fs/nfs/read.c

   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 [email protected]
   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/task_io_accounting_ops.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/sunrpc/clnt.h>
  21 #include <linux/nfs_fs.h>
  22 #include <linux/nfs_page.h>
  23 #include <linux/module.h>
  24
  25 #include "nfs4_fs.h"
  26 #include "internal.h"
  27 #include "iostat.h"
  28 #include "fscache.h"
  29 #include "pnfs.h"
  30 #include "nfstrace.h"
  31
  32 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  33
  34 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops;
  35 static const struct nfs_rw_ops nfs_rw_read_ops;
  36
  37 static struct kmem_cache *nfs_rdata_cachep;
  38
  39 static struct nfs_pgio_header *nfs_readhdr_alloc(void)
  40 {
  41         struct nfs_pgio_header *p = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
  42
  43         if (p)
  44                 p->rw_mode = FMODE_READ;
  45         return p;
  46 }
  47
  48 static void nfs_readhdr_free(struct nfs_pgio_header *rhdr)
  49 {
  50         kmem_cache_free(nfs_rdata_cachep, rhdr);
  51 }
  52
  53 static int nfs_return_empty_folio(struct folio *folio)
  54 {
  55         folio_zero_segment(folio, 0, folio_size(folio));
  56         folio_mark_uptodate(folio);
  57         folio_unlock(folio);
  58         return 0;
  59 }
  60
  61 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
  62                               struct inode *inode, bool force_mds,
  63                               const struct nfs_pgio_completion_ops *compl_ops)
  64 {
  65         struct nfs_server *server = NFS_SERVER(inode);
  66         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
  67
  68 #ifdef CONFIG_NFS_V4_1
  69         if (server->pnfs_curr_ld && !force_mds)
  70                 pg_ops = server->pnfs_curr_ld->pg_read_ops;
  71 #endif
  72         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops,
  73                         server->rsize, 0);
  74 }
  75 EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
  76
  77 static void nfs_pageio_complete_read(struct nfs_pageio_descriptor *pgio)
  78 {
  79         struct nfs_pgio_mirror *pgm;
  80         unsigned long npages;
  81
  82         nfs_pageio_complete(pgio);
  83
  84         /* It doesn't make sense to do mirrored reads! */
  85         WARN_ON_ONCE(pgio->pg_mirror_count != 1);
  86
  87         pgm = &pgio->pg_mirrors[0];
  88         NFS_I(pgio->pg_inode)->read_io += pgm->pg_bytes_written;
  89         npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >> PAGE_SHIFT;
  90         nfs_add_stats(pgio->pg_inode, NFSIOS_READPAGES, npages);
  91 }
  92
  93
  94 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
  95 {
  96         struct nfs_pgio_mirror *mirror;
  97
  98         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
  99                 pgio->pg_ops->pg_cleanup(pgio);
 100
 101         pgio->pg_ops = &nfs_pgio_rw_ops;
 102
 103         /* read path should never have more than one mirror */
 104         WARN_ON_ONCE(pgio->pg_mirror_count != 1);
 105
 106         mirror = &pgio->pg_mirrors[0];
 107         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
 108 }
 109 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
 110
 111 static void nfs_readpage_release(struct nfs_page *req, int error)
 112 {
 113         struct inode *inode = d_inode(nfs_req_openctx(req)->dentry);
 114         struct folio *folio = nfs_page_to_folio(req);
 115
 116         dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb->s_id,
 117                 (unsigned long long)NFS_FILEID(inode), req->wb_bytes,
 118                 (long long)req_offset(req));
 119
 120         if (nfs_error_is_fatal_on_server(error) && error != -ETIMEDOUT)
 121                 folio_set_error(folio);
 122         if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) {
 123                 if (folio_test_uptodate(folio))
