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 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/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
  54 int nfs_return_empty_page(struct page *page)
  55 {
  56         zero_user(page, 0, PAGE_SIZE);
  57         SetPageUptodate(page);
  58         unlock_page(page);
  59         return 0;
  60 }
  61
  62 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
  63                               struct inode *inode, bool force_mds,
  64                               const struct nfs_pgio_completion_ops *compl_ops)
  65 {
  66         struct nfs_server *server = NFS_SERVER(inode);
  67         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
  68
  69 #ifdef CONFIG_NFS_V4_1
  70         if (server->pnfs_curr_ld && !force_mds)
  71                 pg_ops = server->pnfs_curr_ld->pg_read_ops;
  72 #endif
  73         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops,
  74                         server->rsize, 0);
  75 }
  76 EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
  77
  78 static void nfs_pageio_complete_read(struct nfs_pageio_descriptor *pgio)
  79 {
  80         struct nfs_pgio_mirror *pgm;
  81         unsigned long npages;
  82
  83         nfs_pageio_complete(pgio);
  84
  85         /* It doesn't make sense to do mirrored reads! */
  86         WARN_ON_ONCE(pgio->pg_mirror_count != 1);
  87
  88         pgm = &pgio->pg_mirrors[0];
  89         NFS_I(pgio->pg_inode)->read_io += pgm->pg_bytes_written;
  90         npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >> PAGE_SHIFT;
  91         nfs_add_stats(pgio->pg_inode, NFSIOS_READPAGES, npages);
  92 }
  93
  94
  95 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
  96 {
  97         struct nfs_pgio_mirror *mirror;
  98
  99         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
 100                 pgio->pg_ops->pg_cleanup(pgio);
 101
 102         pgio->pg_ops = &nfs_pgio_rw_ops;
 103
 104         /* read path should never have more than one mirror */
 105         WARN_ON_ONCE(pgio->pg_mirror_count != 1);
 106
 107         mirror = &pgio->pg_mirrors[0];
 108         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
 109 }
 110 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
 111
 112 static void nfs_readpage_release(struct nfs_page *req, int error)
 113 {
 114         struct inode *inode = d_inode(nfs_req_openctx(req)->dentry);
 115         struct page *page = req->wb_page;
 116
 117         dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb->s_id,
 118                 (unsigned long long)NFS_FILEID(inode), req->wb_bytes,
 119                 (long long)req_offset(req));
 120
 121         if (nfs_error_is_fatal_on_server(error) && error != -ETIMEDOUT)
 122                 SetPageError(page);
 123         if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) {
 124                 if (PageUptodate(page))
 125                         nfs_fscache_write_page(inode, page);
 126                 unlock_page(page);
 127         }
 128         nfs_release_request(req);
 129 }
 130
 131 struct nfs_readdesc {
 132         struct nfs_pageio_descriptor pgio;
 133         struct nfs_open_context *ctx;
 134 };
 135
 136 static void nfs_page_group_set_uptodate(struct nfs_page *req)
 137 {
 138         if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
 139                 SetPageUptodate(req->wb_page);
 140 }
 141
 142 static void nfs_read_completion(struct nfs_pgio_header *hdr)
 143 {
 144         unsigned long bytes = 0;
 145         int error;
 146
 147         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 148                 goto out;
 149         while (!list_empty(&hdr->pages)) {
 150                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 151                 struct page *page = req->wb_page;
 152                 unsigned long start = req->wb_pgbase;
 153                 unsigned long end = req->wb_pgbase + req->wb_bytes;
 154
 155                 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
 156                         /* note: regions of the page not covered by a
 157                          * request are zeroed in readpage_async_filler */
 158                         if (bytes > hdr->good_bytes) {
 159                                 /* nothing in this request was good, so zero
 160                                  * the full extent of the request */
 161                                 zero_user_segment(page, start, end);
 162
 163                         } else if (hdr->good_bytes - bytes < req->wb_bytes) {
 164                                 /* part of this request has good bytes, but
 165                                  * not all. zero the bad bytes */
 166                                 start += hdr->good_bytes - bytes;
 167                                 WARN_ON(start < req->wb_pgbase);
 168                                 zero_user_segment(page, start, end);
 169                         }
 170                 }
 171                 error = 0;
 172                 bytes += req->wb_bytes;
 173                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 174                         if (bytes <= hdr->good_bytes)
 175                                 nfs_page_group_set_uptodate(req);
 176                         else {
 177                                 error = hdr->error;
 178                                 xchg(&nfs_req_openctx(req)->error, error);
 179                         }
 180                 } else
 181                         nfs_page_group_set_uptodate(req);
 182                 nfs_list_remove_request(req);
 183                 nfs_readpage_release(req, error);
 184         }
 185 out:
 186         hdr->release(hdr);
 187 }
 188
 189 static void nfs_initiate_read(struct nfs_pgio_header *hdr,
 190                               struct rpc_message *msg,
 191                               const struct nfs_rpc_ops *rpc_ops,
 192                               struct rpc_task_setup *task_setup_data, int how)
 193 {
 194         rpc_ops->read_setup(hdr, msg);
 195         trace_nfs_initiate_read(hdr);
 196 }
 197
 198 static void
 199 nfs_async_read_error(struct list_head *head, int error)
 200 {
 201         struct nfs_page *req;
 202
 203         while (!list_empty(head)) {
 204                 req = nfs_list_entry(head->next);
 205                 nfs_list_remove_request(req);
 206                 nfs_readpage_release(req, error);
 207         }
 208 }
 209
 210 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
 211         .error_cleanup = nfs_async_read_error,
 212         .completion = nfs_read_completion,
 213 };
 214
 215 /*
 216  * This is the callback from RPC telling us whether a reply was
 217  * received or some error occurred (timeout or socket shutdown).
