| 1 | /* |
| 2 | * linux/fs/nfs/write.c |
| 3 | * |
| 4 | * Write file data over NFS. |
| 5 | * |
| 6 | * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de> |
| 7 | */ |
| 8 | |
| 9 | #include <linux/types.h> |
| 10 | #include <linux/slab.h> |
| 11 | #include <linux/mm.h> |
| 12 | #include <linux/pagemap.h> |
| 13 | #include <linux/file.h> |
| 14 | #include <linux/writeback.h> |
| 15 | #include <linux/swap.h> |
| 16 | #include <linux/migrate.h> |
| 17 | |
| 18 | #include <linux/sunrpc/clnt.h> |
| 19 | #include <linux/nfs_fs.h> |
| 20 | #include <linux/nfs_mount.h> |
| 21 | #include <linux/nfs_page.h> |
| 22 | #include <linux/backing-dev.h> |
| 23 | #include <linux/export.h> |
| 24 | |
| 25 | #include <asm/uaccess.h> |
| 26 | |
| 27 | #include "delegation.h" |
| 28 | #include "internal.h" |
| 29 | #include "iostat.h" |
| 30 | #include "nfs4_fs.h" |
| 31 | #include "fscache.h" |
| 32 | #include "pnfs.h" |
| 33 | |
| 34 | #include "nfstrace.h" |
| 35 | |
| 36 | #define NFSDBG_FACILITY NFSDBG_PAGECACHE |
| 37 | |
| 38 | #define MIN_POOL_WRITE (32) |
| 39 | #define MIN_POOL_COMMIT (4) |
| 40 | |
| 41 | /* |
| 42 | * Local function declarations |
| 43 | */ |
| 44 | static void nfs_redirty_request(struct nfs_page *req); |
| 45 | static const struct rpc_call_ops nfs_write_common_ops; |
| 46 | static const struct rpc_call_ops nfs_commit_ops; |
| 47 | static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops; |
| 48 | static const struct nfs_commit_completion_ops nfs_commit_completion_ops; |
| 49 | |
| 50 | static struct kmem_cache *nfs_wdata_cachep; |
| 51 | static mempool_t *nfs_wdata_mempool; |
| 52 | static struct kmem_cache *nfs_cdata_cachep; |
| 53 | static mempool_t *nfs_commit_mempool; |
| 54 | |
| 55 | struct nfs_commit_data *nfs_commitdata_alloc(void) |
| 56 | { |
| 57 | struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO); |
| 58 | |
| 59 | if (p) { |
| 60 | memset(p, 0, sizeof(*p)); |
| 61 | INIT_LIST_HEAD(&p->pages); |
| 62 | } |
| 63 | return p; |
| 64 | } |
| 65 | EXPORT_SYMBOL_GPL(nfs_commitdata_alloc); |
| 66 | |
| 67 | void nfs_commit_free(struct nfs_commit_data *p) |
| 68 | { |
| 69 | mempool_free(p, nfs_commit_mempool); |
| 70 | } |
| 71 | EXPORT_SYMBOL_GPL(nfs_commit_free); |
| 72 | |
| 73 | struct nfs_write_header *nfs_writehdr_alloc(void) |
| 74 | { |
| 75 | struct nfs_write_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO); |
| 76 | |
| 77 | if (p) { |
| 78 | struct nfs_pgio_header *hdr = &p->header; |
| 79 | |
| 80 | memset(p, 0, sizeof(*p)); |
| 81 | INIT_LIST_HEAD(&hdr->pages); |
| 82 | INIT_LIST_HEAD(&hdr->rpc_list); |
| 83 | spin_lock_init(&hdr->lock); |
| 84 | atomic_set(&hdr->refcnt, 0); |
| 85 | hdr->verf = &p->verf; |
| 86 | } |
| 87 | return p; |
| 88 | } |
| 89 | EXPORT_SYMBOL_GPL(nfs_writehdr_alloc); |
| 90 | |
| 91 | static struct nfs_write_data *nfs_writedata_alloc(struct nfs_pgio_header *hdr, |
| 92 | unsigned int pagecount) |
| 93 | { |
| 94 | struct nfs_write_data *data, *prealloc; |
| 95 | |
| 96 | prealloc = &container_of(hdr, struct nfs_write_header, header)->rpc_data; |
| 97 | if (prealloc->header == NULL) |
| 98 | data = prealloc; |
| 99 | else |
| 100 | data = kzalloc(sizeof(*data), GFP_KERNEL); |
| 101 | if (!data) |
| 102 | goto out; |
| 103 | |
| 104 | if (nfs_pgarray_set(&data->pages, pagecount)) { |
| 105 | data->header = hdr; |
| 106 | atomic_inc(&hdr->refcnt); |
| 107 | } else { |
| 108 | if (data != prealloc) |
| 109 | kfree(data); |
| 110 | data = NULL; |
| 111 | } |
| 112 | out: |
| 113 | return data; |
| 114 | } |
| 115 | |
| 116 | void nfs_writehdr_free(struct nfs_pgio_header *hdr) |
| 117 | { |
| 118 | struct nfs_write_header *whdr = container_of(hdr, struct nfs_write_header, header); |
| 119 | mempool_free(whdr, nfs_wdata_mempool); |
| 120 | } |
| 121 | EXPORT_SYMBOL_GPL(nfs_writehdr_free); |
| 122 | |
| 123 | void nfs_writedata_release(struct nfs_write_data *wdata) |
| 124 | { |
| 125 | struct nfs_pgio_header *hdr = wdata->header; |
| 126 | struct nfs_write_header *write_header = container_of(hdr, struct nfs_write_header, header); |
| 127 | |
| 128 | put_nfs_open_context(wdata->args.context); |
| 129 | if (wdata->pages.pagevec != wdata->pages.page_array) |
| 130 | kfree(wdata->pages.pagevec); |
| 131 | if (wdata == &write_header->rpc_data) { |
| 132 | wdata->header = NULL; |
| 133 | wdata = NULL; |
| 134 | } |
| 135 | if (atomic_dec_and_test(&hdr->refcnt)) |
| 136 | hdr->completion_ops->completion(hdr); |
| 137 | /* Note: we only free the rpc_task after callbacks are done. |
| 138 | * See the comment in rpc_free_task() for why |
| 139 | */ |
| 140 | kfree(wdata); |
| 141 | } |
| 142 | EXPORT_SYMBOL_GPL(nfs_writedata_release); |
| 143 | |
| 144 | static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error) |
| 145 | { |
| 146 | ctx->error = error; |
| 147 | smp_wmb(); |
| 148 | set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags); |
| 149 | } |
| 150 | |
| 151 | static struct nfs_page * |
| 152 | nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page) |
| 153 | { |
| 154 | struct nfs_page *req = NULL; |
| 155 | |
| 156 | if (PagePrivate(page)) |
| 157 | req = (struct nfs_page *)page_private(page); |
| 158 | else if (unlikely(PageSwapCache(page))) { |
| 159 | struct nfs_page *freq, *t; |
| 160 | |
| 161 | /* Linearly search the commit list for the correct req */ |
| 162 | list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) { |
| 163 | if (freq->wb_page == page) { |
| 164 | req = freq; |
| 165 | break; |
| 166 | } |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | if (req) |
| 171 | kref_get(&req->wb_kref); |
| 172 | |
| 173 | return req; |
| 174 | } |
| 175 | |
| 176 | static struct nfs_page *nfs_page_find_request(struct page *page) |
| 177 | { |
| 178 | struct inode *inode = page_file_mapping(page)->host; |
| 179 | struct nfs_page *req = NULL; |
| 180 | |
| 181 | spin_lock(&inode->i_lock); |
| 182 | req = nfs_page_find_request_locked(NFS_I(inode), page); |
| 183 | spin_unlock(&inode->i_lock); |
| 184 | return req; |
| 185 | } |
| 186 | |
| 187 | /* Adjust the file length if we're writing beyond the end */ |
| 188 | static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count) |
| 189 | { |
| 190 | struct inode *inode = page_file_mapping(page)->host; |
| 191 | loff_t end, i_size; |
| 192 | pgoff_t end_index; |
| 193 | |
| 194 | spin_lock(&inode->i_lock); |
| 195 | i_size = i_size_read(inode); |
| 196 | end_index = (i_size - 1) >> PAGE_CACHE_SHIFT; |
| 197 | if (i_size > 0 && page_file_index(page) < end_index) |
| 198 | goto out; |
| 199 | end = page_file_offset(page) + ((loff_t)offset+count); |
| 200 | if (i_size >= end) |
| 201 | goto out; |
| 202 | i_size_write(inode, end); |
| 203 | nfs_inc_stats(inode, NFSIOS_EXTENDWRITE); |
| 204 | out: |
| 205 | spin_unlock(&inode->i_lock); |
| 206 | } |
| 207 | |
| 208 | /* A writeback failed: mark the page as bad, and invalidate the page cache */ |
| 209 | static void nfs_set_pageerror(struct page *page) |
| 210 | { |
| 211 | nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page)); |
| 212 | } |
| 213 | |
| 214 | /* We can set the PG_uptodate flag if we see that a write request |
| 215 | * covers the full page. |
| 216 | */ |
| 217 | static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count) |
| 218 | { |
| 219 | if (PageUptodate(page)) |
| 220 | return; |
| 221 | if (base != 0) |
| 222 | return; |
| 223 | if (count != nfs_page_length(page)) |
| 224 | return; |
| 225 | SetPageUptodate(page); |
| 226 | } |
| 227 | |
| 228 | static int wb_priority(struct writeback_control *wbc) |
| 229 | { |
| 230 | if (wbc->for_reclaim) |
| 231 | return FLUSH_HIGHPRI | FLUSH_STABLE; |
| 232 | if (wbc->for_kupdate || wbc->for_background) |
| 233 | return FLUSH_LOWPRI | FLUSH_COND_STABLE; |
| 234 | return FLUSH_COND_STABLE; |
| 235 | } |
| 236 | |
| 237 | /* |
| 238 | * NFS congestion control |
| 239 | */ |
| 240 | |
| 241 | int nfs_congestion_kb; |
| 242 | |
| 243 | #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10)) |
| 244 | #define NFS_CONGESTION_OFF_THRESH \ |
| 245 | (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2)) |
| 246 | |
| 247 | static void nfs_set_page_writeback(struct page *page) |
| 248 | { |
| 249 | struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host); |
| 250 | int ret = test_set_page_writeback(page); |
| 251 | |
| 252 | WARN_ON_ONCE(ret != 0); |
| 253 | |
| 254 | if (atomic_long_inc_return(&nfss->writeback) > |
