Pull fluff into release branch
[sfrench/cifs-2.6.git] / fs / nfs / read.c
1 /*
2  * linux/fs/nfs/read.c
3  *
4  * Block I/O for NFS
5  *
6  * Partial copy of Linus' read cache modifications to fs/nfs/file.c
7  * modified for async RPC by okir@monad.swb.de
8  */
9
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>
15 #include <linux/mm.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>
22
23 #include <asm/system.h>
24
25 #include "internal.h"
26 #include "iostat.h"
27
28 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
29
30 static int nfs_pagein_one(struct list_head *, struct inode *);
31 static const struct rpc_call_ops nfs_read_partial_ops;
32 static const struct rpc_call_ops nfs_read_full_ops;
33
34 static struct kmem_cache *nfs_rdata_cachep;
35 static mempool_t *nfs_rdata_mempool;
36
37 #define MIN_POOL_READ   (32)
38
39 struct nfs_read_data *nfs_readdata_alloc(size_t len)
40 {
41         unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
42         struct nfs_read_data *p = mempool_alloc(nfs_rdata_mempool, GFP_NOFS);
43
44         if (p) {
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;
50                 else {
51                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
52                         if (!p->pagevec) {
53                                 mempool_free(p, nfs_rdata_mempool);
54                                 p = NULL;
55                         }
56                 }
57         }
58         return p;
59 }
60
61 static void nfs_readdata_rcu_free(struct rcu_head *head)
62 {
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]))
65                 kfree(p->pagevec);
66         mempool_free(p, nfs_rdata_mempool);
67 }
68
69 static void nfs_readdata_free(struct nfs_read_data *rdata)
70 {
71         call_rcu_bh(&rdata->task.u.tk_rcu, nfs_readdata_rcu_free);
72 }
73
74 void nfs_readdata_release(void *data)
75 {
76         nfs_readdata_free(data);
77 }
78
79 static
80 int nfs_return_empty_page(struct page *page)
81 {
82         memclear_highpage_flush(page, 0, PAGE_CACHE_SIZE);
83         SetPageUptodate(page);
84         unlock_page(page);
85         return 0;
86 }
87
88 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
89 {
90         unsigned int remainder = data->args.count - data->res.count;
91         unsigned int base = data->args.pgbase + data->res.count;
92         unsigned int pglen;
93         struct page **pages;
94
95         if (data->res.eof == 0 || remainder == 0)
96                 return;
97         /*
98          * Note: "remainder" can never be negative, since we check for
99          *      this in the XDR code.
100          */
101         pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
102         base &= ~PAGE_CACHE_MASK;
103         pglen = PAGE_CACHE_SIZE - base;
104         for (;;) {
105                 if (remainder <= pglen) {
106                         memclear_highpage_flush(*pages, base, remainder);
107                         break;
108                 }
109                 memclear_highpage_flush(*pages, base, pglen);
110                 pages++;
111                 remainder -= pglen;
112                 pglen = PAGE_CACHE_SIZE;
113                 base = 0;
114         }
115 }
116
117 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
118                 struct page *page)
119 {
120         LIST_HEAD(one_request);
121         struct nfs_page *new;
122         unsigned int len;
123
124         len = nfs_page_length(page);
125         if (len == 0)
126                 return nfs_return_empty_page(page);
127         new = nfs_create_request(ctx, inode, page, 0, len);
128         if (IS_ERR(new)) {
129                 unlock_page(page);
130                 return PTR_ERR(new);
131         }
132         if (len < PAGE_CACHE_SIZE)
133                 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
134
135         nfs_list_add_request(new, &one_request);
136         nfs_pagein_one(&one_request, inode);
137         return 0;
138 }
139
140 static void nfs_readpage_release(struct nfs_page *req)
141 {
142         unlock_page(req->wb_page);
143
144         dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
145                         req->wb_context->dentry->d_inode->i_sb->s_id,
146                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
147                         req->wb_bytes,
148                         (long long)req_offset(req));
149         nfs_clear_request(req);
150         nfs_release_request(req);
151 }
152
153 /*
154  * Set up the NFS read request struct
155  */
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)
159 {
160         struct inode            *inode;
161         int flags;
162
163         data->req         = req;
164         data->inode       = inode = req->wb_context->dentry->d_inode;
165         data->cred        = req->wb_context->cred;
166
167         data->args.fh     = NFS_FH(inode);
168         data->args.offset = req_offset(req) + offset;
169         data->args.pgbase = req->wb_pgbase + offset;
170         data->args.pages  = data->pagevec;
171         data->args.count  = count;
