243 tmp = tmp->wb_this_page; 244 } while (tmp != req); 245 246 return true; 247} 248 249/* 250 * nfs_page_group_sync_on_bit - set bit on current request, but only 251 * return true if the bit is set for all requests in page group 252 * @req - request in page group 253 * @bit - PG_* bit that is used to sync page group 254 */ 255bool nfs_page_group_sync_on_bit(struct nfs_page *req, unsigned int bit) 256{ 257 bool ret; 258 259 nfs_page_group_lock(req, false); 260 ret = nfs_page_group_sync_on_bit_locked(req, bit); 261 nfs_page_group_unlock(req); 262 263 return ret; 264} 265 266/* 267 * nfs_page_group_init - Initialize the page group linkage for @req 268 * @req - a new nfs request 269 * @prev - the previous request in page group, or NULL if @req is the first 270 * or only request in the group (the head). 271 */ 272static inline void 273nfs_page_group_init(struct nfs_page *req, struct nfs_page *prev) 274{ 275 struct inode *inode; 276 WARN_ON_ONCE(prev == req); 277 278 if (!prev) { 279 /* a head request */ 280 req->wb_head = req; 281 req->wb_this_page = req; 282 } else { 283 /* a subrequest */ 284 WARN_ON_ONCE(prev->wb_this_page != prev->wb_head); 285 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &prev->wb_head->wb_flags)); 286 req->wb_head = prev->wb_head; 287 req->wb_this_page = prev->wb_this_page; 288 prev->wb_this_page = req; 289 290 /* All subrequests take a ref on the head request until 291 * nfs_page_group_destroy is called */ 292 kref_get(&req->wb_head->wb_kref); 293 294 /* grab extra ref and bump the request count if head request 295 * has extra ref from the write/commit path to handle handoff 296 * between write and commit lists. */ 297 if (test_bit(PG_INODE_REF, &prev->wb_head->wb_flags)) { 298 inode = page_file_mapping(req->wb_page)->host; 299 set_bit(PG_INODE_REF, &req->wb_flags); 300 kref_get(&req->wb_kref); 301 spin_lock(&inode->i_lock); 302 NFS_I(inode)->nrequests++; 303 spin_unlock(&inode->i_lock); 304 } 305 } 306} 307 308/* 309 * nfs_page_group_destroy - sync the destruction of page groups 310 * @req - request that no longer needs the page group 311 * 312 * releases the page group reference from each member once all 313 * members have called this function. 314 */ 315static void 316nfs_page_group_destroy(struct kref *kref) 317{ 318 struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref); 319 struct nfs_page *tmp, *next; 320 321 /* subrequests must release the ref on the head request */ 322 if (req->wb_head != req) 323 nfs_release_request(req->wb_head); 324 325 if (!nfs_page_group_sync_on_bit(req, PG_TEARDOWN)) 326 return; 327 328 tmp = req; 329 do { 330 next = tmp->wb_this_page; 331 /* unlink and free */ 332 tmp->wb_this_page = tmp; 333 tmp->wb_head = tmp; 334 nfs_free_request(tmp); 335 tmp = next; 336 } while (tmp != req); 337} 338 339/** 340 * nfs_create_request - Create an NFS read/write request. 341 * @ctx: open context to use 342 * @page: page to write 343 * @last: last nfs request created for this page group or NULL if head 344 * @offset: starting offset within the page for the write 345 * @count: number of bytes to read/write 346 * 347 * The page must be locked by the caller. This makes sure we never 348 * create two different requests for the same page. 349 * User should ensure it is safe to sleep in this function. 350 */ 351struct nfs_page * 352nfs_create_request(struct nfs_open_context *ctx, struct page *page, 353 struct nfs_page *last, unsigned int offset, 354 unsigned int count) 355{ 356 struct nfs_page *req; 357 struct nfs_lock_context *l_ctx; 358 359 if (test_bit(NFS_CONTEXT_BAD, &ctx->flags)) 360 return ERR_PTR(-EBADF); 361 /* try to allocate the request struct */ 362 req = nfs_page_alloc(); 363 if (req == NULL) 364 return ERR_PTR(-ENOMEM); 365 366 /* get lock context early so we can deal with alloc failures */ 367 l_ctx = nfs_get_lock_context(ctx); 368 if (IS_ERR(l_ctx)) { 369 nfs_page_free(req); 