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>
29 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
31 static const struct nfs_pageio_ops nfs_pageio_read_ops;
32 static const struct rpc_call_ops nfs_read_common_ops;
33 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops;
35 static struct kmem_cache *nfs_rdata_cachep;
37 struct nfs_rw_header *nfs_readhdr_alloc(void)
39 struct nfs_rw_header *rhdr;
41 rhdr = kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
43 struct nfs_pgio_header *hdr = &rhdr->header;
45 INIT_LIST_HEAD(&hdr->pages);
46 INIT_LIST_HEAD(&hdr->rpc_list);
47 spin_lock_init(&hdr->lock);
48 atomic_set(&hdr->refcnt, 0);
52 EXPORT_SYMBOL_GPL(nfs_readhdr_alloc);
54 static struct nfs_pgio_data *nfs_readdata_alloc(struct nfs_pgio_header *hdr,
55 unsigned int pagecount)
57 struct nfs_pgio_data *data, *prealloc;
59 prealloc = &container_of(hdr, struct nfs_rw_header, header)->rpc_data;
60 if (prealloc->header == NULL)
63 data = kzalloc(sizeof(*data), GFP_KERNEL);
67 if (nfs_pgarray_set(&data->pages, pagecount)) {
69 atomic_inc(&hdr->refcnt);
79 void nfs_readhdr_free(struct nfs_pgio_header *hdr)
81 struct nfs_rw_header *rhdr = container_of(hdr, struct nfs_rw_header, header);
83 kmem_cache_free(nfs_rdata_cachep, rhdr);
85 EXPORT_SYMBOL_GPL(nfs_readhdr_free);
87 void nfs_readdata_release(struct nfs_pgio_data *rdata)
89 struct nfs_pgio_header *hdr = rdata->header;
90 struct nfs_rw_header *read_header = container_of(hdr, struct nfs_rw_header, header);
92 put_nfs_open_context(rdata->args.context);
93 if (rdata->pages.pagevec != rdata->pages.page_array)
94 kfree(rdata->pages.pagevec);
95 if (rdata == &read_header->rpc_data) {
99 if (atomic_dec_and_test(&hdr->refcnt))
100 hdr->completion_ops->completion(hdr);
101 /* Note: we only free the rpc_task after callbacks are done.
102 * See the comment in rpc_free_task() for why
106 EXPORT_SYMBOL_GPL(nfs_readdata_release);
109 int nfs_return_empty_page(struct page *page)
111 zero_user(page, 0, PAGE_CACHE_SIZE);
112 SetPageUptodate(page);
117 void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
118 struct inode *inode, bool force_mds,
119 const struct nfs_pgio_completion_ops *compl_ops)
121 struct nfs_server *server = NFS_SERVER(inode);
122 const struct nfs_pageio_ops *pg_ops = &nfs_pageio_read_ops;
124 #ifdef CONFIG_NFS_V4_1
125 if (server->pnfs_curr_ld && !force_mds)
126 pg_ops = server->pnfs_curr_ld->pg_read_ops;
128 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, server->rsize, 0);
130 EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
132 void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
134 pgio->pg_ops = &nfs_pageio_read_ops;
135 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
137 EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
139 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
142 struct nfs_page *new;
144 struct nfs_pageio_descriptor pgio;
146 len = nfs_page_length(page);
148 return nfs_return_empty_page(page);
149 new = nfs_create_request(ctx, inode, page, 0, len);
154 if (len < PAGE_CACHE_SIZE)
155 zero_user_segment(page, len, PAGE_CACHE_SIZE);
157 nfs_pageio_init_read(&pgio, inode, false,
158 &nfs_async_read_completion_ops);
159 nfs_pageio_add_request(&pgio, new);
160 nfs_pageio_complete(&pgio);
161 NFS_I(inode)->read_io += pgio.pg_bytes_written;
165 static void nfs_readpage_release(struct nfs_page *req)
167 struct inode *d_inode = req->wb_context->dentry->d_inode;
169 if (PageUptodate(req->wb_page))
170 nfs_readpage_to_fscache(d_inode, req->wb_page, 0);
172 unlock_page(req->wb_page);
174 dprintk("NFS: read done (%s/%Lu %d@%Ld)\n",
175 req->wb_context->dentry->d_inode->i_sb->s_id,
176 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
178 (long long)req_offset(req));
179 nfs_release_request(req);
182 /* Note io was page aligned */
183 static void nfs_read_completion(struct nfs_pgio_header *hdr)
185 unsigned long bytes = 0;
187 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
189 while (!list_empty(&hdr->pages)) {
