6 * Partial copy of Linus' read cache modifications to fs/nfs/file.c
7 * modified for async RPC by okir@monad.swb.de
10 #include <linux/time.h>
11 #include <linux/kernel.h>
12 #include <linux/errno.h>
13 #include <linux/fcntl.h>
14 #include <linux/stat.h>
16 #include <linux/slab.h>
17 #include <linux/pagemap.h>
18 #include <linux/sunrpc/clnt.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_page.h>
21 #include <linux/module.h>
28 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
30 static const struct nfs_pageio_ops nfs_pageio_read_ops;
31 static const struct rpc_call_ops nfs_read_common_ops;
32 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops;
34 static struct kmem_cache *nfs_rdata_cachep;
36 struct nfs_read_header *nfs_readhdr_alloc(void)
38 struct nfs_read_header *rhdr;
40 rhdr = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
42 struct nfs_pgio_header *hdr = &rhdr->header;
44 INIT_LIST_HEAD(&hdr->pages);
45 INIT_LIST_HEAD(&hdr->rpc_list);
46 spin_lock_init(&hdr->lock);
47 atomic_set(&hdr->refcnt, 0);
51 EXPORT_SYMBOL_GPL(nfs_readhdr_alloc);
53 static struct nfs_read_data *nfs_readdata_alloc(struct nfs_pgio_header *hdr,
54 unsigned int pagecount)
56 struct nfs_read_data *data, *prealloc;
58 prealloc = &container_of(hdr, struct nfs_read_header, header)->rpc_data;
59 if (prealloc->header == NULL)
62 data = kzalloc(sizeof(*data), GFP_KERNEL);
66 if (nfs_pgarray_set(&data->pages, pagecount)) {
68 atomic_inc(&hdr->refcnt);
78 void nfs_readhdr_free(struct nfs_pgio_header *hdr)
80 struct nfs_read_header *rhdr = container_of(hdr, struct nfs_read_header, header);
82 kmem_cache_free(nfs_rdata_cachep, rhdr);
84 EXPORT_SYMBOL_GPL(nfs_readhdr_free);
86 void nfs_readdata_release(struct nfs_read_data *rdata)
88 struct nfs_pgio_header *hdr = rdata->header;
89 struct nfs_read_header *read_header = container_of(hdr, struct nfs_read_header, header);
91 put_nfs_open_context(rdata->args.context);
92 if (rdata->pages.pagevec != rdata->pages.page_array)
93 kfree(rdata->pages.pagevec);
94 if (rdata == &read_header->rpc_data) {
98 if (atomic_dec_and_test(&hdr->refcnt))
99 hdr->completion_ops->completion(hdr);
100 /* Note: we only free the rpc_task after callbacks are done.
101 * See the comment in rpc_free_task() for why
105 EXPORT_SYMBOL_GPL(nfs_readdata_release);
108 int nfs_return_empty_page(struct page *page)
110 zero_user(page, 0, PAGE_CACHE_SIZE);
111 SetPageUptodate(page);
116 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
118 const struct nfs_pgio_completion_ops *compl_ops)
120 nfs_pageio_init(pgio, inode, &nfs_pageio_read_ops, compl_ops,
121 NFS_SERVER(inode)->rsize, 0);
123 EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
125 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
127 pgio->pg_ops = &nfs_pageio_read_ops;
128 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
130 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
132 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
135 struct nfs_page *new;
137 struct nfs_pageio_descriptor pgio;
139 len = nfs_page_length(page);
141 return nfs_return_empty_page(page);
142 new = nfs_create_request(ctx, inode, page, 0, len);
147 if (len < PAGE_CACHE_SIZE)
148 zero_user_segment(page, len, PAGE_CACHE_SIZE);
150 NFS_PROTO(inode)->read_pageio_init(&pgio, inode, &nfs_async_read_completion_ops);
151 nfs_pageio_add_request(&pgio, new);
152 nfs_pageio_complete(&pgio);
153 NFS_I(inode)->read_io += pgio.pg_bytes_written;
157 static void nfs_readpage_release(struct nfs_page *req)
159 struct inode *d_inode = req->wb_context->dentry->d_inode;
161 if (PageUptodate(req->wb_page))
162 nfs_readpage_to_fscache(d_inode, req->wb_page, 0);
164 unlock_page(req->wb_page);
166 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
167 req->wb_context->dentry->d_inode->i_sb->s_id,
168 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
170 (long long)req_offset(req));
171 nfs_release_request(req);
174 /* Note io was page aligned */