 124                         nfs_fscache_write_page(inode, &folio->page);
 125                 folio_unlock(folio);
 126         }
 127         nfs_release_request(req);
 128 }
 129
 130 struct nfs_readdesc {
 131         struct nfs_pageio_descriptor pgio;
 132         struct nfs_open_context *ctx;
 133 };
 134
 135 static void nfs_page_group_set_uptodate(struct nfs_page *req)
 136 {
 137         if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
 138                 folio_mark_uptodate(nfs_page_to_folio(req));
 139 }
 140
 141 static void nfs_read_completion(struct nfs_pgio_header *hdr)
 142 {
 143         unsigned long bytes = 0;
 144         int error;
 145
 146         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 147                 goto out;
 148         while (!list_empty(&hdr->pages)) {
 149                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 150                 struct folio *folio = nfs_page_to_folio(req);
 151                 unsigned long start = req->wb_pgbase;
 152                 unsigned long end = req->wb_pgbase + req->wb_bytes;
 153
 154                 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
 155                         /* note: regions of the page not covered by a
 156                          * request are zeroed in readpage_async_filler */
 157                         if (bytes > hdr->good_bytes) {
 158                                 /* nothing in this request was good, so zero
 159                                  * the full extent of the request */
 160                                 folio_zero_segment(folio, start, end);
 161
 162                         } else if (hdr->good_bytes - bytes < req->wb_bytes) {
 163                                 /* part of this request has good bytes, but
 164                                  * not all. zero the bad bytes */
 165                                 start += hdr->good_bytes - bytes;
 166                                 WARN_ON(start < req->wb_pgbase);
 167                                 folio_zero_segment(folio, start, end);
 168                         }
 169                 }
 170                 error = 0;
 171                 bytes += req->wb_bytes;
 172                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 173                         if (bytes <= hdr->good_bytes)
 174                                 nfs_page_group_set_uptodate(req);
 175                         else {
 176                                 error = hdr->error;
 177                                 xchg(&nfs_req_openctx(req)->error, error);
 178                         }
 179                 } else
 180                         nfs_page_group_set_uptodate(req);
 181                 nfs_list_remove_request(req);
 182                 nfs_readpage_release(req, error);
 183         }
 184 out:
 185         hdr->release(hdr);
 186 }
 187
 188 static void nfs_initiate_read(struct nfs_pgio_header *hdr,
 189                               struct rpc_message *msg,
 190                               const struct nfs_rpc_ops *rpc_ops,
 191                               struct rpc_task_setup *task_setup_data, int how)
 192 {
 193         rpc_ops->read_setup(hdr, msg);
 194         trace_nfs_initiate_read(hdr);
 195 }
 196
 197 static void
 198 nfs_async_read_error(struct list_head *head, int error)
 199 {
 200         struct nfs_page *req;
 201
 202         while (!list_empty(head)) {
 203                 req = nfs_list_entry(head->next);
 204                 nfs_list_remove_request(req);
 205                 nfs_readpage_release(req, error);
 206         }
 207 }
 208
 209 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
 210         .error_cleanup = nfs_async_read_error,
 211         .completion = nfs_read_completion,
 212 };
 213
 214 /*
 215  * This is the callback from RPC telling us whether a reply was
 216  * received or some error occurred (timeout or socket shutdown).