190 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
191 struct page *page = req->wb_page;
193 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
194 if (bytes > hdr->good_bytes)
195 zero_user(page, 0, PAGE_SIZE);
196 else if (hdr->good_bytes - bytes < PAGE_SIZE)
197 zero_user_segment(page,
198 hdr->good_bytes & ~PAGE_MASK,
201 bytes += req->wb_bytes;
202 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
203 if (bytes <= hdr->good_bytes)
204 SetPageUptodate(page);
206 SetPageUptodate(page);
207 nfs_list_remove_request(req);
208 nfs_readpage_release(req);
214 int nfs_initiate_read(struct rpc_clnt *clnt,
215 struct nfs_pgio_data *data,
216 const struct rpc_call_ops *call_ops, int flags)
218 struct inode *inode = data->header->inode;
219 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
220 struct rpc_task *task;
221 struct rpc_message msg = {
222 .rpc_argp = &data->args,
223 .rpc_resp = &data->res,
224 .rpc_cred = data->header->cred,
226 struct rpc_task_setup task_setup_data = {
230 .callback_ops = call_ops,
231 .callback_data = data,
232 .workqueue = nfsiod_workqueue,
233 .flags = RPC_TASK_ASYNC | swap_flags | flags,
236 /* Set up the initial task struct. */
237 NFS_PROTO(inode)->read_setup(data, &msg);
239 dprintk("NFS: %5u initiated read call (req %s/%llu, %u bytes @ "
243 (unsigned long long)NFS_FILEID(inode),
245 (unsigned long long)data->args.offset);
247 task = rpc_run_task(&task_setup_data);
249 return PTR_ERR(task);
253 EXPORT_SYMBOL_GPL(nfs_initiate_read);
256 * Set up the NFS read request struct
258 static void nfs_read_rpcsetup(struct nfs_pgio_data *data,
259 unsigned int count, unsigned int offset)
261 struct nfs_page *req = data->header->req;
263 data->args.fh = NFS_FH(data->header->inode);
264 data->args.offset = req_offset(req) + offset;
265 data->args.pgbase = req->wb_pgbase + offset;
266 data->args.pages = data->pages.pagevec;
267 data->args.count = count;
268 data->args.context = get_nfs_open_context(req->wb_context);
269 data->args.lock_context = req->wb_lock_context;
271 data->res.fattr = &data->fattr;
272 data->res.count = count;
274 nfs_fattr_init(&data->fattr);
277 static int nfs_do_read(struct nfs_pgio_data *data,
278 const struct rpc_call_ops *call_ops)
280 struct inode *inode = data->header->inode;
282 return nfs_initiate_read(NFS_CLIENT(inode), data, call_ops, 0);
286 nfs_do_multiple_reads(struct list_head *head,
287 const struct rpc_call_ops *call_ops)
289 struct nfs_pgio_data *data;
292 while (!list_empty(head)) {
295 data = list_first_entry(head, struct nfs_pgio_data, list);
296 list_del_init(&data->list);
298 ret2 = nfs_do_read(data, call_ops);
306 nfs_async_read_error(struct list_head *head)
308 struct nfs_page *req;
310 while (!list_empty(head)) {
311 req = nfs_list_entry(head->next);
312 nfs_list_remove_request(req);
313 nfs_readpage_release(req);
317 static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
318 .error_cleanup = nfs_async_read_error,
319 .completion = nfs_read_completion,
322 static void nfs_pagein_error(struct nfs_pageio_descriptor *desc,
323 struct nfs_pgio_header *hdr)
325 set_bit(NFS_IOHDR_REDO, &hdr->flags);
326 while (!list_empty(&hdr->rpc_list)) {
327 struct nfs_pgio_data *data = list_first_entry(&hdr->rpc_list,
328 struct nfs_pgio_data, list);
329 list_del(&data->list);
330 nfs_readdata_release(data);
332 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
336 * Generate multiple requests to fill a single page.
338 * We optimize to reduce the number of read operations on the wire. If we
339 * detect that we're reading a page, or an area of a page, that is past the
340 * end of file, we do not generate NFS read operations but just clear the
341 * parts of the page that would have come back zero from the server anyway.
343 * We rely on the cached value of i_size to make this determination; another
344 * client can fill pages on the server past our cached end-of-file, but we
345 * won't see the new data until our attribute cache is updated. This is more
346 * or less conventional NFS client behavior.