370 return ERR_CAST(l_ctx); 371 } 372 req->wb_lock_context = l_ctx; 373 nfs_iocounter_inc(&l_ctx->io_count); 374 375 /* Initialize the request struct. Initially, we assume a 376 * long write-back delay. This will be adjusted in 377 * update_nfs_request below if the region is not locked. */ 378 req->wb_page = page; 379 req->wb_index = page_file_index(page); 380 page_cache_get(page); 381 req->wb_offset = offset; 382 req->wb_pgbase = offset; 383 req->wb_bytes = count; 384 req->wb_context = get_nfs_open_context(ctx); 385 kref_init(&req->wb_kref); 386 nfs_page_group_init(req, last); 387 return req; 388} 389 390/** 391 * nfs_unlock_request - Unlock request and wake up sleepers. 392 * @req: 393 */ 394void nfs_unlock_request(struct nfs_page *req) 395{ 396 if (!NFS_WBACK_BUSY(req)) { 397 printk(KERN_ERR "NFS: Invalid unlock attempted\n"); 398 BUG(); 399 } 400 smp_mb__before_atomic(); 401 clear_bit(PG_BUSY, &req->wb_flags); 402 smp_mb__after_atomic(); 403 wake_up_bit(&req->wb_flags, PG_BUSY); 404} 405 406/** 407 * nfs_unlock_and_release_request - Unlock request and release the nfs_page 408 * @req: 409 */ 410void nfs_unlock_and_release_request(struct nfs_page *req) 411{ 412 nfs_unlock_request(req); 413 nfs_release_request(req); 414} 415 416/* 417 * nfs_clear_request - Free up all resources allocated to the request 418 * @req: 419 * 420 * Release page and open context resources associated with a read/write 421 * request after it has completed. 422 */ 423static void nfs_clear_request(struct nfs_page *req) 424{ 425 struct page *page = req->wb_page; 426 struct nfs_open_context *ctx = req->wb_context; 427 struct nfs_lock_context *l_ctx = req->wb_lock_context; 428 429 if (page != NULL) { 430 page_cache_release(page); 431 req->wb_page = NULL; 432 } 433 if (l_ctx != NULL) { 434 nfs_iocounter_dec(&l_ctx->io_count); 435 nfs_put_lock_context(l_ctx); 436 req->wb_lock_context = NULL; 437 } 438 if (ctx != NULL) { 439 put_nfs_open_context(ctx); 440 req->wb_context = NULL; 441 } 442} 443 444/** 445 * nfs_release_request - Release the count on an NFS read/write request 446 * @req: request to release 447 * 448 * Note: Should never be called with the spinlock held! 449 */ 450void nfs_free_request(struct nfs_page *req) 451{ 452 WARN_ON_ONCE(req->wb_this_page != req); 453 454 /* extra debug: make sure no sync bits are still set */ 455 WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags)); 456 WARN_ON_ONCE(test_bit(PG_UNLOCKPAGE, &req->wb_flags)); 457 WARN_ON_ONCE(test_bit(PG_UPTODATE, &req->wb_flags)); 458 WARN_ON_ONCE(test_bit(PG_WB_END, &req->wb_flags)); 459 WARN_ON_ONCE(test_bit(PG_REMOVE, &req->wb_flags)); 460 461 /* Release struct file and open context */ 462 nfs_clear_request(req); 463 nfs_page_free(req); 464} 465 466void nfs_release_request(struct nfs_page *req) 467{ 468 kref_put(&req->wb_kref, nfs_page_group_destroy); 469} 470 471/** 472 * nfs_wait_on_request - Wait for a request to complete. 473 * @req: request to wait upon. 474 * 475 * Interruptible by fatal signals only. 476 * The user is responsible for holding a count on the request. 477 */ 478int 479nfs_wait_on_request(struct nfs_page *req) 480{ 481 return wait_on_bit_io(&req->wb_flags, PG_BUSY, 482 TASK_UNINTERRUPTIBLE); 483} 484 485/* 486 * nfs_generic_pg_test - determine if requests can be coalesced 487 * @desc: pointer to descriptor 488 * @prev: previous request in desc, or NULL 489 * @req: this request 490 * 491 * Returns zero if @req can be coalesced into @desc, otherwise it returns 492 * the size of the request. 493 */ 494size_t nfs_generic_pg_test(struct nfs_pageio_descriptor *desc, 495 struct nfs_page *prev, struct nfs_page *req) 496{ 497 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc); 498 499 500 if (mirror->pg_count > mirror->pg_bsize) { 501 /* should never happen */ 502 WARN_ON_ONCE(1); 503 return 0; 504 } 505 506 /* 507 * Limit the request size so that we can still allocate a page array 508 * for it without upsetting the slab allocator. 