1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * linux/fs/nfs/read.c 4 * 5 * Block I/O for NFS 6 * 7 * Partial copy of Linus' read cache modifications to fs/nfs/file.c 8 * modified for async RPC by okir@monad.swb.de 9 */ 10 11 #include <linux/time.h> 12 #include <linux/kernel.h> 13 #include <linux/errno.h> 14 #include <linux/fcntl.h> 15 #include <linux/stat.h> 16 #include <linux/mm.h> 17 #include <linux/slab.h> 18 #include <linux/pagemap.h> 19 #include <linux/sunrpc/clnt.h> 20 #include <linux/nfs_fs.h> 21 #include <linux/nfs_page.h> 22 #include <linux/module.h> 23 24 #include "nfs4_fs.h" 25 #include "internal.h" 26 #include "iostat.h" 27 #include "fscache.h" 28 #include "pnfs.h" 29 #include "nfstrace.h" 30 31 #define NFSDBG_FACILITY NFSDBG_PAGECACHE 32 33 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops; 34 static const struct nfs_rw_ops nfs_rw_read_ops; 35 36 static struct kmem_cache *nfs_rdata_cachep; 37 38 static struct nfs_pgio_header *nfs_readhdr_alloc(void) 39 { 40 struct nfs_pgio_header *p = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL); 41 42 if (p) 43 p->rw_mode = FMODE_READ; 44 return p; 45 } 46 47 static void nfs_readhdr_free(struct nfs_pgio_header *rhdr) 48 { 49 kmem_cache_free(nfs_rdata_cachep, rhdr); 50 } 51 52 static int nfs_return_empty_folio(struct folio *folio) 53 { 54 folio_zero_segment(folio, 0, folio_size(folio)); 55 folio_mark_uptodate(folio); 56 folio_unlock(folio); 57 return 0; 58 } 59 60 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, 61 struct inode *inode, bool force_mds, 62 const struct nfs_pgio_completion_ops *compl_ops) 63 { 64 struct nfs_server *server = NFS_SERVER(inode); 65 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops; 66 67 #ifdef CONFIG_NFS_V4_1 68 if (server->pnfs_curr_ld && !force_mds) 69 pg_ops = server->pnfs_curr_ld->pg_read_ops; 70 #endif 71 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops, 72 server->rsize, 0); 73 } 74 EXPORT_SYMBOL_GPL(nfs_pageio_init_read); 75 76 static void nfs_pageio_complete_read(struct nfs_pageio_descriptor *pgio) 77 { 78 struct nfs_pgio_mirror *pgm; 79 unsigned long npages; 80 81 nfs_pageio_complete(pgio); 82 83 /* It doesn't make sense to do mirrored reads! */ 84 WARN_ON_ONCE(pgio->pg_mirror_count != 1); 85 86 pgm = &pgio->pg_mirrors[0]; 87 NFS_I(pgio->pg_inode)->read_io += pgm->pg_bytes_written; 88 npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >> PAGE_SHIFT; 89 nfs_add_stats(pgio->pg_inode, NFSIOS_READPAGES, npages); 90 } 91 92 93 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio) 94 { 95 struct nfs_pgio_mirror *mirror; 96 97 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup) 98 pgio->pg_ops->pg_cleanup(pgio); 99 100 pgio->pg_ops = &nfs_pgio_rw_ops; 101 102 /* read path should never have more than one mirror */ 103 WARN_ON_ONCE(pgio->pg_mirror_count != 1); 104 105 mirror = &pgio->pg_mirrors[0]; 106 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize; 107 } 108 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds); 109 110 static void nfs_readpage_release(struct nfs_page *req, int error) 111 { 112 struct inode *inode = d_inode(nfs_req_openctx(req)->dentry); 113 struct folio *folio = nfs_page_to_folio(req); 114 115 dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb->s_id, 116 (unsigned long long)NFS_FILEID(inode), req->wb_bytes, 117 (long long)req_offset(req)); 118 119 if (nfs_error_is_fatal_on_server(error) && error != -ETIMEDOUT) 120 folio_set_error(folio); 121 if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) { 122 if (folio_test_uptodate(folio)) 123 nfs_fscache_write_page(inode, &folio->page); 124 folio_unlock(folio); 125 } 126 nfs_release_request(req); 127 } 128 129 struct nfs_readdesc { 130 struct nfs_pageio_descriptor pgio; 131 struct nfs_open_context *ctx; 132 }; 133 134 static void nfs_page_group_set_uptodate(struct nfs_page *req) 135 { 136 if (nfs_page_group_sync_on_bit(req, PG_UPTODATE)) 137 folio_mark_uptodate(nfs_page_to_folio(req)); 138 } 139 140 static void nfs_read_completion(struct nfs_pgio_header *hdr) 141 { 142 unsigned long bytes = 0; 143 int error; 144 145 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) 146 goto out; 147 while (!list_empty(&hdr->pages)) { 148 