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_free(struct nfs_read_data *p)
63 if (p && (p->pagevec != &p->page_array[0]))
65 mempool_free(p, nfs_rdata_mempool);
68 void nfs_readdata_release(void *data)
70 struct nfs_read_data *rdata = data;
72 put_nfs_open_context(rdata->args.context);
73 nfs_readdata_free(rdata);
77 int nfs_return_empty_page(struct page *page)
79 zero_user(page, 0, PAGE_CACHE_SIZE);
80 SetPageUptodate(page);
85 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
87 unsigned int remainder = data->args.count - data->res.count;
88 unsigned int base = data->args.pgbase + data->res.count;
92 if (data->res.eof == 0 || remainder == 0)
95 * Note: "remainder" can never be negative, since we check for
96 * this in the XDR code.
98 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
99 base &= ~PAGE_CACHE_MASK;
100 pglen = PAGE_CACHE_SIZE - base;
102 if (remainder <= pglen) {
103 zero_user(*pages, base, remainder);
106 zero_user(*pages, base, pglen);
109 pglen = PAGE_CACHE_SIZE;
114 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
117 LIST_HEAD(one_request);
118 struct nfs_page *new;
121 len = nfs_page_length(page);
123 return nfs_return_empty_page(page);
124 new = nfs_create_request(ctx, inode, page, 0, len);
129 if (len < PAGE_CACHE_SIZE)
130 zero_user_segment(page, len, PAGE_CACHE_SIZE);
132 nfs_list_add_request(new, &one_request);
133 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
134 nfs_pagein_multi(inode, &one_request, 1, len, 0);
136 nfs_pagein_one(inode, &one_request, 1, len, 0);
140 static void nfs_readpage_release(struct nfs_page *req)
142 unlock_page(req->wb_page);
144 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
145 req->wb_context->path.dentry->d_inode->i_sb->s_id,
146 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
148 (long long)req_offset(req));
149 nfs_clear_request(req);
150 nfs_release_request(req);
154 * Set up the NFS read request struct
156 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
157 const struct rpc_call_ops *call_ops,
158 unsigned int count, unsigned int offset)
160 struct inode *inode = req->wb_context->path.dentry->d_inode;
161 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
162 struct rpc_task *task;
163 struct rpc_message msg = {
164 .rpc_argp = &data->args,
165 .rpc_resp = &data->res,
166 .rpc_cred = req->wb_context->cred,
168 struct rpc_task_setup task_setup_data = {
170 .rpc_client = NFS_CLIENT(inode),
172 .callback_ops = call_ops,
173 .callback_data = data,
174 .workqueue = nfsiod_workqueue,
175 .flags = RPC_TASK_ASYNC | swap_flags,
180 data->cred = msg.rpc_cred;
182 data->args.fh = NFS_FH(inode);
183 data->args.offset = req_offset(req) + offset;
184 data->args.pgbase = req->wb_pgbase + offset;
185 data->args.pages = data->pagevec;
186 data->args.count = count;
187 data->args.context = get_nfs_open_context(req->wb_context);
189 data->res.fattr = &data->fattr;
190 data->res.count = count;
192 nfs_fattr_init(&data->fattr);
194 /* Set up the initial task struct. */
195 NFS_PROTO(inode)->read_setup(data, &msg);
197 dprintk("NFS: %5u initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
200 (long long)NFS_FILEID(inode),
202 (unsigned long long)data->args.offset);
204 task = rpc_run_task(&task_setup_data);
210 nfs_async_read_error(struct list_head *head)
212 struct nfs_page *req;
214 while (!list_empty(head)) {
215 req = nfs_list_entry(head->next);
216 nfs_list_remove_request(req);
217 SetPageError(req->wb_page);
218 nfs_readpage_release(req);
223 * Generate multiple requests to fill a single page.
225 * We optimize to reduce the number of read operations on the wire. If we
226 * detect that we're reading a page, or an area of a page, that is past the
227 * end of file, we do not generate NFS read operations but just clear the
228 * parts of the page that would have come back zero from the server anyway.