509 */ 510 if (((mirror->pg_count + req->wb_bytes) >> PAGE_SHIFT) * 511 sizeof(struct page *) > PAGE_SIZE) 512 return 0; 513 514 return min(mirror->pg_bsize - mirror->pg_count, (size_t)req->wb_bytes); 515} 516EXPORT_SYMBOL_GPL(nfs_generic_pg_test); 517 518struct nfs_pgio_header *nfs_pgio_header_alloc(const struct nfs_rw_ops *ops) 519{ 520 struct nfs_pgio_header *hdr = ops->rw_alloc_header(); 521 522 if (hdr) { 523 INIT_LIST_HEAD(&hdr->pages); 524 spin_lock_init(&hdr->lock); 525 hdr->rw_ops = ops; 526 } 527 return hdr; 528} 529EXPORT_SYMBOL_GPL(nfs_pgio_header_alloc); 530 531/* 532 * nfs_pgio_header_free - Free a read or write header 533 * @hdr: The header to free 534 */ 535void nfs_pgio_header_free(struct nfs_pgio_header *hdr) 536{ 537 hdr->rw_ops->rw_free_header(hdr); 538} 539EXPORT_SYMBOL_GPL(nfs_pgio_header_free); 540 541/** 542 * nfs_pgio_data_destroy - make @hdr suitable for reuse 543 * 544 * Frees memory and releases refs from nfs_generic_pgio, so that it may 545 * be called again. 546 * 547 * @hdr: A header that has had nfs_generic_pgio called 548 */ 549void nfs_pgio_data_destroy(struct nfs_pgio_header *hdr) 550{ 551 if (hdr->args.context) 552 put_nfs_open_context(hdr->args.context); 553 if (hdr->page_array.pagevec != hdr->page_array.page_array) 554 kfree(hdr->page_array.pagevec); 555} 556EXPORT_SYMBOL_GPL(nfs_pgio_data_destroy); 557 558/** 559 * nfs_pgio_rpcsetup - Set up arguments for a pageio call 560 * @hdr: The pageio hdr 561 * @count: Number of bytes to read 562 * @offset: Initial offset 563 * @how: How to commit data (writes only) 564 * @cinfo: Commit information for the call (writes only) 565 */ 566static void nfs_pgio_rpcsetup(struct nfs_pgio_header *hdr, 567 unsigned int count, unsigned int offset, 568 int how, struct nfs_commit_info *cinfo) 569{ 570 struct nfs_page *req = hdr->req; 571 572 /* Set up the RPC argument and reply structs 573 * NB: take care not to mess about with hdr->commit et al. */ 574 575 hdr->args.fh = NFS_FH(hdr->inode); 576 hdr->args.offset = req_offset(req) + offset; 577 /* pnfs_set_layoutcommit needs this */ 578 hdr->mds_offset = hdr->args.offset; 579 hdr->args.pgbase = req->wb_pgbase + offset; 580 hdr->args.pages = hdr->page_array.pagevec; 581 hdr->args.count = count; 582 hdr->args.context = get_nfs_open_context(req->wb_context); 583 hdr->args.lock_context = req->wb_lock_context; 584 hdr->args.stable = NFS_UNSTABLE; 585 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) { 586 case 0: 587 break; 588 case FLUSH_COND_STABLE: 589 if (nfs_reqs_to_commit(cinfo)) 590 break; 591 default: 592 hdr->args.stable = NFS_FILE_SYNC; 593 } 594 595 hdr->res.fattr = &hdr->fattr; 596 hdr->res.count = count; 597 hdr->res.eof = 0; 598 hdr->res.verf = &hdr->verf; 599 nfs_fattr_init(&hdr->fattr); 600} 601 602/** 603 * nfs_pgio_prepare - Prepare pageio hdr to go over the wire 604 * @task: The current task 605 * @calldata: pageio header to prepare 606 */ 607static void nfs_pgio_prepare(struct rpc_task *task, void *calldata) 608{ 609 struct nfs_pgio_header *hdr = calldata; 610 int err; 611 err = NFS_PROTO(hdr->inode)->pgio_rpc_prepare(task, hdr); 612 if (err) 613 rpc_exit(task, err); 614} 615 616int nfs_initiate_pgio(struct rpc_clnt *clnt, struct nfs_pgio_header *hdr, 617 struct rpc_cred *cred, const struct nfs_rpc_ops *rpc_ops, 618 const struct rpc_call_ops *call_ops, int how, int flags) 619{ 620 struct rpc_task *task; 621 struct rpc_message msg = { 622 .rpc_argp = &hdr->args, 623 .rpc_resp = &hdr->res, 624 .rpc_cred = cred, 625 }; 626 struct rpc_task_setup task_setup_data = { 627 .rpc_client = clnt, 628 .task = &hdr->task, 629 .rpc_message = &msg, 630 .callback_ops = call_ops, 631 .callback_data = hdr, 632 .workqueue = nfsiod_workqueue, 633 .flags = RPC_TASK_ASYNC | flags, 634 }; 635 int ret = 0; 636 637 