6 * Partial copy of Linus' read cache modifications to fs/nfs/file.c
7 * modified for async RPC by okir@monad.swb.de
9 * We do an ugly hack here in order to return proper error codes to the
10 * user program when a read request failed: since generic_file_read
11 * only checks the return value of inode->i_op->readpage() which is always 0
12 * for async RPC, we set the error bit of the page to 1 when an error occurs,
13 * and make nfs_readpage transmit requests synchronously when encountering this.
14 * This is only a small problem, though, since we now retry all operations
15 * within the RPC code when root squashing is suspected.
18 #include <linux/config.h>
19 #include <linux/time.h>
20 #include <linux/kernel.h>
21 #include <linux/errno.h>
22 #include <linux/fcntl.h>
23 #include <linux/stat.h>
25 #include <linux/slab.h>
26 #include <linux/pagemap.h>
27 #include <linux/sunrpc/clnt.h>
28 #include <linux/nfs_fs.h>
29 #include <linux/nfs_page.h>
30 #include <linux/smp_lock.h>
32 #include <asm/system.h>
34 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
36 static int nfs_pagein_one(struct list_head *, struct inode *);
37 static void nfs_readpage_result_partial(struct nfs_read_data *, int);
38 static void nfs_readpage_result_full(struct nfs_read_data *, int);
40 static kmem_cache_t *nfs_rdata_cachep;
41 mempool_t *nfs_rdata_mempool;
43 #define MIN_POOL_READ (32)
45 void nfs_readdata_release(void *data)
47 nfs_readdata_free(data);
51 unsigned int nfs_page_length(struct inode *inode, struct page *page)
53 loff_t i_size = i_size_read(inode);
58 idx = (i_size - 1) >> PAGE_CACHE_SHIFT;
59 if (page->index > idx)
61 if (page->index != idx)
62 return PAGE_CACHE_SIZE;
63 return 1 + ((i_size - 1) & (PAGE_CACHE_SIZE - 1));
67 int nfs_return_empty_page(struct page *page)
69 memclear_highpage_flush(page, 0, PAGE_CACHE_SIZE);
70 SetPageUptodate(page);
76 * Read a page synchronously.
78 static int nfs_readpage_sync(struct nfs_open_context *ctx, struct inode *inode,
81 unsigned int rsize = NFS_SERVER(inode)->rsize;
82 unsigned int count = PAGE_CACHE_SIZE;
84 struct nfs_read_data *rdata;
86 rdata = nfs_readdata_alloc(1);
90 memset(rdata, 0, sizeof(*rdata));
91 rdata->flags = (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
92 rdata->cred = ctx->cred;
94 INIT_LIST_HEAD(&rdata->pages);
95 rdata->args.fh = NFS_FH(inode);
96 rdata->args.context = ctx;
97 rdata->args.pages = &page;
98 rdata->args.pgbase = 0UL;
99 rdata->args.count = rsize;
100 rdata->res.fattr = &rdata->fattr;
102 dprintk("NFS: nfs_readpage_sync(%p)\n", page);
105 * This works now because the socket layer never tries to DMA
106 * into this buffer directly.
110 rdata->args.count = count;
111 rdata->res.count = rdata->args.count;
112 rdata->args.offset = page_offset(page) + rdata->args.pgbase;
114 dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
115 NFS_SERVER(inode)->hostname,
117 (long long)NFS_FILEID(inode),
118 (unsigned long long)rdata->args.pgbase,
122 result = NFS_PROTO(inode)->read(rdata);
126 * Even if we had a partial success we can't mark the page
130 if (result == -EISDIR)
135 rdata->args.pgbase += result;
136 /* Note: result == 0 should only happen if we're caching
137 * a write that extends the file and punches a hole.