175 static void nfs_read_completion(struct nfs_pgio_header *hdr)
177 unsigned long bytes = 0;
179 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
181 while (!list_empty(&hdr->pages)) {
182 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
183 struct page *page = req->wb_page;
185 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
186 if (bytes > hdr->good_bytes)
187 zero_user(page, 0, PAGE_SIZE);
188 else if (hdr->good_bytes - bytes < PAGE_SIZE)
189 zero_user_segment(page,
190 hdr->good_bytes & ~PAGE_MASK,
193 bytes += req->wb_bytes;
194 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
195 if (bytes <= hdr->good_bytes)
196 SetPageUptodate(page);
198 SetPageUptodate(page);
199 nfs_list_remove_request(req);
200 nfs_readpage_release(req);
206 int nfs_initiate_read(struct rpc_clnt *clnt,
207 struct nfs_read_data *data,
208 const struct rpc_call_ops *call_ops, int flags)
210 struct inode *inode = data->header->inode;
211 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
212 struct rpc_task *task;
213 struct rpc_message msg = {
214 .rpc_argp = &data->args,
215 .rpc_resp = &data->res,
216 .rpc_cred = data->header->cred,
218 struct rpc_task_setup task_setup_data = {
222 .callback_ops = call_ops,
223 .callback_data = data,
224 .workqueue = nfsiod_workqueue,
225 .flags = RPC_TASK_ASYNC | swap_flags | flags,
228 /* Set up the initial task struct. */
229 NFS_PROTO(inode)->read_setup(data, &msg);
231 dprintk("NFS: %5u initiated read call (req %s/%lld, %u bytes @ "
235 (long long)NFS_FILEID(inode),
237 (unsigned long long)data->args.offset);
239 task = rpc_run_task(&task_setup_data);
241 return PTR_ERR(task);
245 EXPORT_SYMBOL_GPL(nfs_initiate_read);
248 * Set up the NFS read request struct
250 static void nfs_read_rpcsetup(struct nfs_read_data *data,
251 unsigned int count, unsigned int offset)
253 struct nfs_page *req = data->header->req;
255 data->args.fh = NFS_FH(data->header->inode);
256 data->args.offset = req_offset(req) + offset;
257 data->args.pgbase = req->wb_pgbase + offset;
258 data->args.pages = data->pages.pagevec;
259 data->args.count = count;
260 data->args.context = get_nfs_open_context(req->wb_context);
261 data->args.lock_context = req->wb_lock_context;
263 data->res.fattr = &data->fattr;
264 data->res.count = count;
266 nfs_fattr_init(&data->fattr);
269 static int nfs_do_read(struct nfs_read_data *data,
270 const struct rpc_call_ops *call_ops)
272 struct inode *inode = data->header->inode;
274 return nfs_initiate_read(NFS_CLIENT(inode), data, call_ops, 0);
278 nfs_do_multiple_reads(struct list_head *head,
279 const struct rpc_call_ops *call_ops)
281 struct nfs_read_data *data;
284 while (!list_empty(head)) {
287 data = list_first_entry(head, struct nfs_read_data, list);
288 list_del_init(&data->list);
290 ret2 = nfs_do_read(data, call_ops);
298 nfs_async_read_error(struct list_head *head)
300 struct nfs_page *req;
302 while (!list_empty(head)) {
303 req = nfs_list_entry(head->next);
304 nfs_list_remove_request(req);
305 nfs_readpage_release(req);
309 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
310 .error_cleanup = nfs_async_read_error,
311 .completion = nfs_read_completion,
314 static void nfs_pagein_error(struct nfs_pageio_descriptor *desc,
315 struct nfs_pgio_header *hdr)
317 set_bit(NFS_IOHDR_REDO, &hdr->flags);
318 while (!list_empty(&hdr->rpc_list)) {
319 struct nfs_read_data *data = list_first_entry(&hdr->rpc_list,
320 struct nfs_read_data, list);
321 list_del(&data->list);
322 nfs_readdata_release(data);
324 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
328 * Generate multiple requests to fill a single page.
330 * We optimize to reduce the number of read operations on the wire. If we
331 * detect that we're reading a page, or an area of a page, that is past the
332 * end of file, we do not generate NFS read operations but just clear the
333 * parts of the page that would have come back zero from the server anyway.