hdr->rw_ops->rw_initiate(hdr, &msg, rpc_ops, &task_setup_data, how); 638 639 dprintk("NFS: %5u initiated pgio call " 640 "(req %s/%llu, %u bytes @ offset %llu)\n", 641 hdr->task.tk_pid, 642 hdr->inode->i_sb->s_id, 643 (unsigned long long)NFS_FILEID(hdr->inode), 644 hdr->args.count, 645 (unsigned long long)hdr->args.offset); 646 647 task = rpc_run_task(&task_setup_data); 648 if (IS_ERR(task)) { 649 ret = PTR_ERR(task); 650 goto out; 651 } 652 if (how & FLUSH_SYNC) { 653 ret = rpc_wait_for_completion_task(task); 654 if (ret == 0) 655 ret = task->tk_status; 656 } 657 rpc_put_task(task); 658out: 659 return ret; 660} 661EXPORT_SYMBOL_GPL(nfs_initiate_pgio); 662 663/** 664 * nfs_pgio_error - Clean up from a pageio error 665 * @desc: IO descriptor 666 * @hdr: pageio header 667 */ 668static int nfs_pgio_error(struct nfs_pageio_descriptor *desc, 669 struct nfs_pgio_header *hdr) 670{ 671 struct nfs_pgio_mirror *mirror; 672 u32 midx; 673 674 set_bit(NFS_IOHDR_REDO, &hdr->flags); 675 nfs_pgio_data_destroy(hdr); 676 hdr->completion_ops->completion(hdr); 677 /* TODO: Make sure it's right to clean up all mirrors here 678 * and not just hdr->pgio_mirror_idx */ 679 for (midx = 0; midx < desc->pg_mirror_count; midx++) { 680 mirror = &desc->pg_mirrors[midx]; 681 desc->pg_completion_ops->error_cleanup(&mirror->pg_list); 682 } 683 return -ENOMEM; 684} 685 686/** 687 * nfs_pgio_release - Release pageio data 688 * @calldata: The pageio header to release 689 */ 690static void nfs_pgio_release(void *calldata) 691{ 692 struct nfs_pgio_header *hdr = calldata; 693 if (hdr->rw_ops->rw_release) 694 hdr->rw_ops->rw_release(hdr); 695 nfs_pgio_data_destroy(hdr); 696 hdr->completion_ops->completion(hdr); 697} 698 699static void nfs_pageio_mirror_init(struct nfs_pgio_mirror *mirror, 700 unsigned int bsize) 701{ 702 INIT_LIST_HEAD(&mirror->pg_list); 703 mirror->pg_bytes_written = 0; 704 mirror->pg_count = 0; 705 mirror->pg_bsize = bsize; 706 mirror->pg_base = 0; 707 mirror->pg_recoalesce = 0; 708} 709 710/** 711 * nfs_pageio_init - initialise a page io descriptor 712 * @desc: pointer to descriptor 713 * @inode: pointer to inode 714 * @doio: pointer to io function 715 * @bsize: io block size 716 * @io_flags: extra parameters for the io function 717 */ 718void nfs_pageio_init(struct nfs_pageio_descriptor *desc, 719 struct inode *inode, 720 const struct nfs_pageio_ops *pg_ops, 721 const struct nfs_pgio_completion_ops *compl_ops, 722 const struct nfs_rw_ops *rw_ops, 723 size_t bsize, 724 int io_flags) 725{ 726 struct nfs_pgio_mirror *new; 727 int i; 728 729 desc->pg_moreio = 0; 730 desc->pg_inode = inode; 731 desc->pg_ops = pg_ops; 732 desc->pg_completion_ops = compl_ops; 733 desc->pg_rw_ops = rw_ops; 734 desc->pg_ioflags = io_flags; 735 desc->pg_error = 0; 736 desc->pg_lseg = NULL; 737 desc->pg_dreq = NULL; 738 desc->pg_layout_private = NULL; 739 desc->pg_bsize = bsize; 740 741 desc->pg_mirror_count = 1; 742 desc->pg_mirror_idx = 0; 743 744 if (pg_ops->pg_get_mirror_count) { 745 /* until we have a request, we don't have an lseg and no 746 * idea how many mirrors there will be */ 747 new = kcalloc(NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX, 748 sizeof(struct nfs_pgio_mirror), GFP_KERNEL); 749 desc->pg_mirrors_dynamic = new; 750 desc->pg_mirrors = new; 751 752 for (i = 0; i < NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX; i++) 753 nfs_pageio_mirror_init(&desc->pg_mirrors[i], bsize); 754 } else { 755 desc->pg_mirrors_dynamic = NULL; 756 desc->pg_mirrors = desc->pg_mirrors_static; 757 nfs_pageio_mirror_init(&desc->pg_mirrors[0], bsize); 758 } 759} 760EXPORT_SYMBOL_GPL(nfs_pageio_init); 761 762/** 763 * nfs_pgio_result - Basic pageio error handling 764 * @task: The task that ran 765 * @calldata: Pageio header to check 766 */ 767static void nfs_pgio_result(struct rpc_task *task, void *calldata) 768{ 769 struct nfs_pgio_header *hdr = calldata; 770 struct inode *inode = hdr->inode; 771 772 dprintk("NFS: %s: %5u, (status %d)\n", __func__, 773 task->tk_pid, task->tk_status); 774 775 if (hdr->rw_ops->rw_done(task, hdr, inode) != 0) 776 return; 777 if (task->tk_status < 0) 778 nfs_set_pgio_error(hdr, task->tk_status, hdr->args.offset); 779 else 780 hdr->rw_ops->rw_result(task, hdr); 781} 782 783/* 784 * Create an RPC task for the given read or write request and kick it. 785 * The page must have been locked by the caller. 