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 nfs_readdata_free(data);
80 int nfs_return_empty_page(struct page *page)
82 zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
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_page(*pages, base, remainder, KM_USER0);
109 zero_user_page(*pages, base, pglen, KM_USER0);
112 pglen = PAGE_CACHE_SIZE;
117 static 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_page(page, len, PAGE_CACHE_SIZE - len, KM_USER0);
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);
139 nfs_pagein_one(inode, &one_request, 1, len, 0);
143 static void nfs_readpage_release(struct nfs_page *req)
145 unlock_page(req->wb_page);
147 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
148 req->wb_context->path.dentry->d_inode->i_sb->s_id,
149 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
151 (long long)req_offset(req));
152 nfs_clear_request(req);
153 nfs_release_request(req);
156 static void nfs_execute_read(struct nfs_read_data *data)
158 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
161 rpc_clnt_sigmask(clnt, &oldset);
162 rpc_execute(&data->task);
163 rpc_clnt_sigunmask(clnt, &oldset);
167 * Set up the NFS read request struct
169 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
170 const struct rpc_call_ops *call_ops,
171 unsigned int count, unsigned int offset)
173 struct inode *inode = req->wb_context->path.dentry->d_inode;
174 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
175 struct rpc_message msg = {
176 .rpc_argp = &data->args,
177 .rpc_resp = &data->res,
178 .rpc_cred = req->wb_context->cred,
180 struct rpc_task_setup task_setup_data = {
181 .rpc_client = NFS_CLIENT(inode),
183 .callback_ops = call_ops,
184 .callback_data = data,
185 .flags = RPC_TASK_ASYNC | swap_flags,
190 data->cred = msg.rpc_cred;
192 data->args.fh = NFS_FH(inode);
193 data->args.offset = req_offset(req) + offset;
194 data->args.pgbase = req->wb_pgbase + offset;
195 data->args.pages = data->pagevec;
196 data->args.count = count;
197 data->args.context = req->wb_context;
199 data->res.fattr = &data->fattr;
200 data->res.count = count;
202 nfs_fattr_init(&data->fattr);
204 /* Set up the initial task struct. */
205 NFS_PROTO(inode)->read_setup(data, &msg);
206 rpc_init_task(&data->task, &task_setup_data);
208 dprintk("NFS: %5u initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
211 (long long)NFS_FILEID(inode),
213 (unsigned long long)data->args.offset);
215 nfs_execute_read(data);
219 nfs_async_read_error(struct list_head *head)
221 struct nfs_page *req;
223 while (!list_empty(head)) {
224 req = nfs_list_entry(head->next);
225 nfs_list_remove_request(req);
226 SetPageError(req->wb_page);
227 nfs_readpage_release(req);
232 * Generate multiple requests to fill a single page.
234 * We optimize to reduce the number of read operations on the wire. If we
235 * detect that we're reading a page, or an area of a page, that is past the
236 * end of file, we do not generate NFS read operations but just clear the
237 * parts of the page that would have come back zero from the server anyway.
239 * We rely on the cached value of i_size to make this determination; another
240 * client can fill pages on the server past our cached end-of-file, but we
241 * won't see the new data until our attribute cache is updated. This is more
242 * or less conventional NFS client behavior.
244 static int nfs_pagein_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags)
246 struct nfs_page *req = nfs_list_entry(head->next);
247 struct page *page = req->wb_page;
248 struct nfs_read_data *data;
249 size_t rsize = NFS_SERVER(inode)->rsize, nbytes;
254 nfs_list_remove_request(req);
258 size_t len = min(nbytes,rsize);
260 data = nfs_readdata_alloc(1);
263 INIT_LIST_HEAD(&data->pages);
264 list_add(&data->pages, &list);
267 } while(nbytes != 0);
268 atomic_set(&req->wb_complete, requests);
270 ClearPageError(page);
274 data = list_entry(list.next, struct nfs_read_data, pages);
275 list_del_init(&data->pages);
277 data->pagevec[0] = page;
281 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
285 } while (nbytes != 0);
290 while (!list_empty(&list)) {
291 data = list_entry(list.next, struct nfs_read_data, pages);
292 list_del(&data->pages);
293 nfs_readdata_free(data);
296 nfs_readpage_release(req);
300 static int nfs_pagein_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags)
302 struct nfs_page *req;
304 struct nfs_read_data *data;
306 data = nfs_readdata_alloc(npages);
310 INIT_LIST_HEAD(&data->pages);
311 pages = data->pagevec;
312 while (!list_empty(head)) {
313 req = nfs_list_entry(head->next);
314 nfs_list_remove_request(req);
315 nfs_list_add_request(req, &data->pages);
316 ClearPageError(req->wb_page);
317 *pages++ = req->wb_page;
319 req = nfs_list_entry(data->pages.next);
321 nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
324 nfs_async_read_error(head);
329 * This is the callback from RPC telling us whether a reply was
330 * received or some error occurred (timeout or socket shutdown).
332 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
336 dprintk("NFS: %s: %5u, (status %d)\n", __FUNCTION__, task->tk_pid,
339 status = NFS_PROTO(data->inode)->read_done(task, data);
343 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
345 if (task->tk_status == -ESTALE) {
346 set_bit(NFS_INO_STALE, &NFS_FLAGS(data->inode));
347 nfs_mark_for_revalidate(data->inode);
352 static int nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
354 struct nfs_readargs *argp = &data->args;
355 struct nfs_readres *resp = &data->res;
357 if (resp->eof || resp->count == argp->count)
360 /* This is a short read! */
361 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
362 /* Has the server at least made some progress? */
363 if (resp->count == 0)
366 /* Yes, so retry the read at the end of the data */
367 argp->offset += resp->count;
368 argp->pgbase += resp->count;
369 argp->count -= resp->count;
370 rpc_restart_call(task);
375 * Handle a read reply that fills part of a page.