335 * We rely on the cached value of i_size to make this determination; another
336 * client can fill pages on the server past our cached end-of-file, but we
337 * won't see the new data until our attribute cache is updated. This is more
338 * or less conventional NFS client behavior.
340 static int nfs_pagein_multi(struct nfs_pageio_descriptor *desc,
341 struct nfs_pgio_header *hdr)
343 struct nfs_page *req = hdr->req;
344 struct page *page = req->wb_page;
345 struct nfs_read_data *data;
346 size_t rsize = desc->pg_bsize, nbytes;
350 nbytes = desc->pg_count;
352 size_t len = min(nbytes,rsize);
354 data = nfs_readdata_alloc(hdr, 1);
356 nfs_pagein_error(desc, hdr);
359 data->pages.pagevec[0] = page;
360 nfs_read_rpcsetup(data, len, offset);
361 list_add(&data->list, &hdr->rpc_list);
364 } while (nbytes != 0);
366 nfs_list_remove_request(req);
367 nfs_list_add_request(req, &hdr->pages);
368 desc->pg_rpc_callops = &nfs_read_common_ops;
372 static int nfs_pagein_one(struct nfs_pageio_descriptor *desc,
373 struct nfs_pgio_header *hdr)
375 struct nfs_page *req;
377 struct nfs_read_data *data;
378 struct list_head *head = &desc->pg_list;
380 data = nfs_readdata_alloc(hdr, nfs_page_array_len(desc->pg_base,
383 nfs_pagein_error(desc, hdr);
387 pages = data->pages.pagevec;
388 while (!list_empty(head)) {
389 req = nfs_list_entry(head->next);
390 nfs_list_remove_request(req);
391 nfs_list_add_request(req, &hdr->pages);
392 *pages++ = req->wb_page;
395 nfs_read_rpcsetup(data, desc->pg_count, 0);
396 list_add(&data->list, &hdr->rpc_list);
397 desc->pg_rpc_callops = &nfs_read_common_ops;
401 int nfs_generic_pagein(struct nfs_pageio_descriptor *desc,
402 struct nfs_pgio_header *hdr)
404 if (desc->pg_bsize < PAGE_CACHE_SIZE)
405 return nfs_pagein_multi(desc, hdr);
406 return nfs_pagein_one(desc, hdr);
408 EXPORT_SYMBOL_GPL(nfs_generic_pagein);
410 static int nfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
412 struct nfs_read_header *rhdr;
413 struct nfs_pgio_header *hdr;
416 rhdr = nfs_readhdr_alloc();
418 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
422 nfs_pgheader_init(desc, hdr, nfs_readhdr_free);
423 atomic_inc(&hdr->refcnt);
424 ret = nfs_generic_pagein(desc, hdr);
426 ret = nfs_do_multiple_reads(&hdr->rpc_list,
427 desc->pg_rpc_callops);
428 if (atomic_dec_and_test(&hdr->refcnt))
429 hdr->completion_ops->completion(hdr);
433 static const struct nfs_pageio_ops nfs_pageio_read_ops = {
434 .pg_test = nfs_generic_pg_test,
435 .pg_doio = nfs_generic_pg_readpages,
439 * This is the callback from RPC telling us whether a reply was
440 * received or some error occurred (timeout or socket shutdown).
