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);
157 * Set up the NFS read request struct
159 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
160 const struct rpc_call_ops *call_ops,
161 unsigned int count, unsigned int offset)
167 data->inode = inode = req->wb_context->path.dentry->d_inode;
168 data->cred = req->wb_context->cred;
170 data->args.fh = NFS_FH(inode);
171 data->args.offset = req_offset(req) + offset;
172 data->args.pgbase = req->wb_pgbase + offset;
173 data->args.pages = data->pagevec;
174 data->args.count = count;
175 data->args.context = req->wb_context;
177 data->res.fattr = &data->fattr;
178 data->res.count = count;
180 nfs_fattr_init(&data->fattr);
182 /* Set up the initial task struct. */
183 flags = RPC_TASK_ASYNC | (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
184 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
185 NFS_PROTO(inode)->read_setup(data);
187 data->task.tk_cookie = (unsigned long)inode;
189 dprintk("NFS: %5u initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
192 (long long)NFS_FILEID(inode),
194 (unsigned long long)data->args.offset);
198 nfs_async_read_error(struct list_head *head)
200 struct nfs_page *req;
202 while (!list_empty(head)) {
203 req = nfs_list_entry(head->next);
204 nfs_list_remove_request(req);
205 SetPageError(req->wb_page);
206 nfs_readpage_release(req);
211 * Start an async read operation
213 static void nfs_execute_read(struct nfs_read_data *data)
215 rpc_execute(&data->task);
219 * Generate multiple requests to fill a single page.
221 * We optimize to reduce the number of read operations on the wire. If we
222 * detect that we're reading a page, or an area of a page, that is past the
223 * end of file, we do not generate NFS read operations but just clear the
224 * parts of the page that would have come back zero from the server anyway.
226 * We rely on the cached value of i_size to make this determination; another
227 * client can fill pages on the server past our cached end-of-file, but we
228 * won't see the new data until our attribute cache is updated. This is more
229 * or less conventional NFS client behavior.
231 static int nfs_pagein_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags)
233 struct nfs_page *req = nfs_list_entry(head->next);
234 struct page *page = req->wb_page;
235 struct nfs_read_data *data;
236 size_t rsize = NFS_SERVER(inode)->rsize, nbytes;
241 nfs_list_remove_request(req);
245 size_t len = min(nbytes,rsize);
247 data = nfs_readdata_alloc(1);
250 INIT_LIST_HEAD(&data->pages);
251 list_add(&data->pages, &list);
254 } while(nbytes != 0);
255 atomic_set(&req->wb_complete, requests);
257 ClearPageError(page);
261 data = list_entry(list.next, struct nfs_read_data, pages);
262 list_del_init(&data->pages);
264 data->pagevec[0] = page;
268 nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
272 nfs_execute_read(data);
273 } while (nbytes != 0);
278 while (!list_empty(&list)) {
279 data = list_entry(list.next, struct nfs_read_data, pages);
280 list_del(&data->pages);
281 nfs_readdata_free(data);
284 nfs_readpage_release(req);
288 static int nfs_pagein_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags)
290 struct nfs_page *req;
292 struct nfs_read_data *data;
294 data = nfs_readdata_alloc(npages);
298 INIT_LIST_HEAD(&data->pages);
299 pages = data->pagevec;
300 while (!list_empty(head)) {
301 req = nfs_list_entry(head->next);
302 nfs_list_remove_request(req);
303 nfs_list_add_request(req, &data->pages);
304 ClearPageError(req->wb_page);
305 *pages++ = req->wb_page;
307 req = nfs_list_entry(data->pages.next);
309 nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
311 nfs_execute_read(data);
314 nfs_async_read_error(head);
319 * This is the callback from RPC telling us whether a reply was
320 * received or some error occurred (timeout or socket shutdown).
322 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
326 dprintk("NFS: %s: %5u, (status %d)\n", __FUNCTION__, task->tk_pid,
329 status = NFS_PROTO(data->inode)->read_done(task, data);
333 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
335 if (task->tk_status == -ESTALE) {
336 set_bit(NFS_INO_STALE, &NFS_FLAGS(data->inode));
337 nfs_mark_for_revalidate(data->inode);
342 static int nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
344 struct nfs_readargs *argp = &data->args;
345 struct nfs_readres *resp = &data->res;
347 if (resp->eof || resp->count == argp->count)
350 /* This is a short read! */
351 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
352 /* Has the server at least made some progress? */
353 if (resp->count == 0)
356 /* Yes, so retry the read at the end of the data */
357 argp->offset += resp->count;
358 argp->pgbase += resp->count;
359 argp->count -= resp->count;
360 rpc_restart_call(task);
365 * Handle a read reply that fills part of a page.
