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>
23 #include <asm/system.h>
28 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
30 static int nfs_pagein_multi(struct inode *, struct list_head *, unsigned int, size_t, int);
31 static int nfs_pagein_one(struct inode *, struct list_head *, unsigned int, size_t, int);
32 static const struct rpc_call_ops nfs_read_partial_ops;
33 static const struct rpc_call_ops nfs_read_full_ops;
35 static struct kmem_cache *nfs_rdata_cachep;
36 static mempool_t *nfs_rdata_mempool;
38 #define MIN_POOL_READ (32)
40 struct nfs_read_data *nfs_readdata_alloc(unsigned int pagecount)
42 struct nfs_read_data *p = mempool_alloc(nfs_rdata_mempool, GFP_NOFS);
45 memset(p, 0, sizeof(*p));
46 INIT_LIST_HEAD(&p->pages);
47 p->npages = pagecount;
48 if (pagecount <= ARRAY_SIZE(p->page_array))
49 p->pagevec = p->page_array;
51 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
53 mempool_free(p, nfs_rdata_mempool);
61 static void nfs_readdata_rcu_free(struct rcu_head *head)
63 struct nfs_read_data *p = container_of(head, struct nfs_read_data, task.u.tk_rcu);
64 if (p && (p->pagevec != &p->page_array[0]))
66 mempool_free(p, nfs_rdata_mempool);
69 static void nfs_readdata_free(struct nfs_read_data *rdata)
71 call_rcu_bh(&rdata->task.u.tk_rcu, nfs_readdata_rcu_free);
74 void nfs_readdata_release(void *data)
76 struct nfs_read_data *rdata = data;
78 put_nfs_open_context(rdata->args.context);
79 nfs_readdata_free(rdata);
83 int nfs_return_empty_page(struct page *page)
85 zero_user(page, 0, PAGE_CACHE_SIZE);
86 SetPageUptodate(page);
91 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
93 unsigned int remainder = data->args.count - data->res.count;
94 unsigned int base = data->args.pgbase + data->res.count;
98 if (data->res.eof == 0 || remainder == 0)
101 * Note: "remainder" can never be negative, since we check for
102 * this in the XDR code.
104 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
105 base &= ~PAGE_CACHE_MASK;
106 pglen = PAGE_CACHE_SIZE - base;
108 if (remainder <= pglen) {
109 zero_user(*pages, base, remainder);
112 zero_user(*pages, base, pglen);
115 pglen = PAGE_CACHE_SIZE;
120 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
123 LIST_HEAD(one_request);
124 struct nfs_page *new;
127 len = nfs_page_length(page);
129 return nfs_return_empty_page(page);
130 new = nfs_create_request(ctx, inode, page, 0, len);
135 if (len < PAGE_CACHE_SIZE)
136 zero_user_segment(page, len, PAGE_CACHE_SIZE);
138 nfs_list_add_request(new, &one_request);
139 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
140 nfs_pagein_multi(inode, &one_request, 1, len, 0);
142 nfs_pagein_one(inode, &one_request, 1, len, 0);
146 static void nfs_readpage_release(struct nfs_page *req)
148 unlock_page(req->wb_page);
150 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
151 req->wb_context->path.dentry->d_inode->i_sb->s_id,
152 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
154 (long long)req_offset(req));
155 nfs_clear_request(req);
156 nfs_release_request(req);
160 * Set up the NFS read request struct
162 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
163 const struct rpc_call_ops *call_ops,
164 unsigned int count, unsigned int offset)
166 struct inode *inode = req->wb_context->path.dentry->d_inode;
167 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
168 struct rpc_task *task;
169 struct rpc_message msg = {
170 .rpc_argp = &data->args,
171 .rpc_resp = &data->res,
172 .rpc_cred = req->wb_context->cred,
174 struct rpc_task_setup task_setup_data = {
176 .rpc_client = NFS_CLIENT(inode),
178 .callback_ops = call_ops,
179 .callback_data = data,
180 .flags = RPC_TASK_ASYNC | swap_flags,
185 data->cred = msg.rpc_cred;
187 data->args.fh = NFS_FH(inode);
188 data->args.offset = req_offset(req) + offset;
189 data->args.pgbase = req->wb_pgbase + offset;
190 data->args.pages = data->pagevec;
191 data->args.count = count;
192 data->args.context = get_nfs_open_context(req->wb_context);
194 data->res.fattr = &data->fattr;
195 data->res.count = count;
197 nfs_fattr_init(&data->fattr);
199 /* Set up the initial task struct. */
200 NFS_PROTO(inode)->read_setup(data, &msg);
202 dprintk("NFS: %5u initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
205 (long long)NFS_FILEID(inode),
207 (unsigned long long)data->args.offset);
209 task = rpc_run_task(&task_setup_data);
215 nfs_async_read_error(struct list_head *head)
217 struct nfs_page *req;
219 while (!list_empty(head)) {
220 req = nfs_list_entry(head->next);
221 nfs_list_remove_request(req);
222 SetPageError(req->wb_page);
223 nfs_readpage_release(req);
228 * Generate multiple requests to fill a single page.
