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/smp_lock.h>
22 #include <linux/module.h>
24 #include <asm/system.h>
32 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
34 static int nfs_pagein_multi(struct inode *, struct list_head *, unsigned int, size_t, int, struct pnfs_layout_segment *);
35 static int nfs_pagein_one(struct inode *, struct list_head *, unsigned int, size_t, int, struct pnfs_layout_segment *);
36 static const struct rpc_call_ops nfs_read_partial_ops;
37 static const struct rpc_call_ops nfs_read_full_ops;
39 static struct kmem_cache *nfs_rdata_cachep;
40 static mempool_t *nfs_rdata_mempool;
42 #define MIN_POOL_READ (32)
44 struct nfs_read_data *nfs_readdata_alloc(unsigned int pagecount)
46 struct nfs_read_data *p = mempool_alloc(nfs_rdata_mempool, GFP_KERNEL);
49 memset(p, 0, sizeof(*p));
50 INIT_LIST_HEAD(&p->pages);
51 p->npages = pagecount;
52 if (pagecount <= ARRAY_SIZE(p->page_array))
53 p->pagevec = p->page_array;
55 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_KERNEL);
57 mempool_free(p, nfs_rdata_mempool);
65 void nfs_readdata_free(struct nfs_read_data *p)
67 if (p && (p->pagevec != &p->page_array[0]))
69 mempool_free(p, nfs_rdata_mempool);
72 static void nfs_readdata_release(struct nfs_read_data *rdata)
74 put_lseg(rdata->lseg);
75 put_nfs_open_context(rdata->args.context);
76 nfs_readdata_free(rdata);
80 int nfs_return_empty_page(struct page *page)
82 zero_user(page, 0, PAGE_CACHE_SIZE);
83 SetPageUptodate(page);
88 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
90 unsigned int remainder = data->args.count - data->res.count;
91 unsigned int base = data->args.pgbase + data->res.count;
95 if (data->res.eof == 0 || remainder == 0)
98 * Note: "remainder" can never be negative, since we check for
99 * this in the XDR code.
101 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
102 base &= ~PAGE_CACHE_MASK;
103 pglen = PAGE_CACHE_SIZE - base;
105 if (remainder <= pglen) {
106 zero_user(*pages, base, remainder);
109 zero_user(*pages, base, pglen);
112 pglen = PAGE_CACHE_SIZE;
117 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
120 LIST_HEAD(one_request);
121 struct nfs_page *new;
124 len = nfs_page_length(page);
126 return nfs_return_empty_page(page);
127 new = nfs_create_request(ctx, inode, page, 0, len);
132 if (len < PAGE_CACHE_SIZE)
133 zero_user_segment(page, len, PAGE_CACHE_SIZE);
135 nfs_list_add_request(new, &one_request);
136 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
137 nfs_pagein_multi(inode, &one_request, 1, len, 0, NULL);
139 nfs_pagein_one(inode, &one_request, 1, len, 0, NULL);
143 static void nfs_readpage_release(struct nfs_page *req)
145 struct inode *d_inode = req->wb_context->path.dentry->d_inode;
147 if (PageUptodate(req->wb_page))
148 nfs_readpage_to_fscache(d_inode, req->wb_page, 0);
150 unlock_page(req->wb_page);
152 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
153 req->wb_context->path.dentry->d_inode->i_sb->s_id,
154 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
156 (long long)req_offset(req));
157 nfs_release_request(req);
160 int nfs_initiate_read(struct nfs_read_data *data, struct rpc_clnt *clnt,
161 const struct rpc_call_ops *call_ops)
163 struct inode *inode = data->inode;
164 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
165 struct rpc_task *task;
166 struct rpc_message msg = {
167 .rpc_argp = &data->args,
168 .rpc_resp = &data->res,
169 .rpc_cred = data->cred,
171 struct rpc_task_setup task_setup_data = {
175 .callback_ops = call_ops,
176 .callback_data = data,
177 .workqueue = nfsiod_workqueue,
178 .flags = RPC_TASK_ASYNC | swap_flags,
181 /* Set up the initial task struct. */
182 NFS_PROTO(inode)->read_setup(data, &msg);
184 dprintk("NFS: %5u initiated read call (req %s/%lld, %u bytes @ "
188 (long long)NFS_FILEID(inode),
190 (unsigned long long)data->args.offset);
