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         if (rhdr->res.scratch != NULL)
  51                 kfree(rhdr->res.scratch);
  52         kmem_cache_free(nfs_rdata_cachep, rhdr);
  53 }
  54
  55 static
  56 int nfs_return_empty_page(struct page *page)
  57 {
  58         zero_user(page, 0, PAGE_SIZE);
  59         SetPageUptodate(page);
  60         unlock_page(page);
  61         return 0;
  62 }
  63
  64 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
  65                               struct inode *inode, bool force_mds,
  66                               const struct nfs_pgio_completion_ops *compl_ops)
  67 {
  68         struct nfs_server *server = NFS_SERVER(inode);
  69         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
  70
  71 #ifdef CONFIG_NFS_V4_1
  72         if (server->pnfs_curr_ld && !force_mds)
  73                 pg_ops = server->pnfs_curr_ld->pg_read_ops;
  74 #endif
  75         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops,
  76                         server->rsize, 0);
  77 }
  78 EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
  79
  80 static void nfs_pageio_complete_read(struct nfs_pageio_descriptor *pgio)
  81 {
  82         struct nfs_pgio_mirror *pgm;
  83         unsigned long npages;
  84
  85         nfs_pageio_complete(pgio);
  86
  87         /* It doesn't make sense to do mirrored reads! */
  88         WARN_ON_ONCE(pgio->pg_mirror_count != 1);
  89
  90         pgm = &pgio->pg_mirrors[0];
  91         NFS_I(pgio->pg_inode)->read_io += pgm->pg_bytes_written;
  92         npages = (pgm->pg_bytes_written + PAGE_SIZE - 1) >> PAGE_SHIFT;
  93         nfs_add_stats(pgio->pg_inode, NFSIOS_READPAGES, npages);
  94 }
  95
  96
  97 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
  98 {
  99         struct nfs_pgio_mirror *mirror;
 100
 101         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
 102                 pgio->pg_ops->pg_cleanup(pgio);
 103
 104         pgio->pg_ops = &nfs_pgio_rw_ops;
 105
 106         /* read path should never have more than one mirror */
 107         WARN_ON_ONCE(pgio->pg_mirror_count != 1);
 108
 109         mirror = &pgio->pg_mirrors[0];
 110         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
 111 }
 112 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
 113
 114 bool nfs_read_alloc_scratch(struct nfs_pgio_header *hdr, size_t size)
 115 {
 116         WARN_ON(hdr->res.scratch != NULL);
 117         hdr->res.scratch = kmalloc(size, GFP_KERNEL);
 118         return hdr->res.scratch != NULL;
 119 }
 120 EXPORT_SYMBOL_GPL(nfs_read_alloc_scratch);
 121
 122 static void nfs_readpage_release(struct nfs_page *req, int error)
 123 {
 124         struct inode *inode = d_inode(nfs_req_openctx(req)->dentry);
 125         struct page *page = req->wb_page;
 126
 127         dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb->s_id,
 128                 (unsigned long long)NFS_FILEID(inode), req->wb_bytes,
 129                 (long long)req_offset(req));
 130
 131         if (nfs_error_is_fatal_on_server(error) && error != -ETIMEDOUT)
 132                 SetPageError(page);
 133         if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) {
 134                 if (PageUptodate(page))
 135                         nfs_fscache_write_page(inode, page);
 136                 unlock_page(page);
 137         }
 138         nfs_release_request(req);
 139 }
 140
 141 struct nfs_readdesc {
 142         struct nfs_pageio_descriptor pgio;
 143         struct nfs_open_context *ctx;
 144 };
 145
 146 static void nfs_page_group_set_uptodate(struct nfs_page *req)
 147 {
 148         if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
 149                 SetPageUptodate(req->wb_page);
 150 }
 151
 152 static void nfs_read_completion(struct nfs_pgio_header *hdr)
 153 {
 154         unsigned long bytes = 0;
 155         int error;
 156
 157         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 158                 goto out;
 159         while (!list_empty(&hdr->pages)) {
 160                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 161                 struct page *page = req->wb_page;
 162                 unsigned long start = req->wb_pgbase;
 163                 unsigned long end = req->wb_pgbase + req->wb_bytes;
 164
 165                 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
 166                         /* note: regions of the page not covered by a
 167                          * request are zeroed in readpage_async_filler */
 168                         if (bytes > hdr->good_bytes) {
 169                                 /* nothing in this request was good, so zero
 170                                  * the full extent of the request */
 171                                 zero_user_segment(page, start, end);
 172
 173                         } else if (hdr->good_bytes - bytes < req->wb_bytes) {
 174                                 /* part of this request has good bytes, but
 175                                  * not all. zero the bad bytes */
 176                                 start += hdr->good_bytes - bytes;
 177                                 WARN_ON(start < req->wb_pgbase);
 178                                 zero_user_segment(page, start, end);
 179                         }
 180                 }
 181                 error = 0;
 182                 bytes += req->wb_bytes;
 183                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 184                         if (bytes <= hdr->good_bytes)
 185                                 nfs_page_group_set_uptodate(req);
 186                         else {
 187                                 error = hdr->error;
 188                                 xchg(&nfs_req_openctx(req)->error, error);
 189                         }
 190                 } else
 191                         nfs_page_group_set_uptodate(req);
 192                 nfs_list_remove_request(req);
 193                 nfs_readpage_release(req, error);
 194         }
 195 out:
 196         hdr->release(hdr);
 197 }
 198
 199 static void nfs_initiate_read(struct nfs_pgio_header *hdr,
 200                               struct rpc_message *msg,
 201                               const struct nfs_rpc_ops *rpc_ops,
 202                               struct rpc_task_setup *task_setup_data, int how)
 203 {
 204         rpc_ops->read_setup(hdr, msg);
 205         trace_nfs_initiate_read(hdr);
 206 }
 207
 208 static void
 209 nfs_async_read_error(struct list_head *head, int error)
 210 {
 211         struct nfs_page *req;
 212
 213         while (!list_empty(head)) {
 214                 req = nfs_list_entry(head->next);
 215                 nfs_list_remove_request(req);
 216                 nfs_readpage_release(req, error);
 217         }
 218 }
 219
 220 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
 221         .error_cleanup = nfs_async_read_error,
 222         .completion = nfs_read_completion,
 223 };
 224
 225 /*
 226  * This is the callback from RPC telling us whether a reply was
 227  * received or some error occurred (timeout or socket shutdown).