230 * We optimize to reduce the number of read operations on the wire. If we
231 * detect that we're reading a page, or an area of a page, that is past the
232 * end of file, we do not generate NFS read operations but just clear the
233 * parts of the page that would have come back zero from the server anyway.
235 * We rely on the cached value of i_size to make this determination; another
236 * client can fill pages on the server past our cached end-of-file, but we
237 * won't see the new data until our attribute cache is updated. This is more
238 * or less conventional NFS client behavior.
240 static int nfs_pagein_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags)
242 struct nfs_page *req = nfs_list_entry(head->next);
243 struct page *page = req->wb_page;
244 struct nfs_read_data *data;
245 size_t rsize = NFS_SERVER(inode)->rsize, nbytes;
250 nfs_list_remove_request(req);
254 size_t len = min(nbytes,rsize);
256 data = nfs_readdata_alloc(1);
259 INIT_LIST_HEAD(&data->pages);
260 list_add(&data->pages, &list);
263 } while(nbytes != 0);
264 atomic_set(&req->wb_complete, requests);
266 ClearPageError(page);
270 data = list_entry(list.next, struct nfs_read_data, pages);
271 list_del_init(&data->pages);
273 data->pagevec[0] = page;
277 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
281 } while (nbytes != 0);
286 while (!list_empty(&list)) {
287 data = list_entry(list.next, struct nfs_read_data, pages);
288 list_del(&data->pages);
289 nfs_readdata_free(data);
292 nfs_readpage_release(req);
296 static int nfs_pagein_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags)
298 struct nfs_page *req;
300 struct nfs_read_data *data;
302 data = nfs_readdata_alloc(npages);
306 INIT_LIST_HEAD(&data->pages);
307 pages = data->pagevec;
308 while (!list_empty(head)) {
309 req = nfs_list_entry(head->next);
310 nfs_list_remove_request(req);
311 nfs_list_add_request(req, &data->pages);
312 ClearPageError(req->wb_page);
313 *pages++ = req->wb_page;
315 req = nfs_list_entry(data->pages.next);
317 nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
320 nfs_async_read_error(head);
325 * This is the callback from RPC telling us whether a reply was
326 * received or some error occurred (timeout or socket shutdown).
328 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
332 dprintk("NFS: %s: %5u, (status %d)\n", __FUNCTION__, task->tk_pid,
335 status = NFS_PROTO(data->inode)->read_done(task, data);
339 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
341 if (task->tk_status == -ESTALE) {
342 set_bit(NFS_INO_STALE, &NFS_I(data->inode)->flags);
343 nfs_mark_for_revalidate(data->inode);
348 static int nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
350 struct nfs_readargs *argp = &data->args;
351 struct nfs_readres *resp = &data->res;
353 if (resp->eof || resp->count == argp->count)
356 /* This is a short read! */
357 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
358 /* Has the server at least made some progress? */
359 if (resp->count == 0)
362 /* Yes, so retry the read at the end of the data */
363 argp->offset += resp->count;
364 argp->pgbase += resp->count;
365 argp->count -= resp->count;
366 rpc_restart_call(task);
371 * Handle a read reply that fills part of a page.
373 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
375 struct nfs_read_data *data = calldata;
376 struct nfs_page *req = data->req;
377 struct page *page = req->wb_page;
379 if (nfs_readpage_result(task, data) != 0)
382 if (likely(task->tk_status >= 0)) {
383 nfs_readpage_truncate_uninitialised_page(data);
384 if (nfs_readpage_retry(task, data) != 0)
387 if (unlikely(task->tk_status < 0))
389 if (atomic_dec_and_test(&req->wb_complete)) {
390 if (!PageError(page))
391 SetPageUptodate(page);
392 nfs_readpage_release(req);
396 static const struct rpc_call_ops nfs_read_partial_ops = {
397 .rpc_call_done = nfs_readpage_result_partial,
398 .rpc_release = nfs_readdata_release,
401 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
403 unsigned int count = data->res.count;
404 unsigned int base = data->args.pgbase;
408 count = data->args.count;
409 if (unlikely(count == 0))
411 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
412 base &= ~PAGE_CACHE_MASK;
414 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
415 SetPageUptodate(*pages);
418 /* Was this a short read? */
419 if (data->res.eof || data->res.count == data->args.count)
420 SetPageUptodate(*pages);
424 * This is the callback from RPC telling us whether a reply was
425 * received or some error occurred (timeout or socket shutdown).