| 255 | NFS_CONGESTION_ON_THRESH) { |
| 256 | set_bdi_congested(&nfss->backing_dev_info, |
| 257 | BLK_RW_ASYNC); |
| 258 | } |
| 259 | } |
| 260 | |
| 261 | static void nfs_end_page_writeback(struct page *page) |
| 262 | { |
| 263 | struct inode *inode = page_file_mapping(page)->host; |
| 264 | struct nfs_server *nfss = NFS_SERVER(inode); |
| 265 | |
| 266 | end_page_writeback(page); |
| 267 | if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) |
| 268 | clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC); |
| 269 | } |
| 270 | |
| 271 | static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock) |
| 272 | { |
| 273 | struct inode *inode = page_file_mapping(page)->host; |
| 274 | struct nfs_page *req; |
| 275 | int ret; |
| 276 | |
| 277 | spin_lock(&inode->i_lock); |
| 278 | for (;;) { |
| 279 | req = nfs_page_find_request_locked(NFS_I(inode), page); |
| 280 | if (req == NULL) |
| 281 | break; |
| 282 | if (nfs_lock_request(req)) |
| 283 | break; |
| 284 | /* Note: If we hold the page lock, as is the case in nfs_writepage, |
| 285 | * then the call to nfs_lock_request() will always |
| 286 | * succeed provided that someone hasn't already marked the |
| 287 | * request as dirty (in which case we don't care). |
| 288 | */ |
| 289 | spin_unlock(&inode->i_lock); |
| 290 | if (!nonblock) |
| 291 | ret = nfs_wait_on_request(req); |
| 292 | else |
| 293 | ret = -EAGAIN; |
| 294 | nfs_release_request(req); |
| 295 | if (ret != 0) |
| 296 | return ERR_PTR(ret); |
| 297 | spin_lock(&inode->i_lock); |
| 298 | } |
| 299 | spin_unlock(&inode->i_lock); |
| 300 | return req; |
| 301 | } |
| 302 | |
| 303 | /* |
| 304 | * Find an associated nfs write request, and prepare to flush it out |
| 305 | * May return an error if the user signalled nfs_wait_on_request(). |
| 306 | */ |
| 307 | static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio, |
| 308 | struct page *page, bool nonblock) |
| 309 | { |
| 310 | struct nfs_page *req; |
| 311 | int ret = 0; |
| 312 | |
| 313 | req = nfs_find_and_lock_request(page, nonblock); |
| 314 | if (!req) |
| 315 | goto out; |
| 316 | ret = PTR_ERR(req); |
| 317 | if (IS_ERR(req)) |
| 318 | goto out; |
| 319 | |
| 320 | nfs_set_page_writeback(page); |
| 321 | WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags)); |
| 322 | |
| 323 | ret = 0; |
| 324 | if (!nfs_pageio_add_request(pgio, req)) { |
| 325 | nfs_redirty_request(req); |
| 326 | ret = pgio->pg_error; |
| 327 | } |
| 328 | out: |
| 329 | return ret; |
| 330 | } |
| 331 | |
| 332 | static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio) |
| 333 | { |
| 334 | struct inode *inode = page_file_mapping(page)->host; |
| 335 | int ret; |
| 336 | |
| 337 | nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE); |
| 338 | nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1); |
| 339 | |
| 340 | nfs_pageio_cond_complete(pgio, page_file_index(page)); |
| 341 | ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE); |
| 342 | if (ret == -EAGAIN) { |
| 343 | redirty_page_for_writepage(wbc, page); |
| 344 | ret = 0; |
| 345 | } |
| 346 | return ret; |
| 347 | } |
| 348 | |
| 349 | /* |
| 350 | * Write an mmapped page to the server. |
| 351 | */ |
| 352 | static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc) |
| 353 | { |
| 354 | struct nfs_pageio_descriptor pgio; |
| 355 | int err; |
| 356 | |
| 357 | NFS_PROTO(page_file_mapping(page)->host)->write_pageio_init(&pgio, |
| 358 | page->mapping->host, |
| 359 | wb_priority(wbc), |
| 360 | &nfs_async_write_completion_ops); |
| 361 | err = nfs_do_writepage(page, wbc, &pgio); |
| 362 | nfs_pageio_complete(&pgio); |
| 363 | if (err < 0) |
| 364 | return err; |
| 365 | if (pgio.pg_error < 0) |
| 366 | return pgio.pg_error; |
| 367 | return 0; |
| 368 | } |
| 369 | |
| 370 | int nfs_writepage(struct page *page, struct writeback_control *wbc) |
| 371 | { |
| 372 | int ret; |
| 373 | |
| 374 | ret = nfs_writepage_locked(page, wbc); |
| 375 | unlock_page(page); |
| 376 | return ret; |
| 377 | } |
| 378 | |
| 379 | static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data) |
| 380 | { |
| 381 | int ret; |
| 382 | |
| 383 | ret = nfs_do_writepage(page, wbc, data); |
| 384 | unlock_page(page); |
| 385 | return ret; |
| 386 | } |
| 387 | |
| 388 | int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc) |
| 389 | { |
| 390 | struct inode *inode = mapping->host; |
| 391 | unsigned long *bitlock = &NFS_I(inode)->flags; |
| 392 | struct nfs_pageio_descriptor pgio; |
| 393 | int err; |
| 394 | |
| 395 | /* Stop dirtying of new pages while we sync */ |
| 396 | err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING, |
| 397 | nfs_wait_bit_killable, TASK_KILLABLE); |
| 398 | if (err) |
| 399 | goto out_err; |
| 400 | |
| 401 | nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES); |
| 402 | |
| 403 | NFS_PROTO(inode)->write_pageio_init(&pgio, inode, wb_priority(wbc), &nfs_async_write_completion_ops); |
| 404 | err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio); |
| 405 | nfs_pageio_complete(&pgio); |
| 406 | |
| 407 | clear_bit_unlock(NFS_INO_FLUSHING, bitlock); |
| 408 | smp_mb__after_clear_bit(); |
| 409 | wake_up_bit(bitlock, NFS_INO_FLUSHING); |
| 410 | |
| 411 | if (err < 0) |
| 412 | goto out_err; |
| 413 | err = pgio.pg_error; |
| 414 | if (err < 0) |
| 415 | goto out_err; |
| 416 | return 0; |
| 417 | out_err: |
| 418 | return err; |
| 419 | } |
| 420 | |
| 421 | /* |
| 422 | * Insert a write request into an inode |
| 423 | */ |
| 424 | static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req) |
| 425 | { |
| 426 | struct nfs_inode *nfsi = NFS_I(inode); |
| 427 | |
| 428 | /* Lock the request! */ |
| 429 | nfs_lock_request(req); |
| 430 | |
| 431 | spin_lock(&inode->i_lock); |
| 432 | if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) |
| 433 | inode->i_version++; |
| 434 | /* |
| 435 | * Swap-space should not get truncated. Hence no need to plug the race |
| 436 | * with invalidate/truncate. |
| 437 | */ |
| 438 | if (likely(!PageSwapCache(req->wb_page))) { |
| 439 | set_bit(PG_MAPPED, &req->wb_flags); |
| 440 | SetPagePrivate(req->wb_page); |
| 441 | set_page_private(req->wb_page, (unsigned long)req); |
| 442 | } |
| 443 | nfsi->npages++; |
| 444 | kref_get(&req->wb_kref); |
| 445 | spin_unlock(&inode->i_lock); |
| 446 | } |
| 447 | |
| 448 | /* |
| 449 | * Remove a write request from an inode |
| 450 | */ |
| 451 | static void nfs_inode_remove_request(struct nfs_page *req) |
| 452 | { |
| 453 | struct inode *inode = req->wb_context->dentry->d_inode; |
| 454 | struct nfs_inode *nfsi = NFS_I(inode); |
| 455 | |
| 456 | spin_lock(&inode->i_lock); |
| 457 | if (likely(!PageSwapCache(req->wb_page))) { |
| 458 | set_page_private(req->wb_page, 0); |
| 459 | ClearPagePrivate(req->wb_page); |
| 460 | clear_bit(PG_MAPPED, &req->wb_flags); |
| 461 | } |
| 462 | nfsi->npages--; |
| 463 | spin_unlock(&inode->i_lock); |
| 464 | nfs_release_request(req); |
| 465 | } |
| 466 | |
| 467 | static void |
| 468 | nfs_mark_request_dirty(struct nfs_page *req) |
| 469 | { |
| 470 | __set_page_dirty_nobuffers(req->wb_page); |
| 471 | } |
| 472 | |
| 473 | #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) |
| 474 | /** |
| 475 | * nfs_request_add_commit_list - add request to a commit list |
| 476 | * @req: pointer to a struct nfs_page |
| 477 | * @dst: commit list head |
| 478 | * @cinfo: holds list lock and accounting info |
| 479 | * |
| 480 | * This sets the PG_CLEAN bit, updates the cinfo count of |
| 481 | * number of outstanding requests requiring a commit as well as |
| 482 | * the MM page stats. |
| 483 | * |
| 484 | * The caller must _not_ hold the cinfo->lock, but must be |
| 485 | * holding the nfs_page lock. |
| 486 | */ |
| 487 | void |
| 488 | nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst, |
| 489 | struct nfs_commit_info *cinfo) |
| 490 | { |
| 491 | set_bit(PG_CLEAN, &(req)->wb_flags); |
| 492 | spin_lock(cinfo->lock); |
| 493 | nfs_list_add_request(req, dst); |
| 494 | cinfo->mds->ncommit++; |
| 495 | spin_unlock(cinfo->lock); |
| 496 | if (!cinfo->dreq) { |
| 497 | inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS); |
| 498 | inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info, |
| 499 | BDI_RECLAIMABLE); |
| 500 | __mark_inode_dirty(req->wb_context->dentry->d_inode, |
| 501 | I_DIRTY_DATASYNC); |
| 502 | } |
| 503 | } |
| 504 | EXPORT_SYMBOL_GPL(nfs_request_add_commit_list); |