139 if (rdata->res.eof != 0 || result == 0)
142 spin_lock(&inode->i_lock);
143 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME;
144 spin_unlock(&inode->i_lock);
147 memclear_highpage_flush(page, rdata->args.pgbase, count);
148 SetPageUptodate(page);
150 ClearPageError(page);
155 nfs_readdata_free(rdata);
159 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
162 LIST_HEAD(one_request);
163 struct nfs_page *new;
166 len = nfs_page_length(inode, page);
168 return nfs_return_empty_page(page);
169 new = nfs_create_request(ctx, inode, page, 0, len);
174 if (len < PAGE_CACHE_SIZE)
175 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
177 nfs_list_add_request(new, &one_request);
178 nfs_pagein_one(&one_request, inode);
182 static void nfs_readpage_release(struct nfs_page *req)
184 unlock_page(req->wb_page);
186 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
187 req->wb_context->dentry->d_inode->i_sb->s_id,
188 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
190 (long long)req_offset(req));
191 nfs_clear_request(req);
192 nfs_release_request(req);
196 * Set up the NFS read request struct
198 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
199 unsigned int count, unsigned int offset)
204 data->inode = inode = req->wb_context->dentry->d_inode;
205 data->cred = req->wb_context->cred;
207 data->args.fh = NFS_FH(inode);
208 data->args.offset = req_offset(req) + offset;
209 data->args.pgbase = req->wb_pgbase + offset;
210 data->args.pages = data->pagevec;
211 data->args.count = count;
212 data->args.context = req->wb_context;
214 data->res.fattr = &data->fattr;
215 data->res.count = count;
217 nfs_fattr_init(&data->fattr);
219 NFS_PROTO(inode)->read_setup(data);
221 data->task.tk_cookie = (unsigned long)inode;
223 dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
226 (long long)NFS_FILEID(inode),
228 (unsigned long long)data->args.offset);
232 nfs_async_read_error(struct list_head *head)
234 struct nfs_page *req;
236 while (!list_empty(head)) {
237 req = nfs_list_entry(head->next);
238 nfs_list_remove_request(req);
239 SetPageError(req->wb_page);
240 nfs_readpage_release(req);
245 * Start an async read operation
247 static void nfs_execute_read(struct nfs_read_data *data)
249 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
252 rpc_clnt_sigmask(clnt, &oldset);
254 rpc_execute(&data->task);
256 rpc_clnt_sigunmask(clnt, &oldset);
260 * Generate multiple requests to fill a single page.
262 * We optimize to reduce the number of read operations on the wire. If we
263 * detect that we're reading a page, or an area of a page, that is past the
264 * end of file, we do not generate NFS read operations but just clear the
265 * parts of the page that would have come back zero from the server anyway.
267 * We rely on the cached value of i_size to make this determination; another
268 * client can fill pages on the server past our cached end-of-file, but we
269 * won't see the new data until our attribute cache is updated. This is more
270 * or less conventional NFS client behavior.
272 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
274 struct nfs_page *req = nfs_list_entry(head->next);
275 struct page *page = req->wb_page;
276 struct nfs_read_data *data;
277 unsigned int rsize = NFS_SERVER(inode)->rsize;
278 unsigned int nbytes, offset;
282 nfs_list_remove_request(req);
284 nbytes = req->wb_bytes;
286 data = nfs_readdata_alloc(1);
289 INIT_LIST_HEAD(&data->pages);
290 list_add(&data->pages, &list);
296 atomic_set(&req->wb_complete, requests);
298 ClearPageError(page);
300 nbytes = req->wb_bytes;
302 data = list_entry(list.next, struct nfs_read_data, pages);
303 list_del_init(&data->pages);
305 data->pagevec[0] = page;
306 data->complete = nfs_readpage_result_partial;
308 if (nbytes > rsize) {
309 nfs_read_rpcsetup(req, data, rsize, offset);
313 nfs_read_rpcsetup(req, data, nbytes, offset);
316 nfs_execute_read(data);
317 } while (nbytes != 0);
322 while (!list_empty(&list)) {
323 data = list_entry(list.next, struct nfs_read_data, pages);
324 list_del(&data->pages);
325 nfs_readdata_free(data);
328 nfs_readpage_release(req);
332 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
334 struct nfs_page *req;
336 struct nfs_read_data *data;
339 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
340 return nfs_pagein_multi(head, inode);
342 data = nfs_readdata_alloc(NFS_SERVER(inode)->rpages);
346 INIT_LIST_HEAD(&data->pages);
347 pages = data->pagevec;
349 while (!list_empty(head)) {
350 req = nfs_list_entry(head->next);
351 nfs_list_remove_request(req);
352 nfs_list_add_request(req, &data->pages);
353 ClearPageError(req->wb_page);
354 *pages++ = req->wb_page;
355 count += req->wb_bytes;
357 req = nfs_list_entry(data->pages.next);
359 data->complete = nfs_readpage_result_full;
360 nfs_read_rpcsetup(req, data, count, 0);
362 nfs_execute_read(data);
365 nfs_async_read_error(head);
370 nfs_pagein_list(struct list_head *head, int rpages)
372 LIST_HEAD(one_request);
373 struct nfs_page *req;
375 unsigned int pages = 0;
377 while (!list_empty(head)) {
378 pages += nfs_coalesce_requests(head, &one_request, rpages);
379 req = nfs_list_entry(one_request.next);
380 error = nfs_pagein_one(&one_request, req->wb_context->dentry->d_inode);
387 nfs_async_read_error(head);
392 * Handle a read reply that fills part of a page.