427 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
429 struct nfs_read_data *data = calldata;
431 if (nfs_readpage_result(task, data) != 0)
434 * Note: nfs_readpage_retry may change the values of
435 * data->args. In the multi-page case, we therefore need
436 * to ensure that we call nfs_readpage_set_pages_uptodate()
439 if (likely(task->tk_status >= 0)) {
440 nfs_readpage_truncate_uninitialised_page(data);
441 nfs_readpage_set_pages_uptodate(data);
442 if (nfs_readpage_retry(task, data) != 0)
445 while (!list_empty(&data->pages)) {
446 struct nfs_page *req = nfs_list_entry(data->pages.next);
448 nfs_list_remove_request(req);
449 nfs_readpage_release(req);
453 static const struct rpc_call_ops nfs_read_full_ops = {
454 .rpc_call_done = nfs_readpage_result_full,
455 .rpc_release = nfs_readdata_release,
459 * Read a page over NFS.
460 * We read the page synchronously in the following case:
461 * - The error flag is set for this page. This happens only when a
462 * previous async read operation failed.
464 int nfs_readpage(struct file *file, struct page *page)
466 struct nfs_open_context *ctx;
467 struct inode *inode = page->mapping->host;
470 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
471 page, PAGE_CACHE_SIZE, page->index);
472 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
473 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
476 * Try to flush any pending writes to the file..
478 * NOTE! Because we own the page lock, there cannot
479 * be any new pending writes generated at this point
480 * for this page (other pages can be written to).
482 error = nfs_wb_page(inode, page);
485 if (PageUptodate(page))
489 if (NFS_STALE(inode))
494 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
498 ctx = get_nfs_open_context(nfs_file_open_context(file));
500 error = nfs_readpage_async(ctx, inode, page);
502 put_nfs_open_context(ctx);
509 struct nfs_readdesc {
510 struct nfs_pageio_descriptor *pgio;
511 struct nfs_open_context *ctx;
515 readpage_async_filler(void *data, struct page *page)
517 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
518 struct inode *inode = page->mapping->host;
519 struct nfs_page *new;
523 error = nfs_wb_page(inode, page);
526 if (PageUptodate(page))
529 len = nfs_page_length(page);
531 return nfs_return_empty_page(page);
533 new = nfs_create_request(desc->ctx, inode, page, 0, len);
537 if (len < PAGE_CACHE_SIZE)
538 zero_user_segment(page, len, PAGE_CACHE_SIZE);
539 nfs_pageio_add_request(desc->pgio, new);
542 error = PTR_ERR(new);
549 int nfs_readpages(struct file *filp, struct address_space *mapping,
550 struct list_head *pages, unsigned nr_pages)
552 struct nfs_pageio_descriptor pgio;
553 struct nfs_readdesc desc = {
556 struct inode *inode = mapping->host;
557 struct nfs_server *server = NFS_SERVER(inode);
558 size_t rsize = server->rsize;
559 unsigned long npages;
562 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
564 (long long)NFS_FILEID(inode),
566 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
568 if (NFS_STALE(inode))
572 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
573 if (desc.ctx == NULL)
576 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
577 if (rsize < PAGE_CACHE_SIZE)
578 nfs_pageio_init(&pgio, inode, nfs_pagein_multi, rsize, 0);
580 nfs_pageio_init(&pgio, inode, nfs_pagein_one, rsize, 0);
582 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
584 nfs_pageio_complete(&pgio);
585 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
586 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
587 put_nfs_open_context(desc.ctx);
592 int __init nfs_init_readpagecache(void)
594 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
595 sizeof(struct nfs_read_data),
596 0, SLAB_HWCACHE_ALIGN,
598 if (nfs_rdata_cachep == NULL)
601 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
603 if (nfs_rdata_mempool == NULL)
609 void nfs_destroy_readpagecache(void)
611 mempool_destroy(nfs_rdata_mempool);
612 kmem_cache_destroy(nfs_rdata_cachep);