377 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
379 struct nfs_read_data *data = calldata;
380 struct nfs_page *req = data->req;
381 struct page *page = req->wb_page;
383 if (nfs_readpage_result(task, data) != 0)
386 if (likely(task->tk_status >= 0)) {
387 nfs_readpage_truncate_uninitialised_page(data);
388 if (nfs_readpage_retry(task, data) != 0)
391 if (unlikely(task->tk_status < 0))
393 if (atomic_dec_and_test(&req->wb_complete)) {
394 if (!PageError(page))
395 SetPageUptodate(page);
396 nfs_readpage_release(req);
400 static const struct rpc_call_ops nfs_read_partial_ops = {
401 .rpc_call_done = nfs_readpage_result_partial,
402 .rpc_release = nfs_readdata_release,
405 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
407 unsigned int count = data->res.count;
408 unsigned int base = data->args.pgbase;
412 count = data->args.count;
413 if (unlikely(count == 0))
415 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
416 base &= ~PAGE_CACHE_MASK;
418 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
419 SetPageUptodate(*pages);
422 /* Was this a short read? */
423 if (data->res.eof || data->res.count == data->args.count)
424 SetPageUptodate(*pages);
428 * This is the callback from RPC telling us whether a reply was
429 * received or some error occurred (timeout or socket shutdown).
431 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
433 struct nfs_read_data *data = calldata;
435 if (nfs_readpage_result(task, data) != 0)
438 * Note: nfs_readpage_retry may change the values of
439 * data->args. In the multi-page case, we therefore need
440 * to ensure that we call nfs_readpage_set_pages_uptodate()
443 if (likely(task->tk_status >= 0)) {
444 nfs_readpage_truncate_uninitialised_page(data);
445 nfs_readpage_set_pages_uptodate(data);
446 if (nfs_readpage_retry(task, data) != 0)
449 while (!list_empty(&data->pages)) {
450 struct nfs_page *req = nfs_list_entry(data->pages.next);
452 nfs_list_remove_request(req);
453 nfs_readpage_release(req);
457 static const struct rpc_call_ops nfs_read_full_ops = {
458 .rpc_call_done = nfs_readpage_result_full,
459 .rpc_release = nfs_readdata_release,
463 * Read a page over NFS.
464 * We read the page synchronously in the following case:
465 * - The error flag is set for this page. This happens only when a
466 * previous async read operation failed.
468 int nfs_readpage(struct file *file, struct page *page)
470 struct nfs_open_context *ctx;
471 struct inode *inode = page->mapping->host;
474 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
475 page, PAGE_CACHE_SIZE, page->index);
476 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
477 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
480 * Try to flush any pending writes to the file..
482 * NOTE! Because we own the page lock, there cannot
483 * be any new pending writes generated at this point
484 * for this page (other pages can be written to).
486 error = nfs_wb_page(inode, page);
489 if (PageUptodate(page))
493 if (NFS_STALE(inode))
498 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
502 ctx = get_nfs_open_context(nfs_file_open_context(file));
504 error = nfs_readpage_async(ctx, inode, page);
506 put_nfs_open_context(ctx);
513 struct nfs_readdesc {
514 struct nfs_pageio_descriptor *pgio;
515 struct nfs_open_context *ctx;
519 readpage_async_filler(void *data, struct page *page)
521 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
522 struct inode *inode = page->mapping->host;
523 struct nfs_page *new;
527 error = nfs_wb_page(inode, page);
530 if (PageUptodate(page))
533 len = nfs_page_length(page);
535 return nfs_return_empty_page(page);
537 new = nfs_create_request(desc->ctx, inode, page, 0, len);
541 if (len < PAGE_CACHE_SIZE)
542 zero_user_page(page, len, PAGE_CACHE_SIZE - len, KM_USER0);
543 nfs_pageio_add_request(desc->pgio, new);
546 error = PTR_ERR(new);
553 int nfs_readpages(struct file *filp, struct address_space *mapping,
554 struct list_head *pages, unsigned nr_pages)
556 struct nfs_pageio_descriptor pgio;
557 struct nfs_readdesc desc = {
560 struct inode *inode = mapping->host;
561 struct nfs_server *server = NFS_SERVER(inode);
562 size_t rsize = server->rsize;
563 unsigned long npages;
566 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
568 (long long)NFS_FILEID(inode),
570 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
572 if (NFS_STALE(inode))
576 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
577 if (desc.ctx == NULL)
580 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
581 if (rsize < PAGE_CACHE_SIZE)
582 nfs_pageio_init(&pgio, inode, nfs_pagein_multi, rsize, 0);
584 nfs_pageio_init(&pgio, inode, nfs_pagein_one, rsize, 0);
586 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
588 nfs_pageio_complete(&pgio);
589 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
590 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
591 put_nfs_open_context(desc.ctx);
596 int __init nfs_init_readpagecache(void)
598 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
599 sizeof(struct nfs_read_data),
600 0, SLAB_HWCACHE_ALIGN,
602 if (nfs_rdata_cachep == NULL)
605 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
607 if (nfs_rdata_mempool == NULL)
613 void nfs_destroy_readpagecache(void)
615 mempool_destroy(nfs_rdata_mempool);
616 kmem_cache_destroy(nfs_rdata_cachep);