786 * 787 * It may happen that the page we're passed is not marked dirty. 788 * This is the case if nfs_updatepage detects a conflicting request 789 * that has been written but not committed. 790 */ 791int nfs_generic_pgio(struct nfs_pageio_descriptor *desc, 792 struct nfs_pgio_header *hdr) 793{ 794 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc); 795 796 struct nfs_page *req; 797 struct page **pages, 798 *last_page; 799 struct list_head *head = &mirror->pg_list; 800 struct nfs_commit_info cinfo; 801 unsigned int pagecount, pageused; 802 803 pagecount = nfs_page_array_len(mirror->pg_base, mirror->pg_count); 804 if (!nfs_pgarray_set(&hdr->page_array, pagecount)) 805 return nfs_pgio_error(desc, hdr); 806 807 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq); 808 pages = hdr->page_array.pagevec; 809 last_page = NULL; 810 pageused = 0; 811 while (!list_empty(head)) { 812 req = nfs_list_entry(head->next); 813 nfs_list_remove_request(req); 814 nfs_list_add_request(req, &hdr->pages); 815 816 if (!last_page || last_page != req->wb_page) { 817 pageused++; 818 if (pageused > pagecount) 819 break; 820 *pages++ = last_page = req->wb_page; 821 } 822 } 823 if (WARN_ON_ONCE(pageused != pagecount)) 824 return nfs_pgio_error(desc, hdr); 825 826 if ((desc->pg_ioflags & FLUSH_COND_STABLE) && 827 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo))) 828 desc->pg_ioflags &= ~FLUSH_COND_STABLE; 829 830 /* Set up the argument struct */ 831 nfs_pgio_rpcsetup(hdr, mirror->pg_count, 0, desc->pg_ioflags, &cinfo); 832 desc->pg_rpc_callops = &nfs_pgio_common_ops; 833 return 0; 834} 835EXPORT_SYMBOL_GPL(nfs_generic_pgio); 836 837static int nfs_generic_pg_pgios(struct nfs_pageio_descriptor *desc) 838{ 839 struct nfs_pgio_mirror *mirror; 840 struct nfs_pgio_header *hdr; 841 int ret; 842 843 mirror = nfs_pgio_current_mirror(desc); 844 845 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops); 846 if (!hdr) { 847 /* TODO: make sure this is right with mirroring - or 848 * should it back out all mirrors? */ 849 desc->pg_completion_ops->error_cleanup(&mirror->pg_list); 850 return -ENOMEM; 851 } 852 nfs_pgheader_init(desc, hdr, nfs_pgio_header_free); 853 ret = nfs_generic_pgio(desc, hdr); 854 if (ret == 0) 855 ret = nfs_initiate_pgio(NFS_CLIENT(hdr->inode), 856 hdr, 857 hdr->cred, 858 NFS_PROTO(hdr->inode), 859 desc->pg_rpc_callops, 860 desc->pg_ioflags, 0); 861 return ret; 862} 863 864/* 865 * nfs_pageio_setup_mirroring - determine if mirroring is to be used 866 * by calling the pg_get_mirror_count op 867 */ 868static int nfs_pageio_setup_mirroring(struct nfs_pageio_descriptor *pgio, 869 struct nfs_page *req) 870{ 871 int mirror_count = 1; 872 873 if (!pgio->pg_ops->pg_get_mirror_count) 874 return 0; 875 876 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req); 877 878 if (!mirror_count || mirror_count > NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX) 879 return -EINVAL; 880 881 if (WARN_ON_ONCE(!pgio->pg_mirrors_dynamic)) 882 return -EINVAL; 883 884 pgio->pg_mirror_count = mirror_count; 885 886 return 0; 887} 888 889/* 890 * nfs_pageio_stop_mirroring - stop using mirroring (set mirror count to 1) 891 */ 892void nfs_pageio_stop_mirroring(struct nfs_pageio_descriptor *pgio) 893{ 894 pgio->pg_mirror_count = 1; 895 pgio->pg_mirror_idx = 0; 896} 897 898static void nfs_pageio_cleanup_mirroring(struct nfs_pageio_descriptor *pgio) 899{ 900 pgio->pg_mirror_count = 1; 901 pgio->pg_mirror_idx = 0; 902 