 228  */
 229 static int nfs_readpage_done(struct rpc_task *task,
 230                              struct nfs_pgio_header *hdr,
 231                              struct inode *inode)
 232 {
 233         int status = NFS_PROTO(inode)->read_done(task, hdr);
 234         if (status != 0)
 235                 return status;
 236
 237         nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count);
 238         trace_nfs_readpage_done(task, hdr);
 239
 240         if (task->tk_status == -ESTALE) {
 241                 nfs_set_inode_stale(inode);
 242                 nfs_mark_for_revalidate(inode);
 243         }
 244         return 0;
 245 }
 246
 247 static void nfs_readpage_retry(struct rpc_task *task,
 248                                struct nfs_pgio_header *hdr)
 249 {
 250         struct nfs_pgio_args *argp = &hdr->args;
 251         struct nfs_pgio_res  *resp = &hdr->res;
 252
 253         /* This is a short read! */
 254         nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD);
 255         trace_nfs_readpage_short(task, hdr);
 256
 257         /* Has the server at least made some progress? */
 258         if (resp->count == 0) {
 259                 nfs_set_pgio_error(hdr, -EIO, argp->offset);
 260                 return;
 261         }
 262
 263         /* For non rpc-based layout drivers, retry-through-MDS */
 264         if (!task->tk_ops) {
 265                 hdr->pnfs_error = -EAGAIN;
 266                 return;
 267         }
 268
 269         /* Yes, so retry the read at the end of the hdr */
 270         hdr->mds_offset += resp->count;
 271         argp->offset += resp->count;
 272         argp->pgbase += resp->count;
 273         argp->count -= resp->count;
 274         resp->count = 0;
 275         resp->eof = 0;
 276         rpc_restart_call_prepare(task);
 277 }
 278
 279 static void nfs_readpage_result(struct rpc_task *task,
 280                                 struct nfs_pgio_header *hdr)
 281 {
 282         if (hdr->res.eof) {
 283                 loff_t pos = hdr->args.offset + hdr->res.count;
 284                 unsigned int new = pos - hdr->io_start;
 285
 286                 if (hdr->good_bytes > new) {
 287                         hdr->good_bytes = new;
 288                         set_bit(NFS_IOHDR_EOF, &hdr->flags);
 289                         clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
 290                 }
 291         } else if (hdr->res.count < hdr->args.count)
 292                 nfs_readpage_retry(task, hdr);
 293 }
 294
 295 static int
 296 readpage_async_filler(struct nfs_readdesc *desc, struct page *page)
 297 {
 298         struct inode *inode = page_file_mapping(page)->host;
 299         unsigned int rsize = NFS_SERVER(inode)->rsize;
 300         struct nfs_page *new;
 301         unsigned int len, aligned_len;
 302         int error;
 303
 304         len = nfs_page_length(page);
 305         if (len == 0)
 306                 return nfs_return_empty_page(page);
 307
 308         aligned_len = min_t(unsigned int, ALIGN(len, rsize), PAGE_SIZE);
 309
 310         if (!IS_SYNC(page->mapping->host)) {
 311                 error = nfs_fscache_read_page(page->mapping->host, page);
 312                 if (error == 0)
 313                         goto out_unlock;
 314         }
 315
 316         new = nfs_create_request(desc->ctx, page, 0, aligned_len);
 317         if (IS_ERR(new))
 318                 goto out_error;
 319
 320         if (len < PAGE_SIZE)
 321                 zero_user_segment(page, len, PAGE_SIZE);
 322         if (!nfs_pageio_add_request(&desc->pgio, new)) {
 323                 nfs_list_remove_request(new);
 324                 error = desc->pgio.pg_error;
 325                 nfs_readpage_release(new, error);
 326                 goto out;
 327         }
 328         return 0;
 329 out_error:
 330         error = PTR_ERR(new);
 331 out_unlock:
 332         unlock_page(page);
 333 out:
 334         return error;
 335 }
 336
 337 /*
 338  * Read a page over NFS.