348 static int nfs_pagein_multi(struct nfs_pageio_descriptor *desc,
349 struct nfs_pgio_header *hdr)
351 struct nfs_page *req = hdr->req;
352 struct page *page = req->wb_page;
353 struct nfs_pgio_data *data;
354 size_t rsize = desc->pg_bsize, nbytes;
358 nbytes = desc->pg_count;
360 size_t len = min(nbytes,rsize);
362 data = nfs_readdata_alloc(hdr, 1);
364 nfs_pagein_error(desc, hdr);
367 data->pages.pagevec[0] = page;
368 nfs_read_rpcsetup(data, len, offset);
369 list_add(&data->list, &hdr->rpc_list);
372 } while (nbytes != 0);
374 nfs_list_remove_request(req);
375 nfs_list_add_request(req, &hdr->pages);
376 desc->pg_rpc_callops = &nfs_read_common_ops;
380 static int nfs_pagein_one(struct nfs_pageio_descriptor *desc,
381 struct nfs_pgio_header *hdr)
383 struct nfs_page *req;
385 struct nfs_pgio_data *data;
386 struct list_head *head = &desc->pg_list;
388 data = nfs_readdata_alloc(hdr, nfs_page_array_len(desc->pg_base,
391 nfs_pagein_error(desc, hdr);
395 pages = data->pages.pagevec;
396 while (!list_empty(head)) {
397 req = nfs_list_entry(head->next);
398 nfs_list_remove_request(req);
399 nfs_list_add_request(req, &hdr->pages);
400 *pages++ = req->wb_page;
403 nfs_read_rpcsetup(data, desc->pg_count, 0);
404 list_add(&data->list, &hdr->rpc_list);
405 desc->pg_rpc_callops = &nfs_read_common_ops;
409 int nfs_generic_pagein(struct nfs_pageio_descriptor *desc,
410 struct nfs_pgio_header *hdr)
412 if (desc->pg_bsize < PAGE_CACHE_SIZE)
413 return nfs_pagein_multi(desc, hdr);
414 return nfs_pagein_one(desc, hdr);
416 EXPORT_SYMBOL_GPL(nfs_generic_pagein);
418 static int nfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
420 struct nfs_rw_header *rhdr;
421 struct nfs_pgio_header *hdr;
424 rhdr = nfs_readhdr_alloc();
426 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
430 nfs_pgheader_init(desc, hdr, nfs_readhdr_free);
431 atomic_inc(&hdr->refcnt);
432 ret = nfs_generic_pagein(desc, hdr);
434 ret = nfs_do_multiple_reads(&hdr->rpc_list,
435 desc->pg_rpc_callops);
436 if (atomic_dec_and_test(&hdr->refcnt))
437 hdr->completion_ops->completion(hdr);
441 static const struct nfs_pageio_ops nfs_pageio_read_ops = {
442 .pg_test = nfs_generic_pg_test,
443 .pg_doio = nfs_generic_pg_readpages,
447 * This is the callback from RPC telling us whether a reply was
448 * received or some error occurred (timeout or socket shutdown).
450 int nfs_readpage_result(struct rpc_task *task, struct nfs_pgio_data *data)
452 struct inode *inode = data->header->inode;
455 dprintk("NFS: %s: %5u, (status %d)\n", __func__, task->tk_pid,
458 status = NFS_PROTO(inode)->read_done(task, data);
462 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, data->res.count);
464 if (task->tk_status == -ESTALE) {
465 set_bit(NFS_INO_STALE, &NFS_I(inode)->flags);
466 nfs_mark_for_revalidate(inode);
471 static void nfs_readpage_retry(struct rpc_task *task, struct nfs_pgio_data *data)
473 struct nfs_pgio_args *argp = &data->args;
474 struct nfs_pgio_res *resp = &data->res;
476 /* This is a short read! */
477 nfs_inc_stats(data->header->inode, NFSIOS_SHORTREAD);
478 /* Has the server at least made some progress? */
479 if (resp->count == 0) {
480 nfs_set_pgio_error(data->header, -EIO, argp->offset);
483 /* Yes, so retry the read at the end of the data */
484 data->mds_offset += resp->count;
485 argp->offset += resp->count;
486 argp->pgbase += resp->count;
487 argp->count -= resp->count;
488 rpc_restart_call_prepare(task);
491 static void nfs_readpage_result_common(struct rpc_task *task, void *calldata)
493 struct nfs_pgio_data *data = calldata;
494 struct nfs_pgio_header *hdr = data->header;
496 /* Note the only returns of nfs_readpage_result are 0 and -EAGAIN */
497 if (nfs_readpage_result(task, data) != 0)
499 if (task->tk_status < 0)
500 nfs_set_pgio_error(hdr, task->tk_status, data->args.offset);
501 else if (data->res.eof) {
504 bound = data->args.offset + data->res.count;
505 spin_lock(&hdr->lock);
506 if (bound < hdr->io_start + hdr->good_bytes) {