pgio->pg_mirrors = pgio->pg_mirrors_static; 903 kfree(pgio->pg_mirrors_dynamic); 904 pgio->pg_mirrors_dynamic = NULL; 905} 906 907static bool nfs_match_open_context(const struct nfs_open_context *ctx1, 908 const struct nfs_open_context *ctx2) 909{ 910 return ctx1->cred == ctx2->cred && ctx1->state == ctx2->state; 911} 912 913static bool nfs_match_lock_context(const struct nfs_lock_context *l1, 914 const struct nfs_lock_context *l2) 915{ 916 return l1->lockowner.l_owner == l2->lockowner.l_owner 917 && l1->lockowner.l_pid == l2->lockowner.l_pid; 918} 919 920/** 921 * nfs_can_coalesce_requests - test two requests for compatibility 922 * @prev: pointer to nfs_page 923 * @req: pointer to nfs_page 924 * 925 * The nfs_page structures 'prev' and 'req' are compared to ensure that the 926 * page data area they describe is contiguous, and that their RPC 927 * credentials, NFSv4 open state, and lockowners are the same. 928 * 929 * Return 'true' if this is the case, else return 'false'. 930 */ 931static bool nfs_can_coalesce_requests(struct nfs_page *prev, 932 struct nfs_page *req, 933 struct nfs_pageio_descriptor *pgio) 934{ 935 size_t size; 936 struct file_lock_context *flctx; 937 938 if (prev) { 939 if (!nfs_match_open_context(req->wb_context, prev->wb_context)) 940 return false; 941 flctx = d_inode(req->wb_context->dentry)->i_flctx; 942 if (flctx != NULL && 943 !(list_empty_careful(&flctx->flc_posix) && 944 list_empty_careful(&flctx->flc_flock)) && 945 !nfs_match_lock_context(req->wb_lock_context, 946 prev->wb_lock_context)) 947 return false; 948 if (req_offset(req) != req_offset(prev) + prev->wb_bytes) 949 return false; 950 if (req->wb_page == prev->wb_page) { 951 if (req->wb_pgbase != prev->wb_pgbase + prev->wb_bytes) 952 return false; 953 } else { 954 if (req->wb_pgbase != 0 || 955 prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE) 956 return false; 957 } 958 } 959 size = pgio->pg_ops->pg_test(pgio, prev, req); 960 WARN_ON_ONCE(size > req->wb_bytes); 961 if (size && size < req->wb_bytes) 962 req->wb_bytes = size; 963 return size > 0; 964} 965 966/** 967 * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list. 968 * @desc: destination io descriptor 969 * @req: request 970 * 971 * Returns true if the request 'req' was successfully coalesced into the 972 * existing list of pages 'desc'. 973 */ 974static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc, 975 struct nfs_page *req) 976{ 977 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc); 978 979 struct nfs_page *prev = NULL; 980 981 if (mirror->pg_count != 0) { 982 prev = nfs_list_entry(mirror->pg_list.prev); 983 } else { 984 if (desc->pg_ops->pg_init) 985 desc->pg_ops->pg_init(desc, req); 986 mirror->pg_base = req->wb_pgbase; 987 } 988 if (!nfs_can_coalesce_requests(prev, req, desc)) 989 return 0; 990 nfs_list_remove_request(req); 991 nfs_list_add_request(req, &mirror->pg_list); 992 mirror->pg_count += req->wb_bytes; 993 return 1; 994} 995 996/* 997 * Helper for nfs_pageio_add_request and nfs_pageio_complete 998 */ 999static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc) 1000{ 1001 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc); 1002 1003 1004 if (!list_empty(&mirror->pg_list)) { 1005 int error = desc->pg_ops->pg_doio(desc); 1006 if (error < 0) 1007 desc->pg_error = error; 1008 else 1009 mirror->pg_bytes_written += mirror->pg_count; 1010 } 1011 if (list_empty(&mirror->pg_list)) { 1012 mirror->pg_count = 0; 1013 mirror->pg_base = 0; 1014 } 1015} 1016 1017/** 1018 * nfs_pageio_add_request - Attempt to coalesce a request into a page list. 1019 * @desc: destination io descriptor 1020 * @req: request 1021 * 1022 * This may split a request into subrequests which are all part of the 1023 * same page group. 1024 * 1025 * Returns true if the request 'req' was successfully coalesced into the 1026 * existing list of pages 'desc'. 