192 task = rpc_run_task(&task_setup_data);
194 return PTR_ERR(task);
198 EXPORT_SYMBOL_GPL(nfs_initiate_read);
201 * Set up the NFS read request struct
203 static int nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
204 const struct rpc_call_ops *call_ops,
205 unsigned int count, unsigned int offset,
206 struct pnfs_layout_segment *lseg)
208 struct inode *inode = req->wb_context->path.dentry->d_inode;
212 data->cred = req->wb_context->cred;
213 data->lseg = get_lseg(lseg);
215 data->args.fh = NFS_FH(inode);
216 data->args.offset = req_offset(req) + offset;
217 data->args.pgbase = req->wb_pgbase + offset;
218 data->args.pages = data->pagevec;
219 data->args.count = count;
220 data->args.context = get_nfs_open_context(req->wb_context);
221 data->args.lock_context = req->wb_lock_context;
223 data->res.fattr = &data->fattr;
224 data->res.count = count;
226 nfs_fattr_init(&data->fattr);
229 (pnfs_try_to_read_data(data, call_ops) == PNFS_ATTEMPTED))
232 return nfs_initiate_read(data, NFS_CLIENT(inode), call_ops);
236 nfs_async_read_error(struct list_head *head)
238 struct nfs_page *req;
240 while (!list_empty(head)) {
241 req = nfs_list_entry(head->next);
242 nfs_list_remove_request(req);
243 SetPageError(req->wb_page);
244 nfs_readpage_release(req);
249 * Generate multiple requests to fill a single page.
251 * We optimize to reduce the number of read operations on the wire. If we
252 * detect that we're reading a page, or an area of a page, that is past the
253 * end of file, we do not generate NFS read operations but just clear the
254 * parts of the page that would have come back zero from the server anyway.
256 * We rely on the cached value of i_size to make this determination; another
257 * client can fill pages on the server past our cached end-of-file, but we
258 * won't see the new data until our attribute cache is updated. This is more
259 * or less conventional NFS client behavior.
261 static int nfs_pagein_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags, struct pnfs_layout_segment *lseg)
263 struct nfs_page *req = nfs_list_entry(head->next);
264 struct page *page = req->wb_page;
265 struct nfs_read_data *data;
266 size_t rsize = NFS_SERVER(inode)->rsize, nbytes;
272 nfs_list_remove_request(req);
276 size_t len = min(nbytes,rsize);
278 data = nfs_readdata_alloc(1);
281 list_add(&data->pages, &list);
284 } while(nbytes != 0);
285 atomic_set(&req->wb_complete, requests);
287 /* We know lseg==NULL */
288 lseg = pnfs_update_layout(inode, req->wb_context, IOMODE_READ);
289 ClearPageError(page);
295 data = list_entry(list.next, struct nfs_read_data, pages);
296 list_del_init(&data->pages);
298 data->pagevec[0] = page;
302 ret2 = nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
303 rsize, offset, lseg);
308 } while (nbytes != 0);
314 while (!list_empty(&list)) {
315 data = list_entry(list.next, struct nfs_read_data, pages);
316 list_del(&data->pages);
317 nfs_readdata_free(data);
320 nfs_readpage_release(req);
324 static int nfs_pagein_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags, struct pnfs_layout_segment *lseg)
326 struct nfs_page *req;
328 struct nfs_read_data *data;
331 data = nfs_readdata_alloc(npages);
333 nfs_async_read_error(head);
337 pages = data->pagevec;
338 while (!list_empty(head)) {
339 req = nfs_list_entry(head->next);
340 nfs_list_remove_request(req);
341 nfs_list_add_request(req, &data->pages);
342 ClearPageError(req->wb_page);
343 *pages++ = req->wb_page;
345 req = nfs_list_entry(data->pages.next);
346 if ((!lseg) && list_is_singular(&data->pages))
347 lseg = pnfs_update_layout(inode, req->wb_context, IOMODE_READ);
349 ret = nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0, lseg);
356 * This is the callback from RPC telling us whether a reply was
357 * received or some error occurred (timeout or socket shutdown).