172         data->args.context = req->wb_context;
173
174         data->res.fattr   = &data->fattr;
175         data->res.count   = count;
176         data->res.eof     = 0;
177         nfs_fattr_init(&data->fattr);
178
179         /* Set up the initial task struct. */
180         flags = RPC_TASK_ASYNC | (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
181         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
182         NFS_PROTO(inode)->read_setup(data);
183
184         data->task.tk_cookie = (unsigned long)inode;
185
186         dprintk("NFS: %5u initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
187                         data->task.tk_pid,
188                         inode->i_sb->s_id,
189                         (long long)NFS_FILEID(inode),
190                         count,
191                         (unsigned long long)data->args.offset);
192 }
193
194 static void
195 nfs_async_read_error(struct list_head *head)
196 {
197         struct nfs_page *req;
198
199         while (!list_empty(head)) {
200                 req = nfs_list_entry(head->next);
201                 nfs_list_remove_request(req);
202                 SetPageError(req->wb_page);
203                 nfs_readpage_release(req);
204         }
205 }
206
207 /*
208  * Start an async read operation
209  */
210 static void nfs_execute_read(struct nfs_read_data *data)
211 {
212         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
213         sigset_t oldset;
214
215         rpc_clnt_sigmask(clnt, &oldset);
216         rpc_execute(&data->task);
217         rpc_clnt_sigunmask(clnt, &oldset);
218 }
219
220 /*
221  * Generate multiple requests to fill a single page.
222  *
223  * We optimize to reduce the number of read operations on the wire.  If we
224  * detect that we're reading a page, or an area of a page, that is past the
225  * end of file, we do not generate NFS read operations but just clear the
226  * parts of the page that would have come back zero from the server anyway.
227  *
228  * We rely on the cached value of i_size to make this determination; another
229  * client can fill pages on the server past our cached end-of-file, but we
230  * won't see the new data until our attribute cache is updated.  This is more
231  * or less conventional NFS client behavior.
232  */
233 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
234 {
235         struct nfs_page *req = nfs_list_entry(head->next);
236         struct page *page = req->wb_page;
237         struct nfs_read_data *data;
238         size_t rsize = NFS_SERVER(inode)->rsize, nbytes;
239         unsigned int offset;
240         int requests = 0;
241         LIST_HEAD(list);
242
243         nfs_list_remove_request(req);
244
245         nbytes = req->wb_bytes;
246         do {
247                 size_t len = min(nbytes,rsize);
248
249                 data = nfs_readdata_alloc(len);
250                 if (!data)
251                         goto out_bad;
252                 INIT_LIST_HEAD(&data->pages);
253                 list_add(&data->pages, &list);
254                 requests++;
255                 nbytes -= len;
256         } while(nbytes != 0);
257         atomic_set(&req->wb_complete, requests);
258
259         ClearPageError(page);
260         offset = 0;
261         nbytes = req->wb_bytes;
262         do {
263                 data = list_entry(list.next, struct nfs_read_data, pages);
264                 list_del_init(&data->pages);
265
266                 data->pagevec[0] = page;
267
268                 if (nbytes > rsize) {
269                         nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
270                                         rsize, offset);
271                         offset += rsize;
272                         nbytes -= rsize;
273                 } else {
274                         nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
275                                         nbytes, offset);
276                         nbytes = 0;
277                 }
278                 nfs_execute_read(data);
279         } while (nbytes != 0);
280
281         return 0;
282
283 out_bad:
284         while (!list_empty(&list)) {
285                 data = list_entry(list.next, struct nfs_read_data, pages);
286                 list_del(&data->pages);
287                 nfs_readdata_free(data);
288         }
289         SetPageError(page);
290         nfs_readpage_release(req);
291         return -ENOMEM;
292 }
293
294 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
295 {
296         struct nfs_page         *req;
297         struct page             **pages;
298         struct nfs_read_data    *data;
299         unsigned int            count;
300
301         if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
302                 return nfs_pagein_multi(head, inode);
303
304         data = nfs_readdata_alloc(NFS_SERVER(inode)->rsize);
305         if (!data)
306                 goto out_bad;
307
308         INIT_LIST_HEAD(&data->pages);
309         pages = data->pagevec;
310         count = 0;
311         while (!list_empty(head)) {