1027 */ 1028static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc, 1029 struct nfs_page *req) 1030{ 1031 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc); 1032 1033 struct nfs_page *subreq; 1034 unsigned int bytes_left = 0; 1035 unsigned int offset, pgbase; 1036 1037 nfs_page_group_lock(req, false); 1038 1039 subreq = req; 1040 bytes_left = subreq->wb_bytes; 1041 offset = subreq->wb_offset; 1042 pgbase = subreq->wb_pgbase; 1043 1044 do { 1045 if (!nfs_pageio_do_add_request(desc, subreq)) { 1046 /* make sure pg_test call(s) did nothing */ 1047 WARN_ON_ONCE(subreq->wb_bytes != bytes_left); 1048 WARN_ON_ONCE(subreq->wb_offset != offset); 1049 WARN_ON_ONCE(subreq->wb_pgbase != pgbase); 1050 1051 nfs_page_group_unlock(req); 1052 desc->pg_moreio = 1; 1053 nfs_pageio_doio(desc); 1054 if (desc->pg_error < 0) 1055 return 0; 1056 if (mirror->pg_recoalesce) 1057 return 0; 1058 /* retry add_request for this subreq */ 1059 nfs_page_group_lock(req, false); 1060 continue; 1061 } 1062 1063 /* check for buggy pg_test call(s) */ 1064 WARN_ON_ONCE(subreq->wb_bytes + subreq->wb_pgbase > PAGE_SIZE); 1065 WARN_ON_ONCE(subreq->wb_bytes > bytes_left); 1066 WARN_ON_ONCE(subreq->wb_bytes == 0); 1067 1068 bytes_left -= subreq->wb_bytes; 1069 offset += subreq->wb_bytes; 1070 pgbase += subreq->wb_bytes; 1071 1072 if (bytes_left) { 1073 subreq = nfs_create_request(req->wb_context, 1074 req->wb_page, 1075 subreq, pgbase, bytes_left); 1076 if (IS_ERR(subreq)) 1077 goto err_ptr; 1078 nfs_lock_request(subreq); 1079 subreq->wb_offset = offset; 1080 subreq->wb_index = req->wb_index; 1081 } 1082 } while (bytes_left > 0); 1083 1084 nfs_page_group_unlock(req); 1085 return 1; 1086err_ptr: 1087 desc->pg_error = PTR_ERR(subreq); 1088 nfs_page_group_unlock(req); 1089 return 0; 1090} 1091 1092static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc) 1093{ 1094 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc); 1095 LIST_HEAD(head); 1096 1097 do { 1098 list_splice_init(&mirror->pg_list, &head); 1099 mirror->pg_bytes_written -= mirror->pg_count; 1100 mirror->pg_count = 0; 1101 mirror->pg_base = 0; 1102 mirror->pg_recoalesce = 0; 1103 1104 desc->pg_moreio = 0; 1105 1106 while (!list_empty(&head)) { 1107 struct nfs_page *req; 1108 1109 req = list_first_entry(&head, struct nfs_page, wb_list); 1110 nfs_list_remove_request(req); 1111 if (__nfs_pageio_add_request(desc, req)) 1112 continue; 1113 if (desc->pg_error < 0) { 1114 list_splice_tail(&head, &mirror->pg_list); 1115 mirror->pg_recoalesce = 1; 1116 return 0; 1117 } 1118 break; 1119 } 1120 } while (mirror->pg_recoalesce); 1121 return 1; 1122} 1123 1124static int nfs_pageio_add_request_mirror(struct nfs_pageio_descriptor *desc, 1125 struct nfs_page *req) 1126{ 1127 int ret; 1128 1129 do { 1130 ret = __nfs_pageio_add_request(desc, req); 1131 if (ret) 1132 break; 1133 if (desc->pg_error < 0) 1134 break; 1135 ret = nfs_do_recoalesce(desc); 1136 } while (ret); 1137 1138 return ret; 1139} 1140 1141int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc, 1142 struct nfs_page *req) 1143{ 1144 u32 midx; 1145 unsigned int pgbase, offset, bytes; 1146 struct nfs_page *dupreq, *lastreq; 1147 1148 pgbase = req->wb_pgbase; 1149 offset = req->wb_offset; 1150 bytes = req->wb_bytes; 1151 1152 nfs_pageio_setup_mirroring(desc, req); 1153 1154 for (midx = 0; midx < desc->pg_mirror_count; midx++) { 1155 if (midx) { 1156 nfs_page_group_lock(req, false); 1157 1158 /* find the last request */ 1159 for (lastreq = req->wb_head; 1160 lastreq->wb_this_page != req->wb_head; 1161 lastreq = lastreq->wb_this_page) 1162 ; 1163 1164 dupreq = nfs_create_request(req->wb_context, 1165 req->wb_page, lastreq, pgbase, bytes); 1166 1167 if (IS_ERR(dupreq)) { 1168 nfs_page_group_unlock(req); 1169 return 0; 1170 } 1171 1172 nfs_lock_request(dupreq); 1173 nfs_page_group_unlock(req); 1174 dupreq->wb_offset = offset; 1175 dupreq->wb_index = req->wb_index; 1176 } else 1177 