| 505 | |
| 506 | /** |
| 507 | * nfs_request_remove_commit_list - Remove request from a commit list |
| 508 | * @req: pointer to a nfs_page |
| 509 | * @cinfo: holds list lock and accounting info |
| 510 | * |
| 511 | * This clears the PG_CLEAN bit, and updates the cinfo's count of |
| 512 | * number of outstanding requests requiring a commit |
| 513 | * It does not update the MM page stats. |
| 514 | * |
| 515 | * The caller _must_ hold the cinfo->lock and the nfs_page lock. |
| 516 | */ |
| 517 | void |
| 518 | nfs_request_remove_commit_list(struct nfs_page *req, |
| 519 | struct nfs_commit_info *cinfo) |
| 520 | { |
| 521 | if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) |
| 522 | return; |
| 523 | nfs_list_remove_request(req); |
| 524 | cinfo->mds->ncommit--; |
| 525 | } |
| 526 | EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list); |
| 527 | |
| 528 | static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo, |
| 529 | struct inode *inode) |
| 530 | { |
| 531 | cinfo->lock = &inode->i_lock; |
| 532 | cinfo->mds = &NFS_I(inode)->commit_info; |
| 533 | cinfo->ds = pnfs_get_ds_info(inode); |
| 534 | cinfo->dreq = NULL; |
| 535 | cinfo->completion_ops = &nfs_commit_completion_ops; |
| 536 | } |
| 537 | |
| 538 | void nfs_init_cinfo(struct nfs_commit_info *cinfo, |
| 539 | struct inode *inode, |
| 540 | struct nfs_direct_req *dreq) |
| 541 | { |
| 542 | if (dreq) |
| 543 | nfs_init_cinfo_from_dreq(cinfo, dreq); |
| 544 | else |
| 545 | nfs_init_cinfo_from_inode(cinfo, inode); |
| 546 | } |
| 547 | EXPORT_SYMBOL_GPL(nfs_init_cinfo); |
| 548 | |
| 549 | /* |
| 550 | * Add a request to the inode's commit list. |
| 551 | */ |
| 552 | void |
| 553 | nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg, |
| 554 | struct nfs_commit_info *cinfo) |
| 555 | { |
| 556 | if (pnfs_mark_request_commit(req, lseg, cinfo)) |
| 557 | return; |
| 558 | nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo); |
| 559 | } |
| 560 | |
| 561 | static void |
| 562 | nfs_clear_page_commit(struct page *page) |
| 563 | { |
| 564 | dec_zone_page_state(page, NR_UNSTABLE_NFS); |
| 565 | dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE); |
| 566 | } |
| 567 | |
| 568 | static void |
| 569 | nfs_clear_request_commit(struct nfs_page *req) |
| 570 | { |
| 571 | if (test_bit(PG_CLEAN, &req->wb_flags)) { |
| 572 | struct inode *inode = req->wb_context->dentry->d_inode; |
| 573 | struct nfs_commit_info cinfo; |
| 574 | |
| 575 | nfs_init_cinfo_from_inode(&cinfo, inode); |
| 576 | if (!pnfs_clear_request_commit(req, &cinfo)) { |
| 577 | spin_lock(cinfo.lock); |
| 578 | nfs_request_remove_commit_list(req, &cinfo); |
| 579 | spin_unlock(cinfo.lock); |
| 580 | } |
| 581 | nfs_clear_page_commit(req->wb_page); |
| 582 | } |
| 583 | } |
| 584 | |
| 585 | static inline |
| 586 | int nfs_write_need_commit(struct nfs_write_data *data) |
| 587 | { |
| 588 | if (data->verf.committed == NFS_DATA_SYNC) |
| 589 | return data->header->lseg == NULL; |
| 590 | return data->verf.committed != NFS_FILE_SYNC; |
| 591 | } |
| 592 | |
| 593 | #else |
| 594 | static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo, |
| 595 | struct inode *inode) |
| 596 | { |
| 597 | } |
| 598 | |
| 599 | void nfs_init_cinfo(struct nfs_commit_info *cinfo, |
| 600 | struct inode *inode, |
| 601 | struct nfs_direct_req *dreq) |
| 602 | { |
| 603 | } |
| 604 | |
| 605 | void |
| 606 | nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg, |
| 607 | struct nfs_commit_info *cinfo) |
| 608 | { |
| 609 | } |
| 610 | |
| 611 | static void |
| 612 | nfs_clear_request_commit(struct nfs_page *req) |
| 613 | { |
| 614 | } |
| 615 | |
| 616 | static inline |
| 617 | int nfs_write_need_commit(struct nfs_write_data *data) |
| 618 | { |
| 619 | return 0; |
| 620 | } |
| 621 | |
| 622 | #endif |
| 623 | |
| 624 | static void nfs_write_completion(struct nfs_pgio_header *hdr) |
| 625 | { |
| 626 | struct nfs_commit_info cinfo; |
| 627 | unsigned long bytes = 0; |
| 628 | |
| 629 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) |
| 630 | goto out; |
| 631 | nfs_init_cinfo_from_inode(&cinfo, hdr->inode); |
| 632 | while (!list_empty(&hdr->pages)) { |
| 633 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); |
| 634 | |
| 635 | bytes += req->wb_bytes; |
| 636 | nfs_list_remove_request(req); |
| 637 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && |
| 638 | (hdr->good_bytes < bytes)) { |
| 639 | nfs_set_pageerror(req->wb_page); |
| 640 | nfs_context_set_write_error(req->wb_context, hdr->error); |
| 641 | goto remove_req; |
| 642 | } |
| 643 | if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) { |
| 644 | nfs_mark_request_dirty(req); |
| 645 | goto next; |
| 646 | } |
| 647 | if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) { |
| 648 | memcpy(&req->wb_verf, &hdr->verf->verifier, sizeof(req->wb_verf)); |
| 649 | nfs_mark_request_commit(req, hdr->lseg, &cinfo); |
| 650 | goto next; |
| 651 | } |
| 652 | remove_req: |
| 653 | nfs_inode_remove_request(req); |
| 654 | next: |
| 655 | nfs_unlock_request(req); |
| 656 | nfs_end_page_writeback(req->wb_page); |
| 657 | nfs_release_request(req); |
| 658 | } |
| 659 | out: |
| 660 | hdr->release(hdr); |
| 661 | } |
| 662 | |
| 663 | #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) |
| 664 | static unsigned long |
| 665 | nfs_reqs_to_commit(struct nfs_commit_info *cinfo) |
| 666 | { |
| 667 | return cinfo->mds->ncommit; |
| 668 | } |
| 669 | |
| 670 | /* cinfo->lock held by caller */ |
| 671 | int |
| 672 | nfs_scan_commit_list(struct list_head *src, struct list_head *dst, |
| 673 | struct nfs_commit_info *cinfo, int max) |
| 674 | { |
| 675 | struct nfs_page *req, *tmp; |
| 676 | int ret = 0; |
| 677 | |
| 678 | list_for_each_entry_safe(req, tmp, src, wb_list) { |
| 679 | if (!nfs_lock_request(req)) |
| 680 | continue; |
| 681 | kref_get(&req->wb_kref); |
| 682 | if (cond_resched_lock(cinfo->lock)) |
| 683 | list_safe_reset_next(req, tmp, wb_list); |
| 684 | nfs_request_remove_commit_list(req, cinfo); |
| 685 | nfs_list_add_request(req, dst); |
| 686 | ret++; |
| 687 | if ((ret == max) && !cinfo->dreq) |
| 688 | break; |
| 689 | } |
| 690 | return ret; |
| 691 | } |
| 692 | |
| 693 | /* |
| 694 | * nfs_scan_commit - Scan an inode for commit requests |
| 695 | * @inode: NFS inode to scan |
| 696 | * @dst: mds destination list |
| 697 | * @cinfo: mds and ds lists of reqs ready to commit |
| 698 | * |
| 699 | * Moves requests from the inode's 'commit' request list. |
| 700 | * The requests are *not* checked to ensure that they form a contiguous set. |
| 701 | */ |
| 702 | int |
| 703 | nfs_scan_commit(struct inode *inode, struct list_head *dst, |
| 704 | struct nfs_commit_info *cinfo) |
| 705 | { |
| 706 | int ret = 0; |
| 707 | |
| 708 | spin_lock(cinfo->lock); |
| 709 | if (cinfo->mds->ncommit > 0) { |
| 710 | const int max = INT_MAX; |
| 711 | |
| 712 | ret = nfs_scan_commit_list(&cinfo->mds->list, dst, |
| 713 | cinfo, max); |
| 714 | ret += pnfs_scan_commit_lists(inode, cinfo, max - ret); |
| 715 | } |
| 716 | spin_unlock(cinfo->lock); |
| 717 | return ret; |
| 718 | } |
| 719 | |
| 720 | #else |
| 721 | static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo) |
| 722 | { |
| 723 | return 0; |
| 724 | } |
| 725 | |
| 726 | int nfs_scan_commit(struct inode *inode, struct list_head *dst, |
| 727 | struct nfs_commit_info *cinfo) |
| 728 | { |
| 729 | return 0; |
| 730 | } |
| 731 | #endif |
| 732 | |
| 733 | /* |
| 734 | * Search for an existing write request, and attempt to update |
| 735 | * it to reflect a new dirty region on a given page. |
| 736 | * |
| 737 | * If the attempt fails, then the existing request is flushed out |
| 738 | * to disk. |
| 739 | */ |
| 740 | static struct nfs_page *nfs_try_to_update_request(struct inode *inode, |
| 741 | struct page *page, |
| 742 | unsigned int offset, |
| 743 | unsigned int bytes) |
| 744 | { |
| 745 | struct nfs_page *req; |
| 746 | unsigned int rqend; |
| 747 | unsigned int end; |
| 748 | int error; |
| 749 | |
| 750 | if (!PagePrivate(page)) |
| 751 | return NULL; |
| 752 | |
| 753 | end = offset + bytes; |
| 754 | spin_lock(&inode->i_lock); |
| 755 | |
| 756 | for (;;) { |
| 757 | req = nfs_page_find_request_locked(NFS_I(inode), page); |
| 758 | if (req == NULL) |
| 759 | goto out_unlock; |
| 760 | |
| 761 | rqend = req->wb_offset + req->wb_bytes; |
| 762 | /* |
| 763 | * Tell the caller to flush out the request if |