312                 req = nfs_list_entry(head->next);
313                 nfs_list_remove_request(req);
314                 nfs_list_add_request(req, &data->pages);
315                 ClearPageError(req->wb_page);
316                 *pages++ = req->wb_page;
317                 count += req->wb_bytes;
318         }
319         req = nfs_list_entry(data->pages.next);
320
321         nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
322
323         nfs_execute_read(data);
324         return 0;
325 out_bad:
326         nfs_async_read_error(head);
327         return -ENOMEM;
328 }
329
330 static int
331 nfs_pagein_list(struct list_head *head, int rpages)
332 {
333         LIST_HEAD(one_request);
334         struct nfs_page         *req;
335         int                     error = 0;
336         unsigned int            pages = 0;
337
338         while (!list_empty(head)) {
339                 pages += nfs_coalesce_requests(head, &one_request, rpages);
340                 req = nfs_list_entry(one_request.next);
341                 error = nfs_pagein_one(&one_request, req->wb_context->dentry->d_inode);
342                 if (error < 0)
343                         break;
344         }
345         if (error >= 0)
346                 return pages;
347
348         nfs_async_read_error(head);
349         return error;
350 }
351
352 /*
353  * This is the callback from RPC telling us whether a reply was
354  * received or some error occurred (timeout or socket shutdown).
355  */
356 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
357 {
358         int status;
359
360         dprintk("NFS: %s: %5u, (status %d)\n", __FUNCTION__, task->tk_pid,
361                         task->tk_status);
362
363         status = NFS_PROTO(data->inode)->read_done(task, data);
364         if (status != 0)
365                 return status;
366
367         nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
368
369         if (task->tk_status == -ESTALE) {
370                 set_bit(NFS_INO_STALE, &NFS_FLAGS(data->inode));
371                 nfs_mark_for_revalidate(data->inode);
372         }
373         spin_lock(&data->inode->i_lock);
374         NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
375         spin_unlock(&data->inode->i_lock);
376         return 0;
377 }
378
379 static int nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
380 {
381         struct nfs_readargs *argp = &data->args;
382         struct nfs_readres *resp = &data->res;
383
384         if (resp->eof || resp->count == argp->count)
385                 return 0;
386
387         /* This is a short read! */
388         nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
389         /* Has the server at least made some progress? */
390         if (resp->count == 0)
391                 return 0;
392
393         /* Yes, so retry the read at the end of the data */
394         argp->offset += resp->count;
395         argp->pgbase += resp->count;
396         argp->count -= resp->count;
397         rpc_restart_call(task);
398         return -EAGAIN;
399 }
400
401 /*
402  * Handle a read reply that fills part of a page.
403  */
404 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
405 {
406         struct nfs_read_data *data = calldata;
407         struct nfs_page *req = data->req;
408         struct page *page = req->wb_page;
409  
410         if (nfs_readpage_result(task, data) != 0)
411                 return;
412
413         if (likely(task->tk_status >= 0)) {
414                 nfs_readpage_truncate_uninitialised_page(data);
415                 if (nfs_readpage_retry(task, data) != 0)
416                         return;
417         }
418         if (unlikely(task->tk_status < 0))
419                 SetPageError(page);
420         if (atomic_dec_and_test(&req->wb_complete)) {
421                 if (!PageError(page))
422                         SetPageUptodate(page);
423                 nfs_readpage_release(req);
424         }
425 }
426
427 static const struct rpc_call_ops nfs_read_partial_ops = {
428         .rpc_call_done = nfs_readpage_result_partial,
429         .rpc_release = nfs_readdata_release,
430 };
431
432 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
433 {
434         unsigned int count = data->res.count;
435         unsigned int base = data->args.pgbase;
436         struct page **pages;
437
438         if (data->res.eof)
439                 count = data->args.count;
440         if (unlikely(count == 0))
441                 return;
442         pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
443         base &= ~PAGE_CACHE_MASK;
444         count += base;
445         for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
446                 SetPageUptodate(*pages);
447         if (count == 0)
448                 return;
449         /* Was this a short read? */
450         if (data->res.eof || data->res.count == data->args.count)
451                 SetPageUptodate(*pages);
452 }
453
454 /*
455  * This is the callback from RPC telling us whether a reply was
456  * received or some error occurred (timeout or socket shutdown).
