6 * Partial copy of Linus' read cache modifications to fs/nfs/file.c
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>
23 #include <asm/system.h>
31 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
33 static int nfs_pagein_multi(struct nfs_pageio_descriptor *desc);
34 static int nfs_pagein_one(struct nfs_pageio_descriptor *desc);
35 static const struct rpc_call_ops nfs_read_partial_ops;
36 static const struct rpc_call_ops nfs_read_full_ops;
38 static struct kmem_cache *nfs_rdata_cachep;
39 static mempool_t *nfs_rdata_mempool;
41 #define MIN_POOL_READ (32)
43 struct nfs_read_data *nfs_readdata_alloc(unsigned int pagecount)
45 struct nfs_read_data *p = mempool_alloc(nfs_rdata_mempool, GFP_KERNEL);
48 memset(p, 0, sizeof(*p));
49 INIT_LIST_HEAD(&p->pages);
50 p->npages = pagecount;
51 if (pagecount <= ARRAY_SIZE(p->page_array))
52 p->pagevec = p->page_array;
54 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_KERNEL);
56 mempool_free(p, nfs_rdata_mempool);
64 void nfs_readdata_free(struct nfs_read_data *p)
66 if (p && (p->pagevec != &p->page_array[0]))
68 mempool_free(p, nfs_rdata_mempool);
71 static void nfs_readdata_release(struct nfs_read_data *rdata)
73 put_lseg(rdata->lseg);
74 put_nfs_open_context(rdata->args.context);
75 nfs_readdata_free(rdata);
79 int nfs_return_empty_page(struct page *page)
81 zero_user(page, 0, PAGE_CACHE_SIZE);
82 SetPageUptodate(page);
87 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
89 unsigned int remainder = data->args.count - data->res.count;
90 unsigned int base = data->args.pgbase + data->res.count;
94 if (data->res.eof == 0 || remainder == 0)
97 * Note: "remainder" can never be negative, since we check for
98 * this in the XDR code.
100 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
101 base &= ~PAGE_CACHE_MASK;
102 pglen = PAGE_CACHE_SIZE - base;
104 if (remainder <= pglen) {
105 zero_user(*pages, base, remainder);
108 zero_user(*pages, base, pglen);
111 pglen = PAGE_CACHE_SIZE;
116 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
119 struct nfs_page *new;
121 struct nfs_pageio_descriptor pgio;
123 len = nfs_page_length(page);
125 return nfs_return_empty_page(page);
126 new = nfs_create_request(ctx, inode, page, 0, len);
131 if (len < PAGE_CACHE_SIZE)
132 zero_user_segment(page, len, PAGE_CACHE_SIZE);
134 nfs_pageio_init(&pgio, inode, NULL, 0, 0);
135 nfs_list_add_request(new, &pgio.pg_list);
138 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
139 nfs_pagein_multi(&pgio);
141 nfs_pagein_one(&pgio);
145 static void nfs_readpage_release(struct nfs_page *req)
147 struct inode *d_inode = req->wb_context->dentry->d_inode;
149 if (PageUptodate(req->wb_page))
150 nfs_readpage_to_fscache(d_inode, req->wb_page, 0);
152 unlock_page(req->wb_page);
154 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
155 req->wb_context->dentry->d_inode->i_sb->s_id,
156 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
158 (long long)req_offset(req));
159 nfs_release_request(req);
162 int nfs_initiate_read(struct nfs_read_data *data, struct rpc_clnt *clnt,
163 const struct rpc_call_ops *call_ops)
165 struct inode *inode = data->inode;
166 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
167 struct rpc_task *task;
168 struct rpc_message msg = {
169 .rpc_argp = &data->args,
170 .rpc_resp = &data->res,
171 .rpc_cred = data->cred,
173 struct rpc_task_setup task_setup_data = {
177 .callback_ops = call_ops,
178 .callback_data = data,
179 .workqueue = nfsiod_workqueue,
180 .flags = RPC_TASK_ASYNC | swap_flags,
183 /* Set up the initial task struct. */
184 NFS_PROTO(inode)->read_setup(data, &msg);
186 dprintk("NFS: %5u initiated read call (req %s/%lld, %u bytes @ "
190 (long long)NFS_FILEID(inode),
192 (unsigned long long)data->args.offset);
194 task = rpc_run_task(&task_setup_data);
196 return PTR_ERR(task);
200 EXPORT_SYMBOL_GPL(nfs_initiate_read);
203 * Set up the NFS read request struct
205 static int nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