442 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
444 struct inode *inode = data->header->inode;
447 dprintk("NFS: %s: %5u, (status %d)\n", __func__, task->tk_pid,
450 status = NFS_PROTO(inode)->read_done(task, data);
454 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, data->res.count);
456 if (task->tk_status == -ESTALE) {
457 set_bit(NFS_INO_STALE, &NFS_I(inode)->flags);
458 nfs_mark_for_revalidate(inode);
463 static void nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
465 struct nfs_readargs *argp = &data->args;
466 struct nfs_readres *resp = &data->res;
468 /* This is a short read! */
469 nfs_inc_stats(data->header->inode, NFSIOS_SHORTREAD);
470 /* Has the server at least made some progress? */
471 if (resp->count == 0) {
472 nfs_set_pgio_error(data->header, -EIO, argp->offset);
475 /* Yes, so retry the read at the end of the data */
476 data->mds_offset += resp->count;
477 argp->offset += resp->count;
478 argp->pgbase += resp->count;
479 argp->count -= resp->count;
480 rpc_restart_call_prepare(task);
483 static void nfs_readpage_result_common(struct rpc_task *task, void *calldata)
485 struct nfs_read_data *data = calldata;
486 struct nfs_pgio_header *hdr = data->header;
488 /* Note the only returns of nfs_readpage_result are 0 and -EAGAIN */
489 if (nfs_readpage_result(task, data) != 0)
491 if (task->tk_status < 0)
492 nfs_set_pgio_error(hdr, task->tk_status, data->args.offset);
493 else if (data->res.eof) {
496 bound = data->args.offset + data->res.count;
497 spin_lock(&hdr->lock);
498 if (bound < hdr->io_start + hdr->good_bytes) {
499 set_bit(NFS_IOHDR_EOF, &hdr->flags);
500 clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
501 hdr->good_bytes = bound - hdr->io_start;
503 spin_unlock(&hdr->lock);
504 } else if (data->res.count != data->args.count)
505 nfs_readpage_retry(task, data);
508 static void nfs_readpage_release_common(void *calldata)
510 nfs_readdata_release(calldata);
513 void nfs_read_prepare(struct rpc_task *task, void *calldata)
515 struct nfs_read_data *data = calldata;
516 NFS_PROTO(data->header->inode)->read_rpc_prepare(task, data);
519 static const struct rpc_call_ops nfs_read_common_ops = {
520 .rpc_call_prepare = nfs_read_prepare,
521 .rpc_call_done = nfs_readpage_result_common,
522 .rpc_release = nfs_readpage_release_common,
526 * Read a page over NFS.
527 * We read the page synchronously in the following case:
528 * - The error flag is set for this page. This happens only when a
529 * previous async read operation failed.
531 int nfs_readpage(struct file *file, struct page *page)
533 struct nfs_open_context *ctx;
534 struct inode *inode = page_file_mapping(page)->host;
537 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
538 page, PAGE_CACHE_SIZE, page_file_index(page));
539 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
540 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
543 * Try to flush any pending writes to the file..
545 * NOTE! Because we own the page lock, there cannot
546 * be any new pending writes generated at this point
547 * for this page (other pages can be written to).
549 error = nfs_wb_page(inode, page);
552 if (PageUptodate(page))
556 if (NFS_STALE(inode))
561 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
565 ctx = get_nfs_open_context(nfs_file_open_context(file));
567 if (!IS_SYNC(inode)) {
568 error = nfs_readpage_from_fscache(ctx, inode, page);
573 error = nfs_readpage_async(ctx, inode, page);
576 put_nfs_open_context(ctx);
583 struct nfs_readdesc {
584 struct nfs_pageio_descriptor *pgio;
585 struct nfs_open_context *ctx;
589 readpage_async_filler(void *data, struct page *page)
591 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
592 struct inode *inode = page_file_mapping(page)->host;
593 struct nfs_page *new;
597 len = nfs_page_length(page);
599 return nfs_return_empty_page(page);
601 new = nfs_create_request(desc->ctx, inode, page, 0, len);
605 if (len < PAGE_CACHE_SIZE)
606 zero_user_segment(page, len, PAGE_CACHE_SIZE);
607 if (!nfs_pageio_add_request(desc->pgio, new)) {
608 error = desc->pgio->pg_error;
613 error = PTR_ERR(new);
619 int nfs_readpages(struct file *filp, struct address_space *mapping,
620 struct list_head *pages, unsigned nr_pages)
622 struct nfs_pageio_descriptor pgio;
623 struct nfs_readdesc desc = {
626 struct inode *inode = mapping->host;
627 unsigned long npages;
630 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
632 (long long)NFS_FILEID(inode),
634 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
636 if (NFS_STALE(inode))
640 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
641 if (desc.ctx == NULL)
644 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
646 /* attempt to read as many of the pages as possible from the cache
647 * - this returns -ENOBUFS immediately if the cookie is negative
649 ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
652 goto read_complete; /* all pages were read */
654 NFS_PROTO(inode)->read_pageio_init(&pgio, inode, &nfs_async_read_completion_ops);
656 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
658 nfs_pageio_complete(&pgio);
659 NFS_I(inode)->read_io += pgio.pg_bytes_written;
660 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
661 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
663 put_nfs_open_context(desc.ctx);
668 int __init nfs_init_readpagecache(void)
670 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
671 sizeof(struct nfs_read_header),
672 0, SLAB_HWCACHE_ALIGN,
674 if (nfs_rdata_cachep == NULL)
680 void nfs_destroy_readpagecache(void)
682 kmem_cache_destroy(nfs_rdata_cachep);