struct nfs_page *req = nfs_list_entry(hdr->pages.next); 149 struct folio *folio = nfs_page_to_folio(req); 150 unsigned long start = req->wb_pgbase; 151 unsigned long end = req->wb_pgbase + req->wb_bytes; 152 153 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) { 154 /* note: regions of the page not covered by a 155 * request are zeroed in readpage_async_filler */ 156 if (bytes > hdr->good_bytes) { 157 /* nothing in this request was good, so zero 158 * the full extent of the request */ 159 folio_zero_segment(folio, start, end); 160 161 } else if (hdr->good_bytes - bytes < req->wb_bytes) { 162 /* part of this request has good bytes, but 163 * not all. zero the bad bytes */ 164 start += hdr->good_bytes - bytes; 165 WARN_ON(start < req->wb_pgbase); 166 folio_zero_segment(folio, start, end); 167 } 168 } 169 error = 0; 170 bytes += req->wb_bytes; 171 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) { 172 if (bytes <= hdr->good_bytes) 173 nfs_page_group_set_uptodate(req); 174 else { 175 error = hdr->error; 176 xchg(&nfs_req_openctx(req)->error, error); 177 } 178 } else 179 nfs_page_group_set_uptodate(req); 180 nfs_list_remove_request(req); 181 nfs_readpage_release(req, error); 182 } 183 out: 184 hdr->release(hdr); 185 } 186 187 static void nfs_initiate_read(struct nfs_pgio_header *hdr, 188 struct rpc_message *msg, 189 const struct nfs_rpc_ops *rpc_ops, 190 struct rpc_task_setup *task_setup_data, int how) 191 { 192 rpc_ops->read_setup(hdr, msg); 193 trace_nfs_initiate_read(hdr); 194 } 195 196 static void 197 nfs_async_read_error(struct list_head *head, int error) 198 { 199 struct nfs_page *req; 200 201 while (!list_empty(head)) { 202 req = nfs_list_entry(head->next); 203 nfs_list_remove_request(req); 204 nfs_readpage_release(req, error); 205 } 206 } 207 208 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = { 209 .error_cleanup = nfs_async_read_error, 210 .completion = nfs_read_completion, 211 }; 212 213 /* 214 * This is the callback from RPC telling us whether a reply was 215 * received or some error occurred (timeout or socket shutdown). 216 */ 217 static int nfs_readpage_done(struct rpc_task *task, 218 struct nfs_pgio_header *hdr, 219 struct inode *inode) 220 { 221 int status = NFS_PROTO(inode)->read_done(task, hdr); 222 if (status != 0) 223 return status; 224 225 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count); 226 trace_nfs_readpage_done(task, hdr); 227 228 if (task->tk_status == -ESTALE) { 229 nfs_set_inode_stale(inode); 230 nfs_mark_for_revalidate(inode); 231 } 232 return 0; 233 } 234 235 static void nfs_readpage_retry(struct rpc_task *task, 236 struct nfs_pgio_header *hdr) 237 { 238 struct nfs_pgio_args *argp = &hdr->args; 239 struct nfs_pgio_res *resp = &hdr->res; 240 241 /* This is a short read! */ 242 nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD); 243 trace_nfs_readpage_short(task, hdr); 244 245 /* Has the server at least made some progress? */ 246 if (resp->count == 0) { 247 nfs_set_pgio_error(hdr, -EIO, argp->offset); 248 return; 249 } 250 251 /* For non rpc-based layout drivers, retry-through-MDS */ 252 if (!task->tk_ops) { 253 hdr->pnfs_error = -EAGAIN; 254 return; 255 } 256 257 /* Yes, so retry the read at the end of the hdr */ 258 hdr->mds_offset += resp->count; 259 argp->offset += resp->count; 260 argp->pgbase += resp->count; 261 argp->count -= resp->count; 262 resp->count = 0; 263 resp->eof = 0; 264 rpc_restart_call_prepare(task); 265 } 266 267 static void nfs_readpage_result(struct rpc_task *task, 268 struct nfs_pgio_header *hdr) 269 { 270 if (hdr->res.eof) { 271 loff_t pos = hdr->args.offset + hdr->res.count; 272 unsigned int new = pos - hdr->io_start; 273 274 if (hdr->good_bytes > new) { 275 hdr->good_bytes = new; 276 set_bit(NFS_IOHDR_EOF, &hdr->flags); 277 clear_bit(NFS_IOHDR_ERROR, &hdr->flags); 278 } 279 } else if (hdr->res.count < hdr->args.count) 280 nfs_readpage_retry(task, hdr); 281 } 282 283 static int readpage_async_filler(struct nfs_readdesc *desc, struct folio *folio) 284 { 285 struct inode *inode = folio_file_mapping(folio)->host; 286 struct nfs_server *server = NFS_SERVER(inode); 287 size_t fsize = folio_size(folio); 288 unsigned int rsize = server->rsize; 289 struct nfs_page *new; 