| 764 | * the offsets are non-contiguous. |
| 765 | * Note: nfs_flush_incompatible() will already |
| 766 | * have flushed out requests having wrong owners. |
| 767 | */ |
| 768 | if (offset > rqend |
| 769 | || end < req->wb_offset) |
| 770 | goto out_flushme; |
| 771 | |
| 772 | if (nfs_lock_request(req)) |
| 773 | break; |
| 774 | |
| 775 | /* The request is locked, so wait and then retry */ |
| 776 | spin_unlock(&inode->i_lock); |
| 777 | error = nfs_wait_on_request(req); |
| 778 | nfs_release_request(req); |
| 779 | if (error != 0) |
| 780 | goto out_err; |
| 781 | spin_lock(&inode->i_lock); |
| 782 | } |
| 783 | |
| 784 | /* Okay, the request matches. Update the region */ |
| 785 | if (offset < req->wb_offset) { |
| 786 | req->wb_offset = offset; |
| 787 | req->wb_pgbase = offset; |
| 788 | } |
| 789 | if (end > rqend) |
| 790 | req->wb_bytes = end - req->wb_offset; |
| 791 | else |
| 792 | req->wb_bytes = rqend - req->wb_offset; |
| 793 | out_unlock: |
| 794 | spin_unlock(&inode->i_lock); |
| 795 | if (req) |
| 796 | nfs_clear_request_commit(req); |
| 797 | return req; |
| 798 | out_flushme: |
| 799 | spin_unlock(&inode->i_lock); |
| 800 | nfs_release_request(req); |
| 801 | error = nfs_wb_page(inode, page); |
| 802 | out_err: |
| 803 | return ERR_PTR(error); |
| 804 | } |
| 805 | |
| 806 | /* |
| 807 | * Try to update an existing write request, or create one if there is none. |
| 808 | * |
| 809 | * Note: Should always be called with the Page Lock held to prevent races |
| 810 | * if we have to add a new request. Also assumes that the caller has |
| 811 | * already called nfs_flush_incompatible() if necessary. |
| 812 | */ |
| 813 | static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx, |
| 814 | struct page *page, unsigned int offset, unsigned int bytes) |
| 815 | { |
| 816 | struct inode *inode = page_file_mapping(page)->host; |
| 817 | struct nfs_page *req; |
| 818 | |
| 819 | req = nfs_try_to_update_request(inode, page, offset, bytes); |
| 820 | if (req != NULL) |
| 821 | goto out; |
| 822 | req = nfs_create_request(ctx, inode, page, offset, bytes); |
| 823 | if (IS_ERR(req)) |
| 824 | goto out; |
| 825 | nfs_inode_add_request(inode, req); |
| 826 | out: |
| 827 | return req; |
| 828 | } |
| 829 | |
| 830 | static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page, |
| 831 | unsigned int offset, unsigned int count) |
| 832 | { |
| 833 | struct nfs_page *req; |
| 834 | |
| 835 | req = nfs_setup_write_request(ctx, page, offset, count); |
| 836 | if (IS_ERR(req)) |
| 837 | return PTR_ERR(req); |
| 838 | /* Update file length */ |
| 839 | nfs_grow_file(page, offset, count); |
| 840 | nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes); |
| 841 | nfs_mark_request_dirty(req); |
| 842 | nfs_unlock_and_release_request(req); |
| 843 | return 0; |
| 844 | } |
| 845 | |
| 846 | int nfs_flush_incompatible(struct file *file, struct page *page) |
| 847 | { |
| 848 | struct nfs_open_context *ctx = nfs_file_open_context(file); |
| 849 | struct nfs_lock_context *l_ctx; |
| 850 | struct nfs_page *req; |
| 851 | int do_flush, status; |
| 852 | /* |
| 853 | * Look for a request corresponding to this page. If there |
| 854 | * is one, and it belongs to another file, we flush it out |
| 855 | * before we try to copy anything into the page. Do this |
| 856 | * due to the lack of an ACCESS-type call in NFSv2. |
| 857 | * Also do the same if we find a request from an existing |
| 858 | * dropped page. |
| 859 | */ |
| 860 | do { |
| 861 | req = nfs_page_find_request(page); |
| 862 | if (req == NULL) |
| 863 | return 0; |
| 864 | l_ctx = req->wb_lock_context; |
| 865 | do_flush = req->wb_page != page || req->wb_context != ctx; |
| 866 | if (l_ctx) { |
| 867 | do_flush |= l_ctx->lockowner.l_owner != current->files |
| 868 | || l_ctx->lockowner.l_pid != current->tgid; |
| 869 | } |
| 870 | nfs_release_request(req); |
| 871 | if (!do_flush) |
| 872 | return 0; |
| 873 | status = nfs_wb_page(page_file_mapping(page)->host, page); |
| 874 | } while (status == 0); |
| 875 | return status; |
| 876 | } |
| 877 | |
| 878 | /* |
| 879 | * Avoid buffered writes when a open context credential's key would |
| 880 | * expire soon. |
| 881 | * |
| 882 | * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL. |
| 883 | * |
| 884 | * Return 0 and set a credential flag which triggers the inode to flush |
| 885 | * and performs NFS_FILE_SYNC writes if the key will expired within |
| 886 | * RPC_KEY_EXPIRE_TIMEO. |
| 887 | */ |
| 888 | int |
| 889 | nfs_key_timeout_notify(struct file *filp, struct inode *inode) |
| 890 | { |
| 891 | struct nfs_open_context *ctx = nfs_file_open_context(filp); |
| 892 | struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth; |
| 893 | |
| 894 | return rpcauth_key_timeout_notify(auth, ctx->cred); |
| 895 | } |
| 896 | |
| 897 | /* |
| 898 | * Test if the open context credential key is marked to expire soon. |
| 899 | */ |
| 900 | bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx) |
| 901 | { |
| 902 | return rpcauth_cred_key_to_expire(ctx->cred); |
| 903 | } |
| 904 | |
| 905 | /* |
| 906 | * If the page cache is marked as unsafe or invalid, then we can't rely on |
| 907 | * the PageUptodate() flag. In this case, we will need to turn off |
| 908 | * write optimisations that depend on the page contents being correct. |
| 909 | */ |
| 910 | static bool nfs_write_pageuptodate(struct page *page, struct inode *inode) |
| 911 | { |
| 912 | if (nfs_have_delegated_attributes(inode)) |
| 913 | goto out; |
| 914 | if (NFS_I(inode)->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE)) |
| 915 | return false; |
| 916 | out: |
| 917 | return PageUptodate(page) != 0; |
| 918 | } |
| 919 | |
| 920 | /* If we know the page is up to date, and we're not using byte range locks (or |
| 921 | * if we have the whole file locked for writing), it may be more efficient to |
| 922 | * extend the write to cover the entire page in order to avoid fragmentation |
| 923 | * inefficiencies. |
| 924 | * |
| 925 | * If the file is opened for synchronous writes or if we have a write delegation |
| 926 | * from the server then we can just skip the rest of the checks. |
| 927 | */ |
| 928 | static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode) |
| 929 | { |
| 930 | if (file->f_flags & O_DSYNC) |
| 931 | return 0; |
| 932 | if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) |
| 933 | return 1; |
| 934 | if (nfs_write_pageuptodate(page, inode) && (inode->i_flock == NULL || |
| 935 | (inode->i_flock->fl_start == 0 && |
| 936 | inode->i_flock->fl_end == OFFSET_MAX && |
| 937 | inode->i_flock->fl_type != F_RDLCK))) |
| 938 | return 1; |
| 939 | return 0; |
| 940 | } |
| 941 | |
| 942 | /* |
| 943 | * Update and possibly write a cached page of an NFS file. |
| 944 | * |
| 945 | * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad |
| 946 | * things with a page scheduled for an RPC call (e.g. invalidate it). |
| 947 | */ |
| 948 | int nfs_updatepage(struct file *file, struct page *page, |
| 949 | unsigned int offset, unsigned int count) |
| 950 | { |
| 951 | struct nfs_open_context *ctx = nfs_file_open_context(file); |
| 952 | struct inode *inode = page_file_mapping(page)->host; |
| 953 | int status = 0; |
| 954 | |
| 955 | nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE); |
| 956 | |
| 957 | dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n", |
| 958 | file->f_path.dentry->d_parent->d_name.name, |
| 959 | file->f_path.dentry->d_name.name, count, |
| 960 | (long long)(page_file_offset(page) + offset)); |
| 961 | |
| 962 | if (nfs_can_extend_write(file, page, inode)) { |
| 963 | count = max(count + offset, nfs_page_length(page)); |
| 964 | offset = 0; |
| 965 | } |
| 966 | |
| 967 | status = nfs_writepage_setup(ctx, page, offset, count); |
| 968 | if (status < 0) |
| 969 | nfs_set_pageerror(page); |
| 970 | else |
| 971 | __set_page_dirty_nobuffers(page); |
| 972 | |
| 973 | dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n", |
| 974 | status, (long long)i_size_read(inode)); |
| 975 | return status; |
| 976 | } |
| 977 | |
| 978 | static int flush_task_priority(int how) |
| 979 | { |
| 980 | switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) { |
| 981 | case FLUSH_HIGHPRI: |
| 982 | return RPC_PRIORITY_HIGH; |
| 983 | case FLUSH_LOWPRI: |
| 984 | return RPC_PRIORITY_LOW; |
| 985 | } |
| 986 | return RPC_PRIORITY_NORMAL; |
| 987 | } |
| 988 | |