457  */
458 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
459 {
460         struct nfs_read_data *data = calldata;
461
462         if (nfs_readpage_result(task, data) != 0)
463                 return;
464         /*
465          * Note: nfs_readpage_retry may change the values of
466          * data->args. In the multi-page case, we therefore need
467          * to ensure that we call nfs_readpage_set_pages_uptodate()
468          * first.
469          */
470         if (likely(task->tk_status >= 0)) {
471                 nfs_readpage_truncate_uninitialised_page(data);
472                 nfs_readpage_set_pages_uptodate(data);
473                 if (nfs_readpage_retry(task, data) != 0)
474                         return;
475         }
476         while (!list_empty(&data->pages)) {
477                 struct nfs_page *req = nfs_list_entry(data->pages.next);
478
479                 nfs_list_remove_request(req);
480                 nfs_readpage_release(req);
481         }
482 }
483
484 static const struct rpc_call_ops nfs_read_full_ops = {
485         .rpc_call_done = nfs_readpage_result_full,
486         .rpc_release = nfs_readdata_release,
487 };
488
489 /*
490  * Read a page over NFS.
491  * We read the page synchronously in the following case:
492  *  -   The error flag is set for this page. This happens only when a
493  *      previous async read operation failed.
494  */
495 int nfs_readpage(struct file *file, struct page *page)
496 {
497         struct nfs_open_context *ctx;
498         struct inode *inode = page->mapping->host;
499         int             error;
500
501         dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
502                 page, PAGE_CACHE_SIZE, page->index);
503         nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
504         nfs_add_stats(inode, NFSIOS_READPAGES, 1);
505
506         /*
507          * Try to flush any pending writes to the file..
508          *
509          * NOTE! Because we own the page lock, there cannot
510          * be any new pending writes generated at this point
511          * for this page (other pages can be written to).
512          */
513         error = nfs_wb_page(inode, page);
514         if (error)
515                 goto out_error;
516
517         error = -ESTALE;
518         if (NFS_STALE(inode))
519                 goto out_error;
520
521         if (file == NULL) {
522                 error = -EBADF;
523                 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
524                 if (ctx == NULL)
525                         goto out_error;
526         } else
527                 ctx = get_nfs_open_context((struct nfs_open_context *)
528                                 file->private_data);
529
530         error = nfs_readpage_async(ctx, inode, page);
531
532         put_nfs_open_context(ctx);
533         return error;
534
535 out_error:
536         unlock_page(page);
537         return error;
538 }
539
540 struct nfs_readdesc {
541         struct list_head *head;
542         struct nfs_open_context *ctx;
543 };
544
545 static int
546 readpage_async_filler(void *data, struct page *page)
547 {
548         struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
549         struct inode *inode = page->mapping->host;
550         struct nfs_page *new;
551         unsigned int len;
552
553         nfs_wb_page(inode, page);
554         len = nfs_page_length(page);
555         if (len == 0)
556                 return nfs_return_empty_page(page);
557         new = nfs_create_request(desc->ctx, inode, page, 0, len);
558         if (IS_ERR(new)) {
559                         SetPageError(page);
560                         unlock_page(page);
561                         return PTR_ERR(new);
562         }
563         if (len < PAGE_CACHE_SIZE)
564                 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
565         nfs_list_add_request(new, desc->head);
566         return 0;
567 }
568
569 int nfs_readpages(struct file *filp, struct address_space *mapping,
570                 struct list_head *pages, unsigned nr_pages)
571 {
572         LIST_HEAD(head);
573         struct nfs_readdesc desc = {
574                 .head           = &head,
575         };
576         struct inode *inode = mapping->host;
577         struct nfs_server *server = NFS_SERVER(inode);
578         int ret = -ESTALE;
579
580         dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
581                         inode->i_sb->s_id,
582                         (long long)NFS_FILEID(inode),
583                         nr_pages);
584         nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
585
586         if (NFS_STALE(inode))
587                 goto out;
588
589         if (filp == NULL) {
590                 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
591                 if (desc.