359 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
363 dprintk("NFS: %s: %5u, (status %d)\n", __func__, task->tk_pid,
366 status = NFS_PROTO(data->inode)->read_done(task, data);
370 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
372 if (task->tk_status == -ESTALE) {
373 set_bit(NFS_INO_STALE, &NFS_I(data->inode)->flags);
374 nfs_mark_for_revalidate(data->inode);
379 static void nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
381 struct nfs_readargs *argp = &data->args;
382 struct nfs_readres *resp = &data->res;
384 if (resp->eof || resp->count == argp->count)
387 /* This is a short read! */
388 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
389 /* Has the server at least made some progress? */
390 if (resp->count == 0)
393 /* Yes, so retry the read at the end of the data */
394 argp->offset += resp->count;
395 argp->pgbase += resp->count;
396 argp->count -= resp->count;
397 nfs_restart_rpc(task, NFS_SERVER(data->inode)->nfs_client);
401 * Handle a read reply that fills part of a page.
403 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
405 struct nfs_read_data *data = calldata;
407 if (nfs_readpage_result(task, data) != 0)
409 if (task->tk_status < 0)
412 nfs_readpage_truncate_uninitialised_page(data);
413 nfs_readpage_retry(task, data);
416 static void nfs_readpage_release_partial(void *calldata)
418 struct nfs_read_data *data = calldata;
419 struct nfs_page *req = data->req;
420 struct page *page = req->wb_page;
421 int status = data->task.tk_status;
426 if (atomic_dec_and_test(&req->wb_complete)) {
427 if (!PageError(page))
428 SetPageUptodate(page);
429 nfs_readpage_release(req);
431 nfs_readdata_release(calldata);
434 #if defined(CONFIG_NFS_V4_1)
435 void nfs_read_prepare(struct rpc_task *task, void *calldata)
437 struct nfs_read_data *data = calldata;
439 if (nfs4_setup_sequence(NFS_SERVER(data->inode),
440 &data->args.seq_args, &data->res.seq_res,
443 rpc_call_start(task);
445 #endif /* CONFIG_NFS_V4_1 */
447 static const struct rpc_call_ops nfs_read_partial_ops = {
448 #if defined(CONFIG_NFS_V4_1)
449 .rpc_call_prepare = nfs_read_prepare,
450 #endif /* CONFIG_NFS_V4_1 */
451 .rpc_call_done = nfs_readpage_result_partial,
452 .rpc_release = nfs_readpage_release_partial,
455 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
457 unsigned int count = data->res.count;
458 unsigned int base = data->args.pgbase;
462 count = data->args.count;
463 if (unlikely(count == 0))
465 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
466 base &= ~PAGE_CACHE_MASK;
468 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
469 SetPageUptodate(*pages);
472 /* Was this a short read? */
473 if (data->res.eof || data->res.count == data->args.count)
474 SetPageUptodate(*pages);
478 * This is the callback from RPC telling us whether a reply was
479 * received or some error occurred (timeout or socket shutdown).