 218  */
 219 static int nfs_readpage_done(struct rpc_task *task,
 220                              struct nfs_pgio_header *hdr,
 221                              struct inode *inode)
 222 {
 223         int status = NFS_PROTO(inode)->read_done(task, hdr);
 224         if (status != 0)
 225                 return status;
 226
 227         nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count);
 228         trace_nfs_readpage_done(task, hdr);
 229
 230         if (task->tk_status == -ESTALE) {
 231                 nfs_set_inode_stale(inode);
 232                 nfs_mark_for_revalidate(inode);
 233         }
 234         return 0;
 235 }
 236
 237 static void nfs_readpage_retry(struct rpc_task *task,
 238                                struct nfs_pgio_header *hdr)
 239 {
 240         struct nfs_pgio_args *argp = &hdr->args;
 241         struct nfs_pgio_res  *resp = &hdr->res;
 242
 243         /* This is a short read! */
 244         nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD);
 245         trace_nfs_readpage_short(task, hdr);
 246
 247         /* Has the server at least made some progress? */
 248         if (resp->count == 0) {
 249                 nfs_set_pgio_error(hdr, -EIO, argp->offset);
 250                 return;
 251         }
 252
 253         /* For non rpc-based layout drivers, retry-through-MDS */
 254         if (!task->tk_ops) {
 255                 hdr->pnfs_error = -EAGAIN;
 256                 return;
 257         }
 258
 259         /* Yes, so retry the read at the end of the hdr */
 260         hdr->mds_offset += resp->count;
 261         argp->offset += resp->count;
 262         argp->pgbase += resp->count;
 263         argp->count -= resp->count;
 264         resp->count = 0;
 265         resp->eof = 0;
 266         rpc_restart_call_prepare(task);
 267 }
 268
 269 static void nfs_readpage_result(struct rpc_task *task,
 270                                 struct nfs_pgio_header *hdr)
 271 {
 272         if (hdr->res.eof) {
 273                 loff_t pos = hdr->args.offset + hdr->res.count;
 274                 unsigned int new = pos - hdr->io_start;
 275
 276                 if (hdr->good_bytes > new) {
 277                         hdr->good_bytes = new;
 278                         set_bit(NFS_IOHDR_EOF, &hdr->flags);
 279                         clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
 280                 }
 281         } else if (hdr->res.count < hdr->args.count)
 282                 nfs_readpage_retry(task, hdr);
 283 }
 284
 285 static int
 286 readpage_async_filler(struct nfs_readdesc *desc, struct page *page)
 287 {
 288         struct inode *inode = page_file_mapping(page)->host;
 289         unsigned int rsize = NFS_SERVER(inode)->rsize;
 290         struct nfs_page *new;
 291         unsigned int len, aligned_len;
 292         int error;
 293
 294         len = nfs_page_length(page);
 295         if (len == 0)
 296                 return nfs_return_empty_page(page);
 297
 298         aligned_len = min_t(unsigned int, ALIGN(len, rsize), PAGE_SIZE);
 299
 300         if (!IS_SYNC(page->mapping->host)) {
 301                 error = nfs_fscache_read_page(page->mapping->host, page);
 302                 if (error == 0)
 303                         goto out_unlock;
 304         }
 305
 306         new = nfs_create_request(desc->ctx, page, 0, aligned_len);
 307         if (IS_ERR(new))
 308                 goto out_error;
 309
 310         if (len < PAGE_SIZE)
 311                 zero_user_segment(page, len, PAGE_SIZE);
 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         unlock_page(page);
 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 page *page = &folio->page;
 336         struct nfs_readdesc desc;
 337         struct inode *inode = page_file_mapping(page)->host;
 338         int ret;
 339
 340         trace_nfs_aop_readpage(inode, page);
 341         nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
 342         task_io_account_read(folio_size(folio));
 343
 344         /*
 345          * Try to flush any pending writes to the file..