367 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
369 struct nfs_read_data *data = calldata;
370 struct nfs_page *req = data->req;
371 struct page *page = req->wb_page;
373 if (nfs_readpage_result(task, data) != 0)
376 if (likely(task->tk_status >= 0)) {
377 nfs_readpage_truncate_uninitialised_page(data);
378 if (nfs_readpage_retry(task, data) != 0)
381 if (unlikely(task->tk_status < 0))
383 if (atomic_dec_and_test(&req->wb_complete)) {
384 if (!PageError(page))
385 SetPageUptodate(page);
386 nfs_readpage_release(req);
390 static const struct rpc_call_ops nfs_read_partial_ops = {
391 .rpc_call_done = nfs_readpage_result_partial,
392 .rpc_release = nfs_readdata_release,
395 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
397 unsigned int count = data->res.count;
398 unsigned int base = data->args.pgbase;
402 count = data->args.count;
403 if (unlikely(count == 0))
405 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
406 base &= ~PAGE_CACHE_MASK;
408 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
409 SetPageUptodate(*pages);
412 /* Was this a short read? */
413 if (data->res.eof || data->res.count == data->args.count)
414 SetPageUptodate(*pages);
418 * This is the callback from RPC telling us whether a reply was
419 * received or some error occurred (timeout or socket shutdown).
421 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
423 struct nfs_read_data *data = calldata;
425 if (nfs_readpage_result(task, data) != 0)
428 * Note: nfs_readpage_retry may change the values of
429 * data->args. In the multi-page case, we therefore need
430 * to ensure that we call nfs_readpage_set_pages_uptodate()
433 if (likely(task->tk_status >= 0)) {
434 nfs_readpage_truncate_uninitialised_page(data);
435 nfs_readpage_set_pages_uptodate(data);
436 if (nfs_readpage_retry(task, data) != 0)
439 while (!list_empty(&data->pages)) {
440 struct nfs_page *req = nfs_list_entry(data->pages.next);
442 nfs_list_remove_request(req);
443 nfs_readpage_release(req);
447 static const struct rpc_call_ops nfs_read_full_ops = {
448 .rpc_call_done = nfs_readpage_result_full,
449 .rpc_release = nfs_readdata_release,
453 * Read a page over NFS.
454 * We read the page synchronously in the following case:
455 * - The error flag is set for this page. This happens only when a
456 * previous async read operation failed.
458 int nfs_readpage(struct file *file, struct page *page)
460 struct nfs_open_context *ctx;
461 struct inode *inode = page->mapping->host;
464 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
465 page, PAGE_CACHE_SIZE, page->index);
466 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
467 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
470 * Try to flush any pending writes to the file..
472 * NOTE! Because we own the page lock, there cannot
473 * be any new pending writes generated at this point
474 * for this page (other pages can be written to).
476 error = nfs_wb_page(inode, page);
479 if (PageUptodate(page))
483 if (NFS_STALE(inode))
488 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
492 ctx = get_nfs_open_context(nfs_file_open_context(file));
494 error = nfs_readpage_async(ctx, inode, page);
496 put_nfs_open_context(ctx);
503 struct nfs_readdesc {
504 struct nfs_pageio_descriptor *pgio;
505 struct nfs_open_context *ctx;
509 readpage_async_filler(void *data, struct page *page)
511 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
512 struct inode *inode = page->mapping->host;
513 struct nfs_page *new;
517 error = nfs_wb_page(inode, page);
520 if (PageUptodate(page))
523 len = nfs_page_length(page);
525 return nfs_return_empty_page(page);
527 new = nfs_create_request(desc->ctx, inode, page, 0, len);
531 if (len < PAGE_CACHE_SIZE)
532 zero_user_page(page, len, PAGE_CACHE_SIZE - len, KM_USER0);
533 nfs_pageio_add_request(desc->pgio, new);
536 error = PTR_ERR(new);
543 int nfs_readpages(struct file *filp, struct address_space *mapping,
544 struct list_head *pages, unsigned nr_pages)
546 struct nfs_pageio_descriptor pgio;
547 struct nfs_readdesc desc = {
550 struct inode *inode = mapping->host;
551 struct nfs_server *server = NFS_SERVER(inode);
552 size_t rsize = server->rsize;
553 unsigned long npages;
556 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
558 (long long)NFS_FILEID(inode),
560 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
562 if (NFS_STALE(inode))
566 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
567 if (desc.ctx == NULL)
570 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
571 if (rsize < PAGE_CACHE_SIZE)
572 nfs_pageio_init(&pgio, inode, nfs_pagein_multi, rsize, 0);
574 nfs_pageio_init(&pgio, inode, nfs_pagein_one, rsize, 0);
576 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
578 nfs_pageio_complete(&pgio);
579 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
580 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
581 put_nfs_open_context(desc.ctx);
586 int __init nfs_init_readpagecache(void)
588 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
589 sizeof(struct nfs_read_data),
590 0, SLAB_HWCACHE_ALIGN,
592 if (nfs_rdata_cachep == NULL)
595 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
597 if (nfs_rdata_mempool == NULL)
603 void nfs_destroy_readpagecache(void)
605 mempool_destroy(nfs_rdata_mempool);
606 kmem_cache_destroy(nfs_rdata_cachep);