| 989 | int nfs_initiate_write(struct rpc_clnt *clnt, |
| 990 | struct nfs_write_data *data, |
| 991 | const struct rpc_call_ops *call_ops, |
| 992 | int how, int flags) |
| 993 | { |
| 994 | struct inode *inode = data->header->inode; |
| 995 | int priority = flush_task_priority(how); |
| 996 | struct rpc_task *task; |
| 997 | struct rpc_message msg = { |
| 998 | .rpc_argp = &data->args, |
| 999 | .rpc_resp = &data->res, |
| 1000 | .rpc_cred = data->header->cred, |
| 1001 | }; |
| 1002 | struct rpc_task_setup task_setup_data = { |
| 1003 | .rpc_client = clnt, |
| 1004 | .task = &data->task, |
| 1005 | .rpc_message = &msg, |
| 1006 | .callback_ops = call_ops, |
| 1007 | .callback_data = data, |
| 1008 | .workqueue = nfsiod_workqueue, |
| 1009 | .flags = RPC_TASK_ASYNC | flags, |
| 1010 | .priority = priority, |
| 1011 | }; |
| 1012 | int ret = 0; |
| 1013 | |
| 1014 | /* Set up the initial task struct. */ |
| 1015 | NFS_PROTO(inode)->write_setup(data, &msg); |
| 1016 | |
| 1017 | dprintk("NFS: %5u initiated write call " |
| 1018 | "(req %s/%lld, %u bytes @ offset %llu)\n", |
| 1019 | data->task.tk_pid, |
| 1020 | inode->i_sb->s_id, |
| 1021 | (long long)NFS_FILEID(inode), |
| 1022 | data->args.count, |
| 1023 | (unsigned long long)data->args.offset); |
| 1024 | |
| 1025 | task = rpc_run_task(&task_setup_data); |
| 1026 | if (IS_ERR(task)) { |
| 1027 | ret = PTR_ERR(task); |
| 1028 | goto out; |
| 1029 | } |
| 1030 | if (how & FLUSH_SYNC) { |
| 1031 | ret = rpc_wait_for_completion_task(task); |
| 1032 | if (ret == 0) |
| 1033 | ret = task->tk_status; |
| 1034 | } |
| 1035 | rpc_put_task(task); |
| 1036 | out: |
| 1037 | return ret; |
| 1038 | } |
| 1039 | EXPORT_SYMBOL_GPL(nfs_initiate_write); |
| 1040 | |
| 1041 | /* |
| 1042 | * Set up the argument/result storage required for the RPC call. |
| 1043 | */ |
| 1044 | static void nfs_write_rpcsetup(struct nfs_write_data *data, |
| 1045 | unsigned int count, unsigned int offset, |
| 1046 | int how, struct nfs_commit_info *cinfo) |
| 1047 | { |
| 1048 | struct nfs_page *req = data->header->req; |
| 1049 | |
| 1050 | /* Set up the RPC argument and reply structs |
| 1051 | * NB: take care not to mess about with data->commit et al. */ |
| 1052 | |
| 1053 | data->args.fh = NFS_FH(data->header->inode); |
| 1054 | data->args.offset = req_offset(req) + offset; |
| 1055 | /* pnfs_set_layoutcommit needs this */ |
| 1056 | data->mds_offset = data->args.offset; |
| 1057 | data->args.pgbase = req->wb_pgbase + offset; |
| 1058 | data->args.pages = data->pages.pagevec; |
| 1059 | data->args.count = count; |
| 1060 | data->args.context = get_nfs_open_context(req->wb_context); |
| 1061 | data->args.lock_context = req->wb_lock_context; |
| 1062 | data->args.stable = NFS_UNSTABLE; |
| 1063 | switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) { |
| 1064 | case 0: |
| 1065 | break; |
| 1066 | case FLUSH_COND_STABLE: |
| 1067 | if (nfs_reqs_to_commit(cinfo)) |
| 1068 | break; |
| 1069 | default: |
| 1070 | data->args.stable = NFS_FILE_SYNC; |
| 1071 | } |
| 1072 | |
| 1073 | data->res.fattr = &data->fattr; |
| 1074 | data->res.count = count; |
| 1075 | data->res.verf = &data->verf; |
| 1076 | nfs_fattr_init(&data->fattr); |
| 1077 | } |
| 1078 | |
| 1079 | static int nfs_do_write(struct nfs_write_data *data, |
| 1080 | const struct rpc_call_ops *call_ops, |
| 1081 | int how) |
| 1082 | { |
| 1083 | struct inode *inode = data->header->inode; |
| 1084 | |
| 1085 | return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0); |
| 1086 | } |
| 1087 | |
| 1088 | static int nfs_do_multiple_writes(struct list_head *head, |
| 1089 | const struct rpc_call_ops *call_ops, |
| 1090 | int how) |
| 1091 | { |
| 1092 | struct nfs_write_data *data; |
| 1093 | int ret = 0; |
| 1094 | |
| 1095 | while (!list_empty(head)) { |
| 1096 | int ret2; |
| 1097 | |
| 1098 | data = list_first_entry(head, struct nfs_write_data, list); |
| 1099 | list_del_init(&data->list); |
| 1100 | |
| 1101 | ret2 = nfs_do_write(data, call_ops, how); |
| 1102 | if (ret == 0) |
| 1103 | ret = ret2; |
| 1104 | } |
| 1105 | return ret; |
| 1106 | } |
| 1107 | |
| 1108 | /* If a nfs_flush_* function fails, it should remove reqs from @head and |
| 1109 | * call this on each, which will prepare them to be retried on next |
| 1110 | * writeback using standard nfs. |
| 1111 | */ |
| 1112 | static void nfs_redirty_request(struct nfs_page *req) |
| 1113 | { |
| 1114 | nfs_mark_request_dirty(req); |
| 1115 | nfs_unlock_request(req); |
| 1116 | nfs_end_page_writeback(req->wb_page); |
| 1117 | nfs_release_request(req); |
| 1118 | } |
| 1119 | |
| 1120 | static void nfs_async_write_error(struct list_head *head) |
| 1121 | { |
| 1122 | struct nfs_page *req; |
| 1123 | |
| 1124 | while (!list_empty(head)) { |
| 1125 | req = nfs_list_entry(head->next); |
| 1126 | nfs_list_remove_request(req); |
| 1127 | nfs_redirty_request(req); |
| 1128 | } |
| 1129 | } |
| 1130 | |
| 1131 | static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = { |
| 1132 | .error_cleanup = nfs_async_write_error, |
| 1133 | .completion = nfs_write_completion, |
| 1134 | }; |
| 1135 | |
| 1136 | static void nfs_flush_error(struct nfs_pageio_descriptor *desc, |
| 1137 | struct nfs_pgio_header *hdr) |
| 1138 | { |
| 1139 | set_bit(NFS_IOHDR_REDO, &hdr->flags); |
| 1140 | while (!list_empty(&hdr->rpc_list)) { |
| 1141 | struct nfs_write_data *data = list_first_entry(&hdr->rpc_list, |
| 1142 | struct nfs_write_data, list); |
| 1143 | list_del(&data->list); |
| 1144 | nfs_writedata_release(data); |
| 1145 | } |
| 1146 | desc->pg_completion_ops->error_cleanup(&desc->pg_list); |
| 1147 | } |
| 1148 | |
| 1149 | /* |
| 1150 | * Generate multiple small requests to write out a single |
| 1151 | * contiguous dirty area on one page. |
| 1152 | */ |
| 1153 | static int nfs_flush_multi(struct nfs_pageio_descriptor *desc, |
| 1154 | struct nfs_pgio_header *hdr) |
| 1155 | { |
| 1156 | struct nfs_page *req = hdr->req; |
| 1157 | struct page *page = req->wb_page; |
| 1158 | struct nfs_write_data *data; |
| 1159 | size_t wsize = desc->pg_bsize, nbytes; |
| 1160 | unsigned int offset; |
| 1161 | int requests = 0; |
| 1162 | struct nfs_commit_info cinfo; |
| 1163 | |
| 1164 | nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq); |
| 1165 | |
| 1166 | if ((desc->pg_ioflags & FLUSH_COND_STABLE) && |
| 1167 | (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) || |
| 1168 | desc->pg_count > wsize)) |
| 1169 | desc->pg_ioflags &= ~FLUSH_COND_STABLE; |
| 1170 | |
| 1171 | |
| 1172 | offset = 0; |
| 1173 | nbytes = desc->pg_count; |
| 1174 | do { |
| 1175 | size_t len = min(nbytes, wsize); |
| 1176 | |
| 1177 | data = nfs_writedata_alloc(hdr, 1); |
| 1178 | if (!data) { |
| 1179 | nfs_flush_error(desc, hdr); |
| 1180 | return -ENOMEM; |
| 1181 | } |
| 1182 | data->pages.pagevec[0] = page; |
| 1183 | nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo); |
| 1184 | list_add(&data->list, &hdr->rpc_list); |
| 1185 | requests++; |
| 1186 | nbytes -= len; |
| 1187 | offset += len; |
| 1188 | } while (nbytes != 0); |
| 1189 | nfs_list_remove_request(req); |
| 1190 | nfs_list_add_request(req, &hdr->pages); |
| 1191 | desc->pg_rpc_callops = &nfs_write_common_ops; |
| 1192 | return 0; |
| 1193 | } |
| 1194 | |
| 1195 | /* |
| 1196 | * Create an RPC task for the given write request and kick it. |
| 1197 | * The page must have been locked by the caller. |
| 1198 | * |
| 1199 | * It may happen that the page we're passed is not marked dirty. |
| 1200 | * This is the case if nfs_updatepage detects a conflicting request |
| 1201 | * that has been written but not committed. |
| 1202 | */ |
| 1203 | static int nfs_flush_one(struct nfs_pageio_descriptor *desc, |
| 1204 | struct nfs_pgio_header *hdr) |
| 1205 | { |
| 1206 | struct nfs_page *req; |
| 1207 | struct page **pages; |
| 1208 | struct nfs_write_data *data; |
| 1209 | struct list_head *head = &desc->pg_list; |
| 1210 | struct nfs_commit_info cinfo; |
| 1211 | |
| 1212 | data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base, |
| 1213 | desc->pg_count)); |
| 1214 | if (!data) { |
| 1215 | nfs_flush_error(desc, hdr); |
| 1216 | return -ENOMEM; |
| 1217 | } |
| 1218 | |
| 1219 | nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq); |
| 1220 | pages = data->pages.pagevec; |
| 1221 | while (!list_empty(head)) { |
| 1222 | req = nfs_list_entry(head->next); |
| 1223 | nfs_list_remove_request(req); |
| 1224 | nfs_list_add_request(req, &hdr->pages); |