230 * We rely on the cached value of i_size to make this determination; another
231 * client can fill pages on the server past our cached end-of-file, but we
232 * won't see the new data until our attribute cache is updated. This is more
233 * or less conventional NFS client behavior.
235 static int nfs_pagein_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags)
237 struct nfs_page *req = nfs_list_entry(head->next);
238 struct page *page = req->wb_page;
239 struct nfs_read_data *data;
240 size_t rsize = NFS_SERVER(inode)->rsize, nbytes;
245 nfs_list_remove_request(req);
249 size_t len = min(nbytes,rsize);
251 data = nfs_readdata_alloc(1);
254 INIT_LIST_HEAD(&data->pages);
255 list_add(&data->pages, &list);
258 } while(nbytes != 0);
259 atomic_set(&req->wb_complete, requests);
261 ClearPageError(page);
265 data = list_entry(list.next, struct nfs_read_data, pages);
266 list_del_init(&data->pages);
268 data->pagevec[0] = page;
272 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
276 } while (nbytes != 0);
281 while (!list_empty(&list)) {
282 data = list_entry(list.next, struct nfs_read_data, pages);
283 list_del(&data->pages);
284 nfs_readdata_free(data);
287 nfs_readpage_release(req);
291 static int nfs_pagein_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags)
293 struct nfs_page *req;
295 struct nfs_read_data *data;
297 data = nfs_readdata_alloc(npages);
301 INIT_LIST_HEAD(&data->pages);
302 pages = data->pagevec;
303 while (!list_empty(head)) {
304 req = nfs_list_entry(head->next);
305 nfs_list_remove_request(req);
306 nfs_list_add_request(req, &data->pages);
307 ClearPageError(req->wb_page);
308 *pages++ = req->wb_page;
310 req = nfs_list_entry(data->pages.next);
312 nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
315 nfs_async_read_error(head);
320 * This is the callback from RPC telling us whether a reply was
321 * received or some error occurred (timeout or socket shutdown).
323 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
327 dprintk("NFS: %s: %5u, (status %d)\n", __FUNCTION__, task->tk_pid,
330 status = NFS_PROTO(data->inode)->read_done(task, data);
334 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
336 if (task->tk_status == -ESTALE) {
337 set_bit(NFS_INO_STALE, &NFS_I(data->inode)->flags);
338 nfs_mark_for_revalidate(data->inode);
343 static int nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
345 struct nfs_readargs *argp = &data->args;
346 struct nfs_readres *resp = &data->res;
348 if (resp->eof || resp->count == argp->count)
351 /* This is a short read! */
352 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
353 /* Has the server at least made some progress? */
354 if (resp->count == 0)
357 /* Yes, so retry the read at the end of the data */
358 argp->offset += resp->count;
359 argp->pgbase += resp->count;
360 argp->count -= resp->count;
361 rpc_restart_call(task);
366 * Handle a read reply that fills part of a page.
368 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
370 struct nfs_read_data *data = calldata;
371 struct nfs_page *req = data->req;
372 struct page *page = req->wb_page;
374 if (nfs_readpage_result(task, data) != 0)
377 if (likely(task->tk_status >= 0)) {
378 nfs_readpage_truncate_uninitialised_page(data);
379 if (nfs_readpage_retry(task, data) != 0)
382 if (unlikely(task->tk_status < 0))
384 if (atomic_dec_and_test(&req->wb_complete)) {
385 if (!PageError(page))
386 SetPageUptodate(page);
387 nfs_readpage_release(req);
391 static const struct rpc_call_ops nfs_read_partial_ops = {
392 .rpc_call_done = nfs_readpage_result_partial,
393 .rpc_release = nfs_readdata_release,
396 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
398 unsigned int count = data->res.count;
399 unsigned int base = data->args.pgbase;
403 count = data->args.count;
404 if (unlikely(count == 0))
406 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
407 base &= ~PAGE_CACHE_MASK;
409 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
410 SetPageUptodate(*pages);
413 /* Was this a short read? */
414 if (data->res.eof || data->res.count == data->args.count)
415 SetPageUptodate(*pages);
419 * This is the callback from RPC telling us whether a reply was
420 * received or some error occurred (timeout or socket shutdown).