481 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
483 struct nfs_read_data *data = calldata;
485 if (nfs_readpage_result(task, data) != 0)
487 if (task->tk_status < 0)
490 * Note: nfs_readpage_retry may change the values of
491 * data->args. In the multi-page case, we therefore need
492 * to ensure that we call nfs_readpage_set_pages_uptodate()
495 nfs_readpage_truncate_uninitialised_page(data);
496 nfs_readpage_set_pages_uptodate(data);
497 nfs_readpage_retry(task, data);
500 static void nfs_readpage_release_full(void *calldata)
502 struct nfs_read_data *data = calldata;
504 while (!list_empty(&data->pages)) {
505 struct nfs_page *req = nfs_list_entry(data->pages.next);
507 nfs_list_remove_request(req);
508 nfs_readpage_release(req);
510 nfs_readdata_release(calldata);
513 static const struct rpc_call_ops nfs_read_full_ops = {
514 #if defined(CONFIG_NFS_V4_1)
515 .rpc_call_prepare = nfs_read_prepare,
516 #endif /* CONFIG_NFS_V4_1 */
517 .rpc_call_done = nfs_readpage_result_full,
518 .rpc_release = nfs_readpage_release_full,
522 * Read a page over NFS.
523 * We read the page synchronously in the following case:
524 * - The error flag is set for this page. This happens only when a
525 * previous async read operation failed.
527 int nfs_readpage(struct file *file, struct page *page)
529 struct nfs_open_context *ctx;
530 struct inode *inode = page->mapping->host;
533 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
534 page, PAGE_CACHE_SIZE, page->index);
535 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
536 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
539 * Try to flush any pending writes to the file..
541 * NOTE! Because we own the page lock, there cannot
542 * be any new pending writes generated at this point
543 * for this page (other pages can be written to).
545 error = nfs_wb_page(inode, page);
548 if (PageUptodate(page))
552 if (NFS_STALE(inode))
557 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
561 ctx = get_nfs_open_context(nfs_file_open_context(file));
563 if (!IS_SYNC(inode)) {
564 error = nfs_readpage_from_fscache(ctx, inode, page);
569 error = nfs_readpage_async(ctx, inode, page);
572 put_nfs_open_context(ctx);
579 struct nfs_readdesc {
580 struct nfs_pageio_descriptor *pgio;
581 struct nfs_open_context *ctx;
585 readpage_async_filler(void *data, struct page *page)
587 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
588 struct inode *inode = page->mapping->host;
589 struct nfs_page *new;
593 len = nfs_page_length(page);
595 return nfs_return_empty_page(page);
597 new = nfs_create_request(desc->ctx, inode, page, 0, len);
601 if (len < PAGE_CACHE_SIZE)
602 zero_user_segment(page, len, PAGE_CACHE_SIZE);
603 if (!nfs_pageio_add_request(desc->pgio, new)) {
604 error = desc->pgio->pg_error;
609 error = PTR_ERR(new);
616 int nfs_readpages(struct file *filp, struct address_space *mapping,
617 struct list_head *pages, unsigned nr_pages)
619 struct nfs_pageio_descriptor pgio;
620 struct nfs_readdesc desc = {
623 struct inode *inode = mapping->host;
624 struct nfs_server *server = NFS_SERVER(inode);
625 size_t rsize = server->rsize;
626 unsigned long npages;
629 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
631 (long long)NFS_FILEID(inode),
633 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
635 if (NFS_STALE(inode))
639 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
640 if (desc.ctx == NULL)
643 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
645 /* attempt to read as many of the pages as possible from the cache
646 * - this returns -ENOBUFS immediately if the cookie is negative
648 ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
651 goto read_complete; /* all pages were read */
653 pnfs_pageio_init_read(&pgio, inode);
654 if (rsize < PAGE_CACHE_SIZE)
655 nfs_pageio_init(&pgio, inode, nfs_pagein_multi, rsize, 0);
657 nfs_pageio_init(&pgio, inode, nfs_pagein_one, rsize, 0);
659 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
661 nfs_pageio_complete(&pgio);
662 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
663 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
665 put_nfs_open_context(desc.ctx);
670 int __init nfs_init_readpagecache(void)
672 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
673 sizeof(struct nfs_read_data),
674 0, SLAB_HWCACHE_ALIGN,
676 if (nfs_rdata_cachep == NULL)
679 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
681 if (nfs_rdata_mempool == NULL)
687 void nfs_destroy_readpagecache(void)
689 mempool_destroy(nfs_rdata_mempool);
690 kmem_cache_destroy(nfs_rdata_cachep);