| 1225 | *pages++ = req->wb_page; |
| 1226 | } |
| 1227 | |
| 1228 | if ((desc->pg_ioflags & FLUSH_COND_STABLE) && |
| 1229 | (desc->pg_moreio || nfs_reqs_to_commit(&cinfo))) |
| 1230 | desc->pg_ioflags &= ~FLUSH_COND_STABLE; |
| 1231 | |
| 1232 | /* Set up the argument struct */ |
| 1233 | nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo); |
| 1234 | list_add(&data->list, &hdr->rpc_list); |
| 1235 | desc->pg_rpc_callops = &nfs_write_common_ops; |
| 1236 | return 0; |
| 1237 | } |
| 1238 | |
| 1239 | int nfs_generic_flush(struct nfs_pageio_descriptor *desc, |
| 1240 | struct nfs_pgio_header *hdr) |
| 1241 | { |
| 1242 | if (desc->pg_bsize < PAGE_CACHE_SIZE) |
| 1243 | return nfs_flush_multi(desc, hdr); |
| 1244 | return nfs_flush_one(desc, hdr); |
| 1245 | } |
| 1246 | EXPORT_SYMBOL_GPL(nfs_generic_flush); |
| 1247 | |
| 1248 | static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc) |
| 1249 | { |
| 1250 | struct nfs_write_header *whdr; |
| 1251 | struct nfs_pgio_header *hdr; |
| 1252 | int ret; |
| 1253 | |
| 1254 | whdr = nfs_writehdr_alloc(); |
| 1255 | if (!whdr) { |
| 1256 | desc->pg_completion_ops->error_cleanup(&desc->pg_list); |
| 1257 | return -ENOMEM; |
| 1258 | } |
| 1259 | hdr = &whdr->header; |
| 1260 | nfs_pgheader_init(desc, hdr, nfs_writehdr_free); |
| 1261 | atomic_inc(&hdr->refcnt); |
| 1262 | ret = nfs_generic_flush(desc, hdr); |
| 1263 | if (ret == 0) |
| 1264 | ret = nfs_do_multiple_writes(&hdr->rpc_list, |
| 1265 | desc->pg_rpc_callops, |
| 1266 | desc->pg_ioflags); |
| 1267 | if (atomic_dec_and_test(&hdr->refcnt)) |
| 1268 | hdr->completion_ops->completion(hdr); |
| 1269 | return ret; |
| 1270 | } |
| 1271 | |
| 1272 | static const struct nfs_pageio_ops nfs_pageio_write_ops = { |
| 1273 | .pg_test = nfs_generic_pg_test, |
| 1274 | .pg_doio = nfs_generic_pg_writepages, |
| 1275 | }; |
| 1276 | |
| 1277 | void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, |
| 1278 | struct inode *inode, int ioflags, |
| 1279 | const struct nfs_pgio_completion_ops *compl_ops) |
| 1280 | { |
| 1281 | nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops, |
| 1282 | NFS_SERVER(inode)->wsize, ioflags); |
| 1283 | } |
| 1284 | EXPORT_SYMBOL_GPL(nfs_pageio_init_write); |
| 1285 | |
| 1286 | void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio) |
| 1287 | { |
| 1288 | pgio->pg_ops = &nfs_pageio_write_ops; |
| 1289 | pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize; |
| 1290 | } |
| 1291 | EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds); |
| 1292 | |
| 1293 | |
| 1294 | void nfs_write_prepare(struct rpc_task *task, void *calldata) |
| 1295 | { |
| 1296 | struct nfs_write_data *data = calldata; |
| 1297 | int err; |
| 1298 | err = NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data); |
| 1299 | if (err) |
| 1300 | rpc_exit(task, err); |
| 1301 | } |
| 1302 | |
| 1303 | void nfs_commit_prepare(struct rpc_task *task, void *calldata) |
| 1304 | { |
| 1305 | struct nfs_commit_data *data = calldata; |
| 1306 | |
| 1307 | NFS_PROTO(data->inode)->commit_rpc_prepare(task, data); |
| 1308 | } |
| 1309 | |
| 1310 | /* |
| 1311 | * Handle a write reply that flushes a whole page. |
| 1312 | * |
| 1313 | * FIXME: There is an inherent race with invalidate_inode_pages and |
| 1314 | * writebacks since the page->count is kept > 1 for as long |
| 1315 | * as the page has a write request pending. |
| 1316 | */ |
| 1317 | static void nfs_writeback_done_common(struct rpc_task *task, void *calldata) |
| 1318 | { |
| 1319 | struct nfs_write_data *data = calldata; |
| 1320 | |
| 1321 | nfs_writeback_done(task, data); |
| 1322 | } |
| 1323 | |
| 1324 | static void nfs_writeback_release_common(void *calldata) |
| 1325 | { |
| 1326 | struct nfs_write_data *data = calldata; |
| 1327 | struct nfs_pgio_header *hdr = data->header; |
| 1328 | int status = data->task.tk_status; |
| 1329 | |
| 1330 | if ((status >= 0) && nfs_write_need_commit(data)) { |
| 1331 | spin_lock(&hdr->lock); |
| 1332 | if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) |
| 1333 | ; /* Do nothing */ |
| 1334 | else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) |
| 1335 | memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf)); |
| 1336 | else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf))) |
| 1337 | set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags); |
| 1338 | spin_unlock(&hdr->lock); |
| 1339 | } |
| 1340 | nfs_writedata_release(data); |
| 1341 | } |
| 1342 | |
| 1343 | static const struct rpc_call_ops nfs_write_common_ops = { |
| 1344 | .rpc_call_prepare = nfs_write_prepare, |
| 1345 | .rpc_call_done = nfs_writeback_done_common, |
| 1346 | .rpc_release = nfs_writeback_release_common, |
| 1347 | }; |
| 1348 | |
| 1349 | |
| 1350 | /* |
| 1351 | * This function is called when the WRITE call is complete. |
| 1352 | */ |
| 1353 | void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data) |
| 1354 | { |
| 1355 | struct nfs_writeargs *argp = &data->args; |
| 1356 | struct nfs_writeres *resp = &data->res; |
| 1357 | struct inode *inode = data->header->inode; |
| 1358 | int status; |
| 1359 | |
| 1360 | dprintk("NFS: %5u nfs_writeback_done (status %d)\n", |
| 1361 | task->tk_pid, task->tk_status); |
| 1362 | |
| 1363 | /* |
| 1364 | * ->write_done will attempt to use post-op attributes to detect |
| 1365 | * conflicting writes by other clients. A strict interpretation |
| 1366 | * of close-to-open would allow us to continue caching even if |
| 1367 | * another writer had changed the file, but some applications |
| 1368 | * depend on tighter cache coherency when writing. |
| 1369 | */ |
| 1370 | status = NFS_PROTO(inode)->write_done(task, data); |
| 1371 | if (status != 0) |
| 1372 | return; |
| 1373 | nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count); |
| 1374 | |
| 1375 | #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) |
| 1376 | if (resp->verf->committed < argp->stable && task->tk_status >= 0) { |
| 1377 | /* We tried a write call, but the server did not |
| 1378 | * commit data to stable storage even though we |
| 1379 | * requested it. |
| 1380 | * Note: There is a known bug in Tru64 < 5.0 in which |
| 1381 | * the server reports NFS_DATA_SYNC, but performs |
| 1382 | * NFS_FILE_SYNC. We therefore implement this checking |
| 1383 | * as a dprintk() in order to avoid filling syslog. |
| 1384 | */ |
| 1385 | static unsigned long complain; |
| 1386 | |
| 1387 | /* Note this will print the MDS for a DS write */ |
| 1388 | if (time_before(complain, jiffies)) { |
| 1389 | dprintk("NFS: faulty NFS server %s:" |
| 1390 | " (committed = %d) != (stable = %d)\n", |
| 1391 | NFS_SERVER(inode)->nfs_client->cl_hostname, |
| 1392 | resp->verf->committed, argp->stable); |
| 1393 | complain = jiffies + 300 * HZ; |
| 1394 | } |
| 1395 | } |
| 1396 | #endif |
| 1397 | if (task->tk_status < 0) |
| 1398 | nfs_set_pgio_error(data->header, task->tk_status, argp->offset); |
| 1399 | else if (resp->count < argp->count) { |
| 1400 | static unsigned long complain; |
| 1401 | |
| 1402 | /* This a short write! */ |
| 1403 | nfs_inc_stats(inode, NFSIOS_SHORTWRITE); |
| 1404 | |
| 1405 | /* Has the server at least made some progress? */ |
| 1406 | if (resp->count == 0) { |
| 1407 | if (time_before(complain, jiffies)) { |
| 1408 | printk(KERN_WARNING |
| 1409 | "NFS: Server wrote zero bytes, expected %u.\n", |
| 1410 | argp->count); |
| 1411 | complain = jiffies + 300 * HZ; |
| 1412 | } |
| 1413 | nfs_set_pgio_error(data->header, -EIO, argp->offset); |
| 1414 | task->tk_status = -EIO; |
| 1415 | return; |
| 1416 | } |
| 1417 | /* Was this an NFSv2 write or an NFSv3 stable write? */ |
| 1418 | if (resp->verf->committed != NFS_UNSTABLE) { |
| 1419 | /* Resend from where the server left off */ |
| 1420 | data->mds_offset += resp->count; |
| 1421 | argp->offset += resp->count; |
| 1422 | argp->pgbase += resp->count; |
| 1423 | argp->count -= resp->count; |
| 1424 | } else { |
| 1425 | /* Resend as a stable write in order to avoid |
| 1426 | * headaches in the case of a server crash. |
| 1427 | */ |
| 1428 | argp->stable = NFS_FILE_SYNC; |
| 1429 | } |
| 1430 | rpc_restart_call_prepare(task); |
| 1431 | } |
| 1432 | } |
| 1433 | |
| 1434 | |
| 1435 | #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) |
| 1436 | static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait) |
| 1437 | { |
| 1438 | int ret; |
| 1439 | |