 339  * We read the page synchronously in the following case:
 340  *  -   The error flag is set for this page. This happens only when a
 341  *      previous async read operation failed.
 342  */
 343 int nfs_read_folio(struct file *file, struct folio *folio)
 344 {
 345         struct page *page = &folio->page;
 346         struct nfs_readdesc desc;
 347         struct inode *inode = page_file_mapping(page)->host;
 348         int ret;
 349
 350         trace_nfs_aop_readpage(inode, page);
 351         nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
 352         task_io_account_read(folio_size(folio));
 353
 354         /*
 355          * Try to flush any pending writes to the file..
 356          *
 357          * NOTE! Because we own the page lock, there cannot
 358          * be any new pending writes generated at this point
 359          * for this page (other pages can be written to).
 360          */
 361         ret = nfs_wb_page(inode, page);
 362         if (ret)
 363                 goto out_unlock;
 364         if (PageUptodate(page))
 365                 goto out_unlock;
 366
 367         ret = -ESTALE;
 368         if (NFS_STALE(inode))
 369                 goto out_unlock;
 370
 371         if (file == NULL) {
 372                 ret = -EBADF;
 373                 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 374                 if (desc.ctx == NULL)
 375                         goto out_unlock;
 376         } else
 377                 desc.ctx = get_nfs_open_context(nfs_file_open_context(file));
 378
 379         xchg(&desc.ctx->error, 0);
 380         nfs_pageio_init_read(&desc.pgio, inode, false,
 381                              &nfs_async_read_completion_ops);
 382
 383         ret = readpage_async_filler(&desc, page);
 384         if (ret)
 385                 goto out;
 386
 387         nfs_pageio_complete_read(&desc.pgio);
 388         ret = desc.pgio.pg_error < 0 ? desc.pgio.pg_error : 0;
 389         if (!ret) {
 390                 ret = wait_on_page_locked_killable(page);
 391                 if (!PageUptodate(page) && !ret)
 392                         ret = xchg(&desc.ctx->error, 0);
 393         }
 394 out:
 395         put_nfs_open_context(desc.ctx);
 396         trace_nfs_aop_readpage_done(inode, page, ret);
 397         return ret;
 398 out_unlock:
 399         unlock_page(page);
 400         trace_nfs_aop_readpage_done(inode, page, ret);
 401         return ret;
 402 }
 403
 404 void nfs_readahead(struct readahead_control *ractl)
 405 {
 406         unsigned int nr_pages = readahead_count(ractl);
 407         struct file *file = ractl->file;
 408         struct nfs_readdesc desc;
 409         struct inode *inode = ractl->mapping->host;
 410         struct page *page;
 411         int ret;
 412
 413         trace_nfs_aop_readahead(inode, readahead_pos(ractl), nr_pages);
 414         nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
 415         task_io_account_read(readahead_length(ractl));
 416
 417         ret = -ESTALE;
 418         if (NFS_STALE(inode))
 419                 goto out;
 420
 421         if (file == NULL) {
 422                 ret = -EBADF;
 423                 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
 424                 if (desc.ctx == NULL)
 425                         goto out;
 426         } else
 427                 desc.ctx = get_nfs_open_context(nfs_file_open_context(file));
 428
 429         nfs_pageio_init_read(&desc.pgio, inode, false,
 430                              &nfs_async_read_completion_ops);
 431
 432         while ((page = readahead_page(ractl)) != NULL) {
 433                 ret = readpage_async_filler(&desc, page);
 434                 put_page(page);
 435                 if (ret)
 436                         break;
 437         }
 438
 439         nfs_pageio_complete_read(&desc.pgio);
 440
 441         put_nfs_open_context(desc.ctx);
 442 out:
 443         trace_nfs_aop_readahead_done(inode, nr_pages, ret);
 444 }
 445
 446 int __init nfs_init_readpagecache(void)
 447 {
 448         nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
 449                                              sizeof(struct nfs_pgio_header),
 450                                              0, SLAB_HWCACHE_ALIGN,
 451                                              NULL);
 452         if (nfs_rdata_cachep == NULL)
 453                 return -ENOMEM;
 454
 455         return 0;
 456 }
 457
 458 void nfs_destroy_readpagecache(void)
 459 {
 460         kmem_cache_destroy(nfs_rdata_cachep);
 461 }
 462
 463 static const struct nfs_rw_ops nfs_rw_read_ops = {
 464         .rw_alloc_header        = nfs_readhdr_alloc,
 465         .rw_free_header         = nfs_readhdr_free,
 466         .rw_done                = nfs_readpage_done,
 467         .rw_result              = nfs_readpage_result,
 468         .rw_initiate            = nfs_initiate_read,
 469 };