422 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
424 struct nfs_read_data *data = calldata;
426 if (nfs_readpage_result(task, data) != 0)
429 * Note: nfs_readpage_retry may change the values of
430 * data->args. In the multi-page case, we therefore need
431 * to ensure that we call nfs_readpage_set_pages_uptodate()
434 if (likely(task->tk_status >= 0)) {
435 nfs_readpage_truncate_uninitialised_page(data);
436 nfs_readpage_set_pages_uptodate(data);
437 if (nfs_readpage_retry(task, data) != 0)
440 while (!list_empty(&data->pages)) {
441 struct nfs_page *req = nfs_list_entry(data->pages.next);
443 nfs_list_remove_request(req);
444 nfs_readpage_release(req);
448 static const struct rpc_call_ops nfs_read_full_ops = {
449 .rpc_call_done = nfs_readpage_result_full,
450 .rpc_release = nfs_readdata_release,
454 * Read a page over NFS.
455 * We read the page synchronously in the following case:
456 * - The error flag is set for this page. This happens only when a
457 * previous async read operation failed.
459 int nfs_readpage(struct file *file, struct page *page)
461 struct nfs_open_context *ctx;
462 struct inode *inode = page->mapping->host;
465 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
466 page, PAGE_CACHE_SIZE, page->index);
467 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
468 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
471 * Try to flush any pending writes to the file..
473 * NOTE! Because we own the page lock, there cannot
474 * be any new pending writes generated at this point
475 * for this page (other pages can be written to).
477 error = nfs_wb_page(inode, page);
480 if (PageUptodate(page))
484 if (NFS_STALE(inode))
489 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
493 ctx = get_nfs_open_context(nfs_file_open_context(file));
495 error = nfs_readpage_async(ctx, inode, page);
497 put_nfs_open_context(ctx);
504 struct nfs_readdesc {
505 struct nfs_pageio_descriptor *pgio;
506 struct nfs_open_context *ctx;
510 readpage_async_filler(void *data, struct page *page)
512 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
513 struct inode *inode = page->mapping->host;
514 struct nfs_page *new;
518 error = nfs_wb_page(inode, page);
521 if (PageUptodate(page))
524 len = nfs_page_length(page);
526 return nfs_return_empty_page(page);
528 new = nfs_create_request(desc->ctx, inode, page, 0, len);
532 if (len < PAGE_CACHE_SIZE)
533 zero_user_segment(page, len, PAGE_CACHE_SIZE);
534 nfs_pageio_add_request(desc->pgio, new);
537 error = PTR_ERR(new);
544 int nfs_readpages(struct file *filp, struct address_space *mapping,
545 struct list_head *pages, unsigned nr_pages)
547 struct nfs_pageio_descriptor pgio;
548 struct nfs_readdesc desc = {
551 struct inode *inode = mapping->host;
552 struct nfs_server *server = NFS_SERVER(inode);
553 size_t rsize = server->rsize;
554 unsigned long npages;
557 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
559 (long long)NFS_FILEID(inode),
561 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
563 if (NFS_STALE(inode))
567 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
568 if (desc.ctx == NULL)
571 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
572 if (rsize < PAGE_CACHE_SIZE)
573 nfs_pageio_init(&pgio, inode, nfs_pagein_multi, rsize, 0);
575 nfs_pageio_init(&pgio, inode, nfs_pagein_one, rsize, 0);
577 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
579 nfs_pageio_complete(&pgio);
580 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
581 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
582 put_nfs_open_context(desc.ctx);
587 int __init nfs_init_readpagecache(void)
589 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
590 sizeof(struct nfs_read_data),
591 0, SLAB_HWCACHE_ALIGN,
593 if (nfs_rdata_cachep == NULL)
596 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
598 if (nfs_rdata_mempool == NULL)
604 void nfs_destroy_readpagecache(void)
606 mempool_destroy(nfs_rdata_mempool);
607 kmem_cache_destroy(nfs_rdata_cachep);