| 1440 | if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags)) |
| 1441 | return 1; |
| 1442 | if (!may_wait) |
| 1443 | return 0; |
| 1444 | ret = out_of_line_wait_on_bit_lock(&nfsi->flags, |
| 1445 | NFS_INO_COMMIT, |
| 1446 | nfs_wait_bit_killable, |
| 1447 | TASK_KILLABLE); |
| 1448 | return (ret < 0) ? ret : 1; |
| 1449 | } |
| 1450 | |
| 1451 | static void nfs_commit_clear_lock(struct nfs_inode *nfsi) |
| 1452 | { |
| 1453 | clear_bit(NFS_INO_COMMIT, &nfsi->flags); |
| 1454 | smp_mb__after_clear_bit(); |
| 1455 | wake_up_bit(&nfsi->flags, NFS_INO_COMMIT); |
| 1456 | } |
| 1457 | |
| 1458 | void nfs_commitdata_release(struct nfs_commit_data *data) |
| 1459 | { |
| 1460 | put_nfs_open_context(data->context); |
| 1461 | nfs_commit_free(data); |
| 1462 | } |
| 1463 | EXPORT_SYMBOL_GPL(nfs_commitdata_release); |
| 1464 | |
| 1465 | int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data, |
| 1466 | const struct rpc_call_ops *call_ops, |
| 1467 | int how, int flags) |
| 1468 | { |
| 1469 | struct rpc_task *task; |
| 1470 | int priority = flush_task_priority(how); |
| 1471 | struct rpc_message msg = { |
| 1472 | .rpc_argp = &data->args, |
| 1473 | .rpc_resp = &data->res, |
| 1474 | .rpc_cred = data->cred, |
| 1475 | }; |
| 1476 | struct rpc_task_setup task_setup_data = { |
| 1477 | .task = &data->task, |
| 1478 | .rpc_client = clnt, |
| 1479 | .rpc_message = &msg, |
| 1480 | .callback_ops = call_ops, |
| 1481 | .callback_data = data, |
| 1482 | .workqueue = nfsiod_workqueue, |
| 1483 | .flags = RPC_TASK_ASYNC | flags, |
| 1484 | .priority = priority, |
| 1485 | }; |
| 1486 | /* Set up the initial task struct. */ |
| 1487 | NFS_PROTO(data->inode)->commit_setup(data, &msg); |
| 1488 | |
| 1489 | dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid); |
| 1490 | |
| 1491 | task = rpc_run_task(&task_setup_data); |
| 1492 | if (IS_ERR(task)) |
| 1493 | return PTR_ERR(task); |
| 1494 | if (how & FLUSH_SYNC) |
| 1495 | rpc_wait_for_completion_task(task); |
| 1496 | rpc_put_task(task); |
| 1497 | return 0; |
| 1498 | } |
| 1499 | EXPORT_SYMBOL_GPL(nfs_initiate_commit); |
| 1500 | |
| 1501 | /* |
| 1502 | * Set up the argument/result storage required for the RPC call. |
| 1503 | */ |
| 1504 | void nfs_init_commit(struct nfs_commit_data *data, |
| 1505 | struct list_head *head, |
| 1506 | struct pnfs_layout_segment *lseg, |
| 1507 | struct nfs_commit_info *cinfo) |
| 1508 | { |
| 1509 | struct nfs_page *first = nfs_list_entry(head->next); |
| 1510 | struct inode *inode = first->wb_context->dentry->d_inode; |
| 1511 | |
| 1512 | /* Set up the RPC argument and reply structs |
| 1513 | * NB: take care not to mess about with data->commit et al. */ |
| 1514 | |
| 1515 | list_splice_init(head, &data->pages); |
| 1516 | |
| 1517 | data->inode = inode; |
| 1518 | data->cred = first->wb_context->cred; |
| 1519 | data->lseg = lseg; /* reference transferred */ |
| 1520 | data->mds_ops = &nfs_commit_ops; |
| 1521 | data->completion_ops = cinfo->completion_ops; |
| 1522 | data->dreq = cinfo->dreq; |
| 1523 | |
| 1524 | data->args.fh = NFS_FH(data->inode); |
| 1525 | /* Note: we always request a commit of the entire inode */ |
| 1526 | data->args.offset = 0; |
| 1527 | data->args.count = 0; |
| 1528 | data->context = get_nfs_open_context(first->wb_context); |
| 1529 | data->res.fattr = &data->fattr; |
| 1530 | data->res.verf = &data->verf; |
| 1531 | nfs_fattr_init(&data->fattr); |
| 1532 | } |
| 1533 | EXPORT_SYMBOL_GPL(nfs_init_commit); |
| 1534 | |
| 1535 | void nfs_retry_commit(struct list_head *page_list, |
| 1536 | struct pnfs_layout_segment *lseg, |
| 1537 | struct nfs_commit_info *cinfo) |
| 1538 | { |
| 1539 | struct nfs_page *req; |
| 1540 | |
| 1541 | while (!list_empty(page_list)) { |
| 1542 | req = nfs_list_entry(page_list->next); |
| 1543 | nfs_list_remove_request(req); |
| 1544 | nfs_mark_request_commit(req, lseg, cinfo); |
| 1545 | if (!cinfo->dreq) { |
| 1546 | dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS); |
| 1547 | dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info, |
| 1548 | BDI_RECLAIMABLE); |
| 1549 | } |
| 1550 | nfs_unlock_and_release_request(req); |
| 1551 | } |
| 1552 | } |
| 1553 | EXPORT_SYMBOL_GPL(nfs_retry_commit); |
| 1554 | |
| 1555 | /* |
| 1556 | * Commit dirty pages |
| 1557 | */ |
| 1558 | static int |
| 1559 | nfs_commit_list(struct inode *inode, struct list_head *head, int how, |
| 1560 | struct nfs_commit_info *cinfo) |
| 1561 | { |
| 1562 | struct nfs_commit_data *data; |
| 1563 | |
| 1564 | data = nfs_commitdata_alloc(); |
| 1565 | |
| 1566 | if (!data) |
| 1567 | goto out_bad; |
| 1568 | |
| 1569 | /* Set up the argument struct */ |
| 1570 | nfs_init_commit(data, head, NULL, cinfo); |
| 1571 | atomic_inc(&cinfo->mds->rpcs_out); |
| 1572 | return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops, |
| 1573 | how, 0); |
| 1574 | out_bad: |
| 1575 | nfs_retry_commit(head, NULL, cinfo); |
| 1576 | cinfo->completion_ops->error_cleanup(NFS_I(inode)); |
| 1577 | return -ENOMEM; |
| 1578 | } |
| 1579 | |
| 1580 | /* |
| 1581 | * COMMIT call returned |
| 1582 | */ |
| 1583 | static void nfs_commit_done(struct rpc_task *task, void *calldata) |
| 1584 | { |
| 1585 | struct nfs_commit_data *data = calldata; |
| 1586 | |
| 1587 | dprintk("NFS: %5u nfs_commit_done (status %d)\n", |
| 1588 | task->tk_pid, task->tk_status); |
| 1589 | |
| 1590 | /* Call the NFS version-specific code */ |
| 1591 | NFS_PROTO(data->inode)->commit_done(task, data); |
| 1592 | } |
| 1593 | |
| 1594 | static void nfs_commit_release_pages(struct nfs_commit_data *data) |
| 1595 | { |
| 1596 | struct nfs_page *req; |
| 1597 | int status = data->task.tk_status; |
| 1598 | struct nfs_commit_info cinfo; |
| 1599 | |
| 1600 | while (!list_empty(&data->pages)) { |
| 1601 | req = nfs_list_entry(data->pages.next); |
| 1602 | nfs_list_remove_request(req); |
| 1603 | nfs_clear_page_commit(req->wb_page); |
| 1604 | |
| 1605 | dprintk("NFS: commit (%s/%lld %d@%lld)", |
| 1606 | req->wb_context->dentry->d_sb->s_id, |
| 1607 | (long long)NFS_FILEID(req->wb_context->dentry->d_inode), |
| 1608 | req->wb_bytes, |
| 1609 | (long long)req_offset(req)); |
| 1610 | if (status < 0) { |
| 1611 | nfs_context_set_write_error(req->wb_context, status); |
| 1612 | nfs_inode_remove_request(req); |
| 1613 | dprintk(", error = %d\n", status); |
| 1614 | goto next; |
| 1615 | } |
| 1616 | |
| 1617 | /* Okay, COMMIT succeeded, apparently. Check the verifier |
| 1618 | * returned by the server against all stored verfs. */ |
| 1619 | if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) { |
| 1620 | /* We have a match */ |
| 1621 | nfs_inode_remove_request(req); |
| 1622 | dprintk(" OK\n"); |
| 1623 | goto next; |
| 1624 | } |
| 1625 | /* We have a mismatch. Write the page again */ |
| 1626 | dprintk(" mismatch\n"); |
| 1627 | nfs_mark_request_dirty(req); |
| 1628 | set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags); |
| 1629 | next: |
| 1630 | nfs_unlock_and_release_request(req); |
| 1631 | } |
| 1632 | nfs_init_cinfo(&cinfo, data->inode, data->dreq); |
| 1633 | if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) |
| 1634 | nfs_commit_clear_lock(NFS_I(data->inode)); |
| 1635 | } |
| 1636 | |
| 1637 | static void nfs_commit_release(void *calldata) |
| 1638 | { |
| 1639 | struct nfs_commit_data *data = calldata; |
| 1640 | |
| 1641 | data->completion_ops->completion(data); |
| 1642 | nfs_commitdata_release(calldata); |
| 1643 | } |
| 1644 | |
| 1645 | static const struct rpc_call_ops nfs_commit_ops = { |
| 1646 | .rpc_call_prepare = nfs_commit_prepare, |
| 1647 | .rpc_call_done = nfs_commit_done, |
| 1648 | .rpc_release = nfs_commit_release, |
| 1649 | }; |
| 1650 | |
| 1651 | static const struct nfs_commit_completion_ops nfs_commit_completion_ops = { |
| 1652 | .completion = nfs_commit_release_pages, |
| 1653 | .error_cleanup = nfs_commit_clear_lock, |
| 1654 | }; |
| 1655 | |
| 1656 | int nfs_generic_commit_list(struct inode *inode, struct list_head *head, |
| 1657 | int how, struct nfs_commit_info *cinfo) |
| 1658 | { |
| 1659 | int status; |
| 1660 | |
| 1661 | status = pnfs_commit_list(inode, head, how, cinfo); |
| 1662 | if (status == PNFS_NOT_ATTEMPTED) |
| 1663 | status = nfs_commit_list(inode, head, how, cinfo); |
| 1664 | return status; |
| 1665 | } |
| 1666 | |
| 1667 | int nfs_commit_inode(struct inode *inode, int how) |
| 1668 | { |
| 1669 | LIST_HEAD(head); |
| 1670 | struct nfs_commit_info cinfo; |