 217  */
 218 static int nfs_readpage_done(struct rpc_task *task,
 219                              struct nfs_pgio_header *hdr,
 220                              struct inode *inode)
 221 {
 222         int status = NFS_PROTO(inode)->read_done(task, hdr);
 223         if (status != 0)
 224                 return status;
 225
 226         nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count);
 227         trace_nfs_readpage_done(task, hdr);
 228
 229         if (task->tk_status == -ESTALE) {
 230                 nfs_set_inode_stale(inode);
 231                 nfs_mark_for_revalidate(inode);
 232         }
 233         return 0;
 234 }
 235
 236 static void nfs_readpage_retry(struct rpc_task *task,
 237                                struct nfs_pgio_header *hdr)
 238 {
 239         struct nfs_pgio_args *argp = &hdr->args;
 240         struct nfs_pgio_res  *resp = &hdr->res;
 241
 242         /* This is a short read! */
 243         nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD);
 244         trace_nfs_readpage_short(task, hdr);
 245
 246         /* Has the server at least made some progress? */
 247         if (resp->count == 0) {
 248                 nfs_set_pgio_error(hdr, -EIO, argp->offset);
 249                 return;
 250         }
 251
 252         /* For non rpc-based layout drivers, retry-through-MDS */
 253         if (!task->tk_ops) {
 254                 hdr->pnfs_error = -EAGAIN;
 255                 return;
 256         }
 257
 258         /* Yes, so retry the read at the end of the hdr */
 259         hdr->mds_offset += resp->count;
 260         argp->offset += resp->count;
 261         argp->pgbase += resp->count;
 262         argp->count -= resp->count;
 263         resp->count = 0;
 264         resp->eof = 0;
 265         rpc_restart_call_prepare(task);
 266 }
 267
 268 static void nfs_readpage_result(struct rpc_task *task,
 269                                 struct nfs_pgio_header *hdr)
 270 {
 271         if (hdr->res.eof) {
 272                 loff_t pos = hdr->args.offset + hdr->res.count;
 273                 unsigned int new = pos - hdr->io_start;
 274
 275                 if (hdr->good_bytes > new) {
 276                         hdr->good_bytes = new;
 277                         set_bit(NFS_IOHDR_EOF, &hdr->flags);
 278                         clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
 279                 }
 280         } else if (hdr->res.count < hdr->args.count)
 281                 nfs_readpage_retry(task, hdr);
 282 }
 283
 284 static int readpage_async_filler(struct nfs_readdesc *desc, struct folio *folio)
 285 {
 286         struct inode *inode = folio_file_mapping(folio)->host;
 287         struct nfs_server *server = NFS_SERVER(inode);
 288         size_t fsize = folio_size(folio);
 289         unsigned int rsize = server->rsize;
 290         struct nfs_page *new;
 291         unsigned int len, aligned_len;
 292         int error;
 293
 294         len = nfs_folio_length(folio);
 295         if (len == 0)
 296                 return nfs_return_empty_folio(folio);
 297
 298         aligned_len = min_t(unsigned int, ALIGN(len, rsize), fsize);
 299
 300         if (!IS_SYNC(inode)) {
 301                 error = nfs_fscache_read_page(inode, &folio->page);
 302                 if (error == 0)
 303                         goto out_unlock;
 304         }
 305
 306         new = nfs_page_create_from_folio(desc->ctx, folio, 0, aligned_len);
 307         if (IS_ERR(new))
 308                 goto out_error;
 309
 310         if (len < fsize)
 311                 folio_zero_segment(folio, len, fsize);
 312         if (!nfs_pageio_add_request(&desc->pgio, new)) {
 313                 nfs_list_remove_request(new);
 314                 error = desc->pgio.pg_error;
 315                 nfs_readpage_release(new, error);
 316                 goto out;
 317         }
 318         return 0;
 319 out_error:
 320         error = PTR_ERR(new);
 321 out_unlock:
 322         folio_unlock(folio);
 323 out:
 324         return error;
 325 }
 326
 327 /*
 328  * Read a page over NFS.
 329  * We read the page synchronously in the following case:
 330  *  -   The error flag is set for this page. This happens only when a
 331  *      previous async read operation failed.
 332  */
 333 int nfs_read_folio(struct file *file, struct folio *folio)
 334 {
 335         struct nfs_readdesc desc;
 336         struct inode *inode = file_inode(file);
 337         int ret;
 338
 339         trace_nfs_aop_readpage(inode, folio);
 340         nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
 341         task_io_account_read(folio_size(folio));
 342
 343         /*
 344          * Try to flush any pending writes to the file..