206 const struct rpc_call_ops *call_ops,
207 unsigned int count, unsigned int offset,
208 struct pnfs_layout_segment *lseg)
210 struct inode *inode = req->wb_context->dentry->d_inode;
214 data->cred = req->wb_context->cred;
215 data->lseg = get_lseg(lseg);
217 data->args.fh = NFS_FH(inode);
218 data->args.offset = req_offset(req) + offset;
219 data->args.pgbase = req->wb_pgbase + offset;
220 data->args.pages = data->pagevec;
221 data->args.count = count;
222 data->args.context = get_nfs_open_context(req->wb_context);
223 data->args.lock_context = req->wb_lock_context;
225 data->res.fattr = &data->fattr;
226 data->res.count = count;
228 nfs_fattr_init(&data->fattr);
231 (pnfs_try_to_read_data(data, call_ops) == PNFS_ATTEMPTED))
234 return nfs_initiate_read(data, NFS_CLIENT(inode), call_ops);
238 nfs_async_read_error(struct list_head *head)
240 struct nfs_page *req;
242 while (!list_empty(head)) {
243 req = nfs_list_entry(head->next);
244 nfs_list_remove_request(req);
245 SetPageError(req->wb_page);
246 nfs_readpage_release(req);
251 * Generate multiple requests to fill a single page.
253 * We optimize to reduce the number of read operations on the wire. If we
254 * detect that we're reading a page, or an area of a page, that is past the
255 * end of file, we do not generate NFS read operations but just clear the
256 * parts of the page that would have come back zero from the server anyway.
258 * We rely on the cached value of i_size to make this determination; another
259 * client can fill pages on the server past our cached end-of-file, but we
260 * won't see the new data until our attribute cache is updated. This is more
261 * or less conventional NFS client behavior.
263 static int nfs_pagein_multi(struct nfs_pageio_descriptor *desc)
265 struct nfs_page *req = nfs_list_entry(desc->pg_list.next);
266 struct page *page = req->wb_page;
267 struct nfs_read_data *data;
268 size_t rsize = NFS_SERVER(desc->pg_inode)->rsize, nbytes;
272 struct pnfs_layout_segment *lseg;
275 nfs_list_remove_request(req);
277 nbytes = desc->pg_count;
279 size_t len = min(nbytes,rsize);
281 data = nfs_readdata_alloc(1);
284 list_add(&data->pages, &list);
287 } while(nbytes != 0);
288 atomic_set(&req->wb_complete, requests);
290 BUG_ON(desc->pg_lseg != NULL);
291 lseg = pnfs_update_layout(desc->pg_inode, req->wb_context,
292 req_offset(req), desc->pg_count,
293 IOMODE_READ, GFP_KERNEL);
294 ClearPageError(page);
296 nbytes = desc->pg_count;
300 data = list_entry(list.next, struct nfs_read_data, pages);
301 list_del_init(&data->pages);
303 data->pagevec[0] = page;
307 ret2 = nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
308 rsize, offset, lseg);
313 } while (nbytes != 0);
315 desc->pg_lseg = NULL;
320 while (!list_empty(&list)) {
321 data = list_entry(list.next, struct nfs_read_data, pages);
322 list_del(&data->pages);
323 nfs_readdata_free(data);
326 nfs_readpage_release(req);
330 static int nfs_pagein_one(struct nfs_pageio_descriptor *desc)
332 struct nfs_page *req;
334 struct nfs_read_data *data;
335 struct list_head *head = &desc->pg_list;
336 struct pnfs_layout_segment *lseg = desc->pg_lseg;
339 data = nfs_readdata_alloc(nfs_page_array_len(desc->pg_base,
342 nfs_async_read_error(head);
346 pages = data->pagevec;
347 while (!list_empty(head)) {
348 req = nfs_list_entry(head->next);
349 nfs_list_remove_request(req);
350 nfs_list_add_request(req, &data->pages);
351 ClearPageError(req->wb_page);
352 *pages++ = req->wb_page;
354 req = nfs_list_entry(data->pages.next);
355 if ((!lseg) && list_is_singular(&data->pages))
356 lseg = pnfs_update_layout(desc->pg_inode, req->wb_context,
357 req_offset(req), desc->pg_count,
358 IOMODE_READ, GFP_KERNEL);
360 ret = nfs_read_rpcsetup(req, data, &nfs_read_full_ops, desc->pg_count,
364 desc->pg_lseg = NULL;
369 * This is the callback from RPC telling us whether a reply was
370 * received or some error occurred (timeout or socket shutdown).