| 1671 | int may_wait = how & FLUSH_SYNC; |
| 1672 | int res; |
| 1673 | |
| 1674 | res = nfs_commit_set_lock(NFS_I(inode), may_wait); |
| 1675 | if (res <= 0) |
| 1676 | goto out_mark_dirty; |
| 1677 | nfs_init_cinfo_from_inode(&cinfo, inode); |
| 1678 | res = nfs_scan_commit(inode, &head, &cinfo); |
| 1679 | if (res) { |
| 1680 | int error; |
| 1681 | |
| 1682 | error = nfs_generic_commit_list(inode, &head, how, &cinfo); |
| 1683 | if (error < 0) |
| 1684 | return error; |
| 1685 | if (!may_wait) |
| 1686 | goto out_mark_dirty; |
| 1687 | error = wait_on_bit(&NFS_I(inode)->flags, |
| 1688 | NFS_INO_COMMIT, |
| 1689 | nfs_wait_bit_killable, |
| 1690 | TASK_KILLABLE); |
| 1691 | if (error < 0) |
| 1692 | return error; |
| 1693 | } else |
| 1694 | nfs_commit_clear_lock(NFS_I(inode)); |
| 1695 | return res; |
| 1696 | /* Note: If we exit without ensuring that the commit is complete, |
| 1697 | * we must mark the inode as dirty. Otherwise, future calls to |
| 1698 | * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure |
| 1699 | * that the data is on the disk. |
| 1700 | */ |
| 1701 | out_mark_dirty: |
| 1702 | __mark_inode_dirty(inode, I_DIRTY_DATASYNC); |
| 1703 | return res; |
| 1704 | } |
| 1705 | |
| 1706 | static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc) |
| 1707 | { |
| 1708 | struct nfs_inode *nfsi = NFS_I(inode); |
| 1709 | int flags = FLUSH_SYNC; |
| 1710 | int ret = 0; |
| 1711 | |
| 1712 | /* no commits means nothing needs to be done */ |
| 1713 | if (!nfsi->commit_info.ncommit) |
| 1714 | return ret; |
| 1715 | |
| 1716 | if (wbc->sync_mode == WB_SYNC_NONE) { |
| 1717 | /* Don't commit yet if this is a non-blocking flush and there |
| 1718 | * are a lot of outstanding writes for this mapping. |
| 1719 | */ |
| 1720 | if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1)) |
| 1721 | goto out_mark_dirty; |
| 1722 | |
| 1723 | /* don't wait for the COMMIT response */ |
| 1724 | flags = 0; |
| 1725 | } |
| 1726 | |
| 1727 | ret = nfs_commit_inode(inode, flags); |
| 1728 | if (ret >= 0) { |
| 1729 | if (wbc->sync_mode == WB_SYNC_NONE) { |
| 1730 | if (ret < wbc->nr_to_write) |
| 1731 | wbc->nr_to_write -= ret; |
| 1732 | else |
| 1733 | wbc->nr_to_write = 0; |
| 1734 | } |
| 1735 | return 0; |
| 1736 | } |
| 1737 | out_mark_dirty: |
| 1738 | __mark_inode_dirty(inode, I_DIRTY_DATASYNC); |
| 1739 | return ret; |
| 1740 | } |
| 1741 | #else |
| 1742 | static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc) |
| 1743 | { |
| 1744 | return 0; |
| 1745 | } |
| 1746 | #endif |
| 1747 | |
| 1748 | int nfs_write_inode(struct inode *inode, struct writeback_control *wbc) |
| 1749 | { |
| 1750 | return nfs_commit_unstable_pages(inode, wbc); |
| 1751 | } |
| 1752 | EXPORT_SYMBOL_GPL(nfs_write_inode); |
| 1753 | |
| 1754 | /* |
| 1755 | * flush the inode to disk. |
| 1756 | */ |
| 1757 | int nfs_wb_all(struct inode *inode) |
| 1758 | { |
| 1759 | struct writeback_control wbc = { |
| 1760 | .sync_mode = WB_SYNC_ALL, |
| 1761 | .nr_to_write = LONG_MAX, |
| 1762 | .range_start = 0, |
| 1763 | .range_end = LLONG_MAX, |
| 1764 | }; |
| 1765 | int ret; |
| 1766 | |
| 1767 | trace_nfs_writeback_inode_enter(inode); |
| 1768 | |
| 1769 | ret = sync_inode(inode, &wbc); |
| 1770 | |
| 1771 | trace_nfs_writeback_inode_exit(inode, ret); |
| 1772 | return ret; |
| 1773 | } |
| 1774 | EXPORT_SYMBOL_GPL(nfs_wb_all); |
| 1775 | |
| 1776 | int nfs_wb_page_cancel(struct inode *inode, struct page *page) |
| 1777 | { |
| 1778 | struct nfs_page *req; |
| 1779 | int ret = 0; |
| 1780 | |
| 1781 | for (;;) { |
| 1782 | wait_on_page_writeback(page); |
| 1783 | req = nfs_page_find_request(page); |
| 1784 | if (req == NULL) |
| 1785 | break; |
| 1786 | if (nfs_lock_request(req)) { |
| 1787 | nfs_clear_request_commit(req); |
| 1788 | nfs_inode_remove_request(req); |
| 1789 | /* |
| 1790 | * In case nfs_inode_remove_request has marked the |
| 1791 | * page as being dirty |
| 1792 | */ |
| 1793 | cancel_dirty_page(page, PAGE_CACHE_SIZE); |
| 1794 | nfs_unlock_and_release_request(req); |
| 1795 | break; |
| 1796 | } |
| 1797 | ret = nfs_wait_on_request(req); |
| 1798 | nfs_release_request(req); |
| 1799 | if (ret < 0) |
| 1800 | break; |
| 1801 | } |
| 1802 | return ret; |
| 1803 | } |
| 1804 | |
| 1805 | /* |
| 1806 | * Write back all requests on one page - we do this before reading it. |
| 1807 | */ |
| 1808 | int nfs_wb_page(struct inode *inode, struct page *page) |
| 1809 | { |
| 1810 | loff_t range_start = page_file_offset(page); |
| 1811 | loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1); |
| 1812 | struct writeback_control wbc = { |
| 1813 | .sync_mode = WB_SYNC_ALL, |
| 1814 | .nr_to_write = 0, |
| 1815 | .range_start = range_start, |
| 1816 | .range_end = range_end, |
| 1817 | }; |
| 1818 | int ret; |
| 1819 | |
| 1820 | trace_nfs_writeback_page_enter(inode); |
| 1821 | |
| 1822 | for (;;) { |
| 1823 | wait_on_page_writeback(page); |
| 1824 | if (clear_page_dirty_for_io(page)) { |
| 1825 | ret = nfs_writepage_locked(page, &wbc); |
| 1826 | if (ret < 0) |
| 1827 | goto out_error; |
| 1828 | continue; |
| 1829 | } |
| 1830 | ret = 0; |
| 1831 | if (!PagePrivate(page)) |
| 1832 | break; |
| 1833 | ret = nfs_commit_inode(inode, FLUSH_SYNC); |
| 1834 | if (ret < 0) |
| 1835 | goto out_error; |
| 1836 | } |
| 1837 | out_error: |
| 1838 | trace_nfs_writeback_page_exit(inode, ret); |
| 1839 | return ret; |
| 1840 | } |
| 1841 | |
| 1842 | #ifdef CONFIG_MIGRATION |
| 1843 | int nfs_migrate_page(struct address_space *mapping, struct page *newpage, |
| 1844 | struct page *page, enum migrate_mode mode) |
| 1845 | { |
| 1846 | /* |
| 1847 | * If PagePrivate is set, then the page is currently associated with |
| 1848 | * an in-progress read or write request. Don't try to migrate it. |
| 1849 | * |
| 1850 | * FIXME: we could do this in principle, but we'll need a way to ensure |
| 1851 | * that we can safely release the inode reference while holding |
| 1852 | * the page lock. |
| 1853 | */ |
| 1854 | if (PagePrivate(page)) |
| 1855 | return -EBUSY; |
| 1856 | |
| 1857 | if (!nfs_fscache_release_page(page, GFP_KERNEL)) |
| 1858 | return -EBUSY; |
| 1859 | |
| 1860 | return migrate_page(mapping, newpage, page, mode); |
| 1861 | } |
| 1862 | #endif |
| 1863 | |
| 1864 | int __init nfs_init_writepagecache(void) |
| 1865 | { |
| 1866 | nfs_wdata_cachep = kmem_cache_create("nfs_write_data", |
| 1867 | sizeof(struct nfs_write_header), |
| 1868 | 0, SLAB_HWCACHE_ALIGN, |
| 1869 | NULL); |
| 1870 | if (nfs_wdata_cachep == NULL) |
| 1871 | return -ENOMEM; |
| 1872 | |
| 1873 | nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE, |
| 1874 | nfs_wdata_cachep); |
| 1875 | if (nfs_wdata_mempool == NULL) |
| 1876 | goto out_destroy_write_cache; |
| 1877 | |
| 1878 | nfs_cdata_cachep = kmem_cache_create("nfs_commit_data", |
| 1879 | sizeof(struct nfs_commit_data), |
| 1880 | 0, SLAB_HWCACHE_ALIGN, |
| 1881 | NULL); |
| 1882 | if (nfs_cdata_cachep == NULL) |
| 1883 | goto out_destroy_write_mempool; |
| 1884 | |
| 1885 | nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT, |
| 1886 | nfs_cdata_cachep); |
| 1887 | if (nfs_commit_mempool == NULL) |
| 1888 | goto out_destroy_commit_cache; |
| 1889 | |
| 1890 | /* |
| 1891 | * NFS congestion size, scale with available memory. |
| 1892 | * |
| 1893 | * 64MB: 8192k |
| 1894 | * 128MB: 11585k |
| 1895 | * 256MB: 16384k |
| 1896 | * 512MB: 23170k |
| 1897 | * 1GB: 32768k |
| 1898 | * 2GB: 46340k |
| 1899 | * 4GB: 65536k |
| 1900 | * 8GB: 92681k |
| 1901 | * 16GB: 131072k |
| 1902 | * |
| 1903 | * This allows larger machines to have larger/more transfers. |
| 1904 | * Limit the default to 256M |
| 1905 | */ |
| 1906 | nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10); |
| 1907 | if (nfs_congestion_kb > 256*1024) |
| 1908 | nfs_congestion_kb = 256*1024; |
| 1909 | |
| 1910 | return 0; |
| 1911 | |
| 1912 | out_destroy_commit_cache: |
| 1913 | kmem_cache_destroy(nfs_cdata_cachep); |
| 1914 | out_destroy_write_mempool: |
| 1915 | mempool_destroy(nfs_wdata_mempool); |
| 1916 | out_destroy_write_cache: |
| 1917 | kmem_cache_destroy(nfs_wdata_cachep); |
| 1918 | return -ENOMEM; |
| 1919 | } |
| 1920 | |
| 1921 | void nfs_destroy_writepagecache(void) |
| 1922 | { |
| 1923 | mempool_destroy(nfs_commit_mempool); |
| 1924 | kmem_cache_destroy(nfs_cdata_cachep); |
| 1925 | mempool_destroy(nfs_wdata_mempool); |
| 1926 | kmem_cache_destroy(nfs_wdata_cachep); |
| 1927 | } |
| 1928 | |