290 unsigned int len, aligned_len; 291 int error; 292 293 len = nfs_folio_length(folio); 294 if (len == 0) 295 return nfs_return_empty_folio(folio); 296 297 aligned_len = min_t(unsigned int, ALIGN(len, rsize), fsize); 298 299 if (!IS_SYNC(inode)) { 300 error = nfs_fscache_read_page(inode, &folio->page); 301 if (error == 0) 302 goto out_unlock; 303 } 304 305 new = nfs_page_create_from_folio(desc->ctx, folio, 0, aligned_len); 306 if (IS_ERR(new)) 307 goto out_error; 308 309 if (len < fsize) 310 folio_zero_segment(folio, len, fsize); 311 if (!nfs_pageio_add_request(&desc->pgio, new)) { 312 nfs_list_remove_request(new); 313 error = desc->pgio.pg_error; 314 nfs_readpage_release(new, error); 315 goto out; 316 } 317 return 0; 318 out_error: 319 error = PTR_ERR(new); 320 out_unlock: 321 folio_unlock(folio); 322 out: 323 return error; 324 } 325 326 /* 327 * Read a page over NFS. 328 * We read the page synchronously in the following case: 329 * - The error flag is set for this page. This happens only when a 330 * previous async read operation failed. 331 */ 332 int nfs_read_folio(struct file *file, struct folio *folio) 333 { 334 struct nfs_readdesc desc; 335 struct inode *inode = file_inode(file); 336 int ret; 337 338 trace_nfs_aop_readpage(inode, folio); 339 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE); 340 341 /* 342 * Try to flush any pending writes to the file.. 343 * 344 * NOTE! Because we own the folio lock, there cannot 345 * be any new pending writes generated at this point 346 * for this folio (other folios can be written to). 347 */ 348 ret = nfs_wb_folio(inode, folio); 349 if (ret) 350 goto out_unlock; 351 if (folio_test_uptodate(folio)) 352 goto out_unlock; 353 354 ret = -ESTALE; 355 if (NFS_STALE(inode)) 356 goto out_unlock; 357 358 desc.ctx = get_nfs_open_context(nfs_file_open_context(file)); 359 360 xchg(&desc.ctx->error, 0); 361 nfs_pageio_init_read(&desc.pgio, inode, false, 362 &nfs_async_read_completion_ops); 363 364 ret = readpage_async_filler(&desc, folio); 365 if (ret) 366 goto out; 367 368 nfs_pageio_complete_read(&desc.pgio); 369 ret = desc.pgio.pg_error < 0 ? desc.pgio.pg_error : 0; 370 if (!ret) { 371 ret = folio_wait_locked_killable(folio); 372 if (!folio_test_uptodate(folio) && !ret) 373 ret = xchg(&desc.ctx->error, 0); 374 } 375 out: 376 put_nfs_open_context(desc.ctx); 377 trace_nfs_aop_readpage_done(inode, folio, ret); 378 return ret; 379 out_unlock: 380 folio_unlock(folio); 381 trace_nfs_aop_readpage_done(inode, folio, ret); 382 return ret; 383 } 384 385 void nfs_readahead(struct readahead_control *ractl) 386 { 387 unsigned int nr_pages = readahead_count(ractl); 388 struct file *file = ractl->file; 389 struct nfs_readdesc desc; 390 struct inode *inode = ractl->mapping->host; 391 struct folio *folio; 392 int ret; 393 394 trace_nfs_aop_readahead(inode, readahead_pos(ractl), nr_pages); 395 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES); 396 397 ret = -ESTALE; 398 if (NFS_STALE(inode)) 399 goto out; 400 401 if (file == NULL) { 402 ret = -EBADF; 403 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ); 404 if (desc.ctx == NULL) 405 goto out; 406 } else 407 desc.ctx = get_nfs_open_context(nfs_file_open_context(file)); 408 409 nfs_pageio_init_read(&desc.pgio, inode, false, 410 &nfs_async_read_completion_ops); 411 412 while ((folio = readahead_folio(ractl)) != NULL) { 413 ret = readpage_async_filler(&desc, folio); 414 if (ret) 415 break; 416 } 417 418 nfs_pageio_complete_read(&desc.pgio); 419 420 put_nfs_open_context(desc.ctx); 421 out: 422 trace_nfs_aop_readahead_done(inode, nr_pages, ret); 423 } 424 425 int __init nfs_init_readpagecache(void) 426 { 427 nfs_rdata_cachep = kmem_cache_create("nfs_read_data", 428 sizeof(struct nfs_pgio_header), 429 0, SLAB_HWCACHE_ALIGN, 430 NULL); 431 if (nfs_rdata_cachep == NULL) 432 return -ENOMEM; 433 434 return 0; 435 } 436 437 void nfs_destroy_readpagecache(void) 438 { 439 kmem_cache_destroy(nfs_rdata_cachep); 440 } 441 442 static const struct nfs_rw_ops nfs_rw_read_ops = { 443 .rw_alloc_header = nfs_readhdr_alloc, 444 .rw_free_header = nfs_readhdr_free, 445 .rw_done = nfs_readpage_done, 446 .rw_result = nfs_readpage_result, 447 .rw_initiate = nfs_initiate_read, 448 }; 449