372 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
376 dprintk("NFS: %s: %5u, (status %d)\n", __func__, task->tk_pid,
379 status = NFS_PROTO(data->inode)->read_done(task, data);
383 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
385 if (task->tk_status == -ESTALE) {
386 set_bit(NFS_INO_STALE, &NFS_I(data->inode)->flags);
387 nfs_mark_for_revalidate(data->inode);
392 static void nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
394 struct nfs_readargs *argp = &data->args;
395 struct nfs_readres *resp = &data->res;
397 if (resp->eof || resp->count == argp->count)
400 /* This is a short read! */
401 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
402 /* Has the server at least made some progress? */
403 if (resp->count == 0)
406 /* Yes, so retry the read at the end of the data */
407 data->mds_offset += resp->count;
408 argp->offset += resp->count;
409 argp->pgbase += resp->count;
410 argp->count -= resp->count;
411 nfs_restart_rpc(task, NFS_SERVER(data->inode)->nfs_client);
415 * Handle a read reply that fills part of a page.
417 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
419 struct nfs_read_data *data = calldata;
421 if (nfs_readpage_result(task, data) != 0)
423 if (task->tk_status < 0)
426 nfs_readpage_truncate_uninitialised_page(data);
427 nfs_readpage_retry(task, data);
430 static void nfs_readpage_release_partial(void *calldata)
432 struct nfs_read_data *data = calldata;
433 struct nfs_page *req = data->req;
434 struct page *page = req->wb_page;
435 int status = data->task.tk_status;
440 if (atomic_dec_and_test(&req->wb_complete)) {
441 if (!PageError(page))
442 SetPageUptodate(page);
443 nfs_readpage_release(req);
445 nfs_readdata_release(calldata);
448 #if defined(CONFIG_NFS_V4_1)
449 void nfs_read_prepare(struct rpc_task *task, void *calldata)
451 struct nfs_read_data *data = calldata;
453 if (nfs4_setup_sequence(NFS_SERVER(data->inode),
454 &data->args.seq_args, &data->res.seq_res,
457 rpc_call_start(task);
459 #endif /* CONFIG_NFS_V4_1 */
461 static const struct rpc_call_ops nfs_read_partial_ops = {
462 #if defined(CONFIG_NFS_V4_1)
463 .rpc_call_prepare = nfs_read_prepare,
464 #endif /* CONFIG_NFS_V4_1 */
465 .rpc_call_done = nfs_readpage_result_partial,
466 .rpc_release = nfs_readpage_release_partial,
469 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
471 unsigned int count = data->res.count;
472 unsigned int base = data->args.pgbase;
476 count = data->args.count;
477 if (unlikely(count == 0))
479 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
480 base &= ~PAGE_CACHE_MASK;
482 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
483 SetPageUptodate(*pages);
486 /* Was this a short read? */
487 if (data->res.eof || data->res.count == data->args.count)
488 SetPageUptodate(*pages);
492 * This is the callback from RPC telling us whether a reply was
493 * received or some error occurred (timeout or socket shutdown).