507 set_bit(NFS_IOHDR_EOF, &hdr->flags);
508 clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
509 hdr->good_bytes = bound - hdr->io_start;
511 spin_unlock(&hdr->lock);
512 } else if (data->res.count != data->args.count)
513 nfs_readpage_retry(task, data);
516 static void nfs_readpage_release_common(void *calldata)
518 nfs_readdata_release(calldata);
521 void nfs_read_prepare(struct rpc_task *task, void *calldata)
523 struct nfs_pgio_data *data = calldata;
525 err = NFS_PROTO(data->header->inode)->read_rpc_prepare(task, data);
530 static const struct rpc_call_ops nfs_read_common_ops = {
531 .rpc_call_prepare = nfs_read_prepare,
532 .rpc_call_done = nfs_readpage_result_common,
533 .rpc_release = nfs_readpage_release_common,
537 * Read a page over NFS.
538 * We read the page synchronously in the following case:
539 * - The error flag is set for this page. This happens only when a
540 * previous async read operation failed.
542 int nfs_readpage(struct file *file, struct page *page)
544 struct nfs_open_context *ctx;
545 struct inode *inode = page_file_mapping(page)->host;
548 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
549 page, PAGE_CACHE_SIZE, page_file_index(page));
550 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
551 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
554 * Try to flush any pending writes to the file..
556 * NOTE! Because we own the page lock, there cannot
557 * be any new pending writes generated at this point
558 * for this page (other pages can be written to).
560 error = nfs_wb_page(inode, page);
563 if (PageUptodate(page))
567 if (NFS_STALE(inode))
572 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
576 ctx = get_nfs_open_context(nfs_file_open_context(file));
578 if (!IS_SYNC(inode)) {
579 error = nfs_readpage_from_fscache(ctx, inode, page);
584 error = nfs_readpage_async(ctx, inode, page);
587 put_nfs_open_context(ctx);
594 struct nfs_readdesc {
595 struct nfs_pageio_descriptor *pgio;
596 struct nfs_open_context *ctx;
600 readpage_async_filler(void *data, struct page *page)
602 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
603 struct inode *inode = page_file_mapping(page)->host;
604 struct nfs_page *new;
608 len = nfs_page_length(page);
610 return nfs_return_empty_page(page);
612 new = nfs_create_request(desc->ctx, inode, page, 0, len);
616 if (len < PAGE_CACHE_SIZE)
617 zero_user_segment(page, len, PAGE_CACHE_SIZE);
618 if (!nfs_pageio_add_request(desc->pgio, new)) {
619 error = desc->pgio->pg_error;
624 error = PTR_ERR(new);
630 int nfs_readpages(struct file *filp, struct address_space *mapping,
631 struct list_head *pages, unsigned nr_pages)
633 struct nfs_pageio_descriptor pgio;
634 struct nfs_readdesc desc = {
637 struct inode *inode = mapping->host;
638 unsigned long npages;
641 dprintk("NFS: nfs_readpages (%s/%Lu %d)\n",
643 (unsigned long long)NFS_FILEID(inode),
645 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
647 if (NFS_STALE(inode))
651 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
652 if (desc.ctx == NULL)
655 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
657 /* attempt to read as many of the pages as possible from the cache
658 * - this returns -ENOBUFS immediately if the cookie is negative
660 ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
663 goto read_complete; /* all pages were read */
665 nfs_pageio_init_read(&pgio, inode, false,
666 &nfs_async_read_completion_ops);
668 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
670 nfs_pageio_complete(&pgio);
671 NFS_I(inode)->read_io += pgio.pg_bytes_written;
672 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
673 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
675 put_nfs_open_context(desc.ctx);
680 int __init nfs_init_readpagecache(void)
682 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
683 sizeof(struct nfs_rw_header),
684 0, SLAB_HWCACHE_ALIGN,
686 if (nfs_rdata_cachep == NULL)
692 void nfs_destroy_readpagecache(void)
694 kmem_cache_destroy(nfs_rdata_cachep);