dupreq = req; 1178 1179 if (nfs_pgio_has_mirroring(desc)) 1180 desc->pg_mirror_idx = midx; 1181 if (!nfs_pageio_add_request_mirror(desc, dupreq)) 1182 return 0; 1183 } 1184 1185 return 1; 1186} 1187 1188/* 1189 * nfs_pageio_complete_mirror - Complete I/O on the current mirror of an 1190 * nfs_pageio_descriptor 1191 * @desc: pointer to io descriptor 1192 */ 1193static void nfs_pageio_complete_mirror(struct nfs_pageio_descriptor *desc, 1194 u32 mirror_idx) 1195{ 1196 struct nfs_pgio_mirror *mirror = &desc->pg_mirrors[mirror_idx]; 1197 u32 restore_idx = desc->pg_mirror_idx; 1198 1199 if (nfs_pgio_has_mirroring(desc)) 1200 desc->pg_mirror_idx = mirror_idx; 1201 for (;;) { 1202 nfs_pageio_doio(desc); 1203 if (!mirror->pg_recoalesce) 1204 break; 1205 if (!nfs_do_recoalesce(desc)) 1206 break; 1207 } 1208 desc->pg_mirror_idx = restore_idx; 1209} 1210 1211/* 1212 * nfs_pageio_resend - Transfer requests to new descriptor and resend 1213 * @hdr - the pgio header to move request from 1214 * @desc - the pageio descriptor to add requests to 1215 * 1216 * Try to move each request (nfs_page) from @hdr to @desc then attempt 1217 * to send them. 1218 * 1219 * Returns 0 on success and < 0 on error. 1220 */ 1221int nfs_pageio_resend(struct nfs_pageio_descriptor *desc, 1222 struct nfs_pgio_header *hdr) 1223{ 1224 LIST_HEAD(failed); 1225 1226 desc->pg_dreq = hdr->dreq; 1227 while (!list_empty(&hdr->pages)) { 1228 struct nfs_page *req = nfs_list_entry(hdr->pages.next); 1229 1230 nfs_list_remove_request(req); 1231 if (!nfs_pageio_add_request(desc, req)) 1232 nfs_list_add_request(req, &failed); 1233 } 1234 nfs_pageio_complete(desc); 1235 if (!list_empty(&failed)) { 1236 list_move(&failed, &hdr->pages); 1237 return -EIO; 1238 } 1239 return 0; 1240} 1241EXPORT_SYMBOL_GPL(nfs_pageio_resend); 1242 1243/** 1244 * nfs_pageio_complete - Complete I/O then cleanup an nfs_pageio_descriptor 1245 * @desc: pointer to io descriptor 1246 */ 1247void nfs_pageio_complete(struct nfs_pageio_descriptor *desc) 1248{ 1249 u32 midx; 1250 1251 for (midx = 0; midx < desc->pg_mirror_count; midx++) 1252 nfs_pageio_complete_mirror(desc, midx); 1253 1254 if (desc->pg_ops->pg_cleanup) 1255 desc->pg_ops->pg_cleanup(desc); 1256 nfs_pageio_cleanup_mirroring(desc); 1257} 1258 1259/** 1260 * nfs_pageio_cond_complete - Conditional I/O completion 1261 * @desc: pointer to io descriptor 1262 * @index: page index 1263 * 1264 * It is important to ensure that processes don't try to take locks 1265 * on non-contiguous ranges of pages as that might deadlock. This 1266 * function should be called before attempting to wait on a locked 1267 * nfs_page. It will complete the I/O if the page index 'index' 1268 * is not contiguous with the existing list of pages in 'desc'. 1269 */ 1270void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index) 1271{ 1272 struct nfs_pgio_mirror *mirror; 1273 struct nfs_page *prev; 1274 u32 midx; 1275 1276 for (midx = 0; midx < desc->pg_mirror_count; midx++) { 1277 mirror = &desc->pg_mirrors[midx]; 1278 if (!list_empty(&mirror->pg_list)) { 1279 prev = nfs_list_entry(mirror->pg_list.prev); 1280 if (index != prev->wb_index + 1) 1281 nfs_pageio_complete_mirror(desc, midx); 1282 } 1283 } 1284} 1285 1286int __init nfs_init_nfspagecache(void) 1287{ 1288 nfs_page_cachep = kmem_cache_create("nfs_page", 1289 sizeof(struct nfs_page), 1290 0, SLAB_HWCACHE_ALIGN, 1291 NULL); 1292 if (nfs_page_cachep == NULL) 1293 return -ENOMEM; 1294 1295 return 0; 1296} 1297 1298void nfs_destroy_nfspagecache(void) 1299{ 1300 kmem_cache_destroy(nfs_page_cachep); 1301} 1302 1303static const struct rpc_call_ops nfs_pgio_common_ops = { 1304 .rpc_call_prepare = nfs_pgio_prepare, 1305 .rpc_call_done = nfs_pgio_result, 1306 .rpc_release = nfs_pgio_release, 1307}; 1308 1309const struct nfs_pageio_ops nfs_pgio_rw_ops = { 1310 .pg_test = nfs_generic_pg_test, 1311 .pg_doio = nfs_generic_pg_pgios, 1312}; 1313