495 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
497 struct nfs_read_data *data = calldata;
499 if (nfs_readpage_result(task, data) != 0)
501 if (task->tk_status < 0)
504 * Note: nfs_readpage_retry may change the values of
505 * data->args. In the multi-page case, we therefore need
506 * to ensure that we call nfs_readpage_set_pages_uptodate()
509 nfs_readpage_truncate_uninitialised_page(data);
510 nfs_readpage_set_pages_uptodate(data);
511 nfs_readpage_retry(task, data);
514 static void nfs_readpage_release_full(void *calldata)
516 struct nfs_read_data *data = calldata;
518 while (!list_empty(&data->pages)) {
519 struct nfs_page *req = nfs_list_entry(data->pages.next);
521 nfs_list_remove_request(req);
522 nfs_readpage_release(req);
524 nfs_readdata_release(calldata);
527 static const struct rpc_call_ops nfs_read_full_ops = {
528 #if defined(CONFIG_NFS_V4_1)
529 .rpc_call_prepare = nfs_read_prepare,
530 #endif /* CONFIG_NFS_V4_1 */
531 .rpc_call_done = nfs_readpage_result_full,
532 .rpc_release = nfs_readpage_release_full,
536 * Read a page over NFS.
537 * We read the page synchronously in the following case:
538 * - The error flag is set for this page. This happens only when a
539 * previous async read operation failed.
541 int nfs_readpage(struct file *file, struct page *page)
543 struct nfs_open_context *ctx;
544 struct inode *inode = page->mapping->host;
547 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
548 page, PAGE_CACHE_SIZE, page->index);
549 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
550 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
553 * Try to flush any pending writes to the file..
555 * NOTE! Because we own the page lock, there cannot
556 * be any new pending writes generated at this point
557 * for this page (other pages can be written to).
559 error = nfs_wb_page(inode, page);
562 if (PageUptodate(page))
566 if (NFS_STALE(inode))
571 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
575 ctx = get_nfs_open_context(nfs_file_open_context(file));
577 if (!IS_SYNC(inode)) {
578 error = nfs_readpage_from_fscache(ctx, inode, page);
583 error = nfs_readpage_async(ctx, inode, page);
586 put_nfs_open_context(ctx);
593 struct nfs_readdesc {
594 struct nfs_pageio_descriptor *pgio;
595 struct nfs_open_context *ctx;
599 readpage_async_filler(void *data, struct page *page)
601 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
602 struct inode *inode = page->mapping->host;
603 struct nfs_page *new;
607 len = nfs_page_length(page);
609 return nfs_return_empty_page(page);
611 new = nfs_create_request(desc->ctx, inode, page, 0, len);
615 if (len < PAGE_CACHE_SIZE)
616 zero_user_segment(page, len, PAGE_CACHE_SIZE);
617 if (!nfs_pageio_add_request(desc->pgio, new)) {
618 error = desc->pgio->pg_error;
623 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 struct nfs_server *server = NFS_SERVER(inode);
639 size_t rsize = server->rsize;
640 unsigned long npages;
643 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
645 (long long)NFS_FILEID(inode),
647 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
649 if (NFS_STALE(inode))
653 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
654 if (desc.ctx == NULL)
657 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
659 /* attempt to read as many of the pages as possible from the cache
660 * - this returns -ENOBUFS immediately if the cookie is negative
662 ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
665 goto read_complete; /* all pages were read */
667 if (rsize < PAGE_CACHE_SIZE)
668 nfs_pageio_init(&pgio, inode, nfs_pagein_multi, rsize, 0);
670 nfs_pageio_init(&pgio, inode, nfs_pagein_one, rsize, 0);
672 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
674 nfs_pageio_complete(&pgio);
675 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
676 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
678 put_nfs_open_context(desc.ctx);
683 int __init nfs_init_readpagecache(void)
685 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
686 sizeof(struct nfs_read_data),
687 0, SLAB_HWCACHE_ALIGN,
689 if (nfs_rdata_cachep == NULL)
692 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
694 if (nfs_rdata_mempool == NULL)
700 void nfs_destroy_readpagecache(void)
702 mempool_destroy(nfs_rdata_mempool);
703 kmem_cache_destroy(nfs_rdata_cachep);