 345          *
 346          * NOTE! Because we own the folio lock, there cannot
 347          * be any new pending writes generated at this point
 348          * for this folio (other folios can be written to).
 349          */
 350         ret = nfs_wb_folio(inode, folio);
 351         if (ret)
 352                 goto out_unlock;
 353         if (folio_test_uptodate(folio))
 354                 goto out_unlock;
 355
 356         ret = -ESTALE;
 357         if (NFS_STALE(inode))
 358                 goto out_unlock;
 359
 360         desc.ctx = get_nfs_open_context(nfs_file_open_context(file));
 361
 362         xchg(&desc.ctx->error, 0);
 363         nfs_pageio_init_read(&desc.pgio, inode, false,
 364                              &nfs_async_read_completion_ops);
 365
 366         ret = readpage_async_filler(&desc, folio);
 367         if (ret)
 368                 goto out;
 369
 370         nfs_pageio_complete_read(&desc.pgio);
 371         ret = desc.pgio.pg_error < 0 ? desc.pgio.pg_error : 0;
 372         if (!ret) {
 373                 ret = folio_wait_locked_killable(folio);
 374                 if (!folio_test_uptodate(folio) && !ret)
 375                         ret = xchg(&desc.ctx->error, 0);
 376         }
 377 out:
 378         put_nfs_open_context(desc.ctx);
 379         trace_nfs_aop_readpage_done(inode, folio, ret);
 380         return ret;
 381 out_unlock:
 382         folio_unlock(folio);
 383         trace_nfs_aop_readpage_done(inode, folio, ret);
 384         return ret;
 385 }
 386
 387 void nfs_readahead(struct readahead_control *ractl)
 388 {
 389         unsigned int nr_pages = readahead_count(ractl);
 390         struct file *file = ractl->file;
 391         struct nfs_readdesc desc;
 392         struct inode *inode = ractl->mapping->host;
 393         struct folio *folio;
 394         int ret;
 395
 396         trace_nfs_aop_readahead(inode, readahead_pos(ractl), nr_pages);
 397         nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
 398         task_io_account_read(readahead_length(ractl));
 399
 400         ret = -ESTALE;
 401         if (NFS_STALE(inode))
 402                 goto out;
 403
 404         if (file == NULL) {
 405                 ret = -EBADF;
 406                 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 407                 if (desc.ctx == NULL)
 408                         goto out;
 409         } else
 410                 desc.ctx = get_nfs_open_context(nfs_file_open_context(file));
 411
 412         nfs_pageio_init_read(&desc.pgio, inode, false,
 413                              &nfs_async_read_completion_ops);
 414
 415         while ((folio = readahead_folio(ractl)) != NULL) {
 416                 ret = readpage_async_filler(&desc, folio);
 417                 if (ret)
 418                         break;
 419         }
 420
 421         nfs_pageio_complete_read(&desc.pgio);
 422
 423         put_nfs_open_context(desc.ctx);
 424 out:
 425         trace_nfs_aop_readahead_done(inode, nr_pages, ret);
 426 }
 427
 428 int __init nfs_init_readpagecache(void)
 429 {
 430         nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
 431                                              sizeof(struct nfs_pgio_header),
 432                                              0, SLAB_HWCACHE_ALIGN,
 433                                              NULL);
 434         if (nfs_rdata_cachep == NULL)
 435                 return -ENOMEM;
 436
 437         return 0;
 438 }
 439
 440 void nfs_destroy_readpagecache(void)
 441 {
 442         kmem_cache_destroy(nfs_rdata_cachep);
 443 }
 444
 445 static const struct nfs_rw_ops nfs_rw_read_ops = {
 446         .rw_alloc_header        = nfs_readhdr_alloc,
 447         .rw_free_header         = nfs_readhdr_free,
 448         .rw_done                = nfs_readpage_done,
 449         .rw_result              = nfs_readpage_result,
 450         .rw_initiate            = nfs_initiate_read,
 451 };