 346          *
 347          * NOTE! Because we own the page lock, there cannot
 348          * be any new pending writes generated at this point
 349          * for this page (other pages can be written to).
 350          */
 351         ret = nfs_wb_page(inode, page);
 352         if (ret)
 353                 goto out_unlock;
 354         if (PageUptodate(page))
 355                 goto out_unlock;
 356
 357         ret = -ESTALE;
 358         if (NFS_STALE(inode))
 359                 goto out_unlock;
 360
 361         if (file == NULL) {
 362                 ret = -EBADF;
 363                 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 364                 if (desc.ctx == NULL)
 365                         goto out_unlock;
 366         } else
 367                 desc.ctx = get_nfs_open_context(nfs_file_open_context(file));
 368
 369         xchg(&desc.ctx->error, 0);
 370         nfs_pageio_init_read(&desc.pgio, inode, false,
 371                              &nfs_async_read_completion_ops);
 372
 373         ret = readpage_async_filler(&desc, page);
 374         if (ret)
 375                 goto out;
 376
 377         nfs_pageio_complete_read(&desc.pgio);
 378         ret = desc.pgio.pg_error < 0 ? desc.pgio.pg_error : 0;
 379         if (!ret) {
 380                 ret = wait_on_page_locked_killable(page);
 381                 if (!PageUptodate(page) && !ret)
 382                         ret = xchg(&desc.ctx->error, 0);
 383         }
 384 out:
 385         put_nfs_open_context(desc.ctx);
 386         trace_nfs_aop_readpage_done(inode, page, ret);
 387         return ret;
 388 out_unlock:
 389         unlock_page(page);
 390         trace_nfs_aop_readpage_done(inode, page, ret);
 391         return ret;
 392 }
 393
 394 void nfs_readahead(struct readahead_control *ractl)
 395 {
 396         unsigned int nr_pages = readahead_count(ractl);
 397         struct file *file = ractl->file;
 398         struct nfs_readdesc desc;
 399         struct inode *inode = ractl->mapping->host;
 400         struct page *page;
 401         int ret;
 402
 403         trace_nfs_aop_readahead(inode, readahead_pos(ractl), nr_pages);
 404         nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
 405         task_io_account_read(readahead_length(ractl));
 406
 407         ret = -ESTALE;
 408         if (NFS_STALE(inode))
 409                 goto out;
 410
 411         if (file == NULL) {
 412                 ret = -EBADF;
 413                 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 414                 if (desc.ctx == NULL)
 415                         goto out;
 416         } else
 417                 desc.ctx = get_nfs_open_context(nfs_file_open_context(file));
 418
 419         nfs_pageio_init_read(&desc.pgio, inode, false,
 420                              &nfs_async_read_completion_ops);
 421
 422         while ((page = readahead_page(ractl)) != NULL) {
 423                 ret = readpage_async_filler(&desc, page);
 424                 put_page(page);
 425                 if (ret)
 426                         break;
 427         }
 428
 429         nfs_pageio_complete_read(&desc.pgio);
 430
 431         put_nfs_open_context(desc.ctx);
 432 out:
 433         trace_nfs_aop_readahead_done(inode, nr_pages, ret);
 434 }
 435
 436 int __init nfs_init_readpagecache(void)
 437 {
 438         nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
 439                                              sizeof(struct nfs_pgio_header),
 440                                              0, SLAB_HWCACHE_ALIGN,
 441                                              NULL);
 442         if (nfs_rdata_cachep == NULL)
 443                 return -ENOMEM;
 444
 445         return 0;
 446 }
 447
 448 void nfs_destroy_readpagecache(void)
 449 {
 450         kmem_cache_destroy(nfs_rdata_cachep);
 451 }
 452
 453 static const struct nfs_rw_ops nfs_rw_read_ops = {
 454         .rw_alloc_header        = nfs_readhdr_alloc,
 455         .rw_free_header         = nfs_readhdr_free,
 456         .rw_done                = nfs_readpage_done,
 457         .rw_result              = nfs_readpage_result,
 458         .rw_initiate            = nfs_initiate_read,
 459 };