394 static void nfs_readpage_result_partial(struct nfs_read_data *data, int status)
396 struct nfs_page *req = data->req;
397 struct page *page = req->wb_page;
400 unsigned int request = data->args.count;
401 unsigned int result = data->res.count;
403 if (result < request) {
404 memclear_highpage_flush(page,
405 data->args.pgbase + result,
411 if (atomic_dec_and_test(&req->wb_complete)) {
412 if (!PageError(page))
413 SetPageUptodate(page);
414 nfs_readpage_release(req);
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 nfs_read_data *data, int status)
424 unsigned int count = data->res.count;
426 while (!list_empty(&data->pages)) {
427 struct nfs_page *req = nfs_list_entry(data->pages.next);
428 struct page *page = req->wb_page;
429 nfs_list_remove_request(req);
432 if (count < PAGE_CACHE_SIZE) {
433 if (count < req->wb_bytes)
434 memclear_highpage_flush(page,
435 req->wb_pgbase + count,
436 req->wb_bytes - count);
439 count -= PAGE_CACHE_SIZE;
440 SetPageUptodate(page);
443 nfs_readpage_release(req);
448 * This is the callback from RPC telling us whether a reply was
449 * received or some error occurred (timeout or socket shutdown).
451 void nfs_readpage_result(struct rpc_task *task, void *calldata)
453 struct nfs_read_data *data = calldata;
454 struct nfs_readargs *argp = &data->args;
455 struct nfs_readres *resp = &data->res;
456 int status = task->tk_status;
458 dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
459 task->tk_pid, status);
461 /* Is this a short read? */
462 if (task->tk_status >= 0 && resp->count < argp->count && !resp->eof) {
463 /* Has the server at least made some progress? */
464 if (resp->count != 0) {
465 /* Yes, so retry the read at the end of the data */
466 argp->offset += resp->count;
467 argp->pgbase += resp->count;
468 argp->count -= resp->count;
469 rpc_restart_call(task);
472 task->tk_status = -EIO;
474 spin_lock(&data->inode->i_lock);
475 NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
476 spin_unlock(&data->inode->i_lock);
477 data->complete(data, status);
481 * Read a page over NFS.
482 * We read the page synchronously in the following case:
483 * - The error flag is set for this page. This happens only when a
484 * previous async read operation failed.
486 int nfs_readpage(struct file *file, struct page *page)
488 struct nfs_open_context *ctx;
489 struct inode *inode = page->mapping->host;
492 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
493 page, PAGE_CACHE_SIZE, page->index);
495 * Try to flush any pending writes to the file..
497 * NOTE! Because we own the page lock, there cannot
498 * be any new pending writes generated at this point
499 * for this page (other pages can be written to).
501 error = nfs_wb_page(inode, page);
506 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
510 ctx = get_nfs_open_context((struct nfs_open_context *)
512 if (!IS_SYNC(inode)) {
513 error = nfs_readpage_async(ctx, inode, page);
517 error = nfs_readpage_sync(ctx, inode, page);
518 if (error < 0 && IS_SWAPFILE(inode))
519 printk("Aiee.. nfs swap-in of page failed!\n");
521 put_nfs_open_context(ctx);
529 struct nfs_readdesc {
530 struct list_head *head;
531 struct nfs_open_context *ctx;
535 readpage_async_filler(void *data, struct page *page)
537 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
538 struct inode *inode = page->mapping->host;
539 struct nfs_page *new;
542 nfs_wb_page(inode, page);
543 len = nfs_page_length(inode, page);
545 return nfs_return_empty_page(page);
546 new = nfs_create_request(desc->ctx, inode, page, 0, len);
552 if (len < PAGE_CACHE_SIZE)
553 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
554 nfs_list_add_request(new, desc->head);
558 int nfs_readpages(struct file *filp, struct address_space *mapping,
559 struct list_head *pages, unsigned nr_pages)
562 struct nfs_readdesc desc = {
565 struct inode *inode = mapping->host;
566 struct nfs_server *server = NFS_SERVER(inode);
569 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
571 (long long)NFS_FILEID(inode),
575 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
576 if (desc.ctx == NULL)
579 desc.ctx = get_nfs_open_context((struct nfs_open_context *)
581 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
582 if (!list_empty(&head)) {
583 int err = nfs_pagein_list(&head, server->rpages);
587 put_nfs_open_context(desc.ctx);
591 int nfs_init_readpagecache(void)
593 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
594 sizeof(struct nfs_read_data),
595 0, SLAB_HWCACHE_ALIGN,
597 if (nfs_rdata_cachep == NULL)
600 nfs_rdata_mempool = mempool_create(MIN_POOL_READ,
604 if (nfs_rdata_mempool == NULL)
610 void nfs_destroy_readpagecache(void)
612 mempool_destroy(nfs_rdata_mempool);
613 if (kmem_cache_destroy(nfs_rdata_cachep))
614 printk(KERN_INFO "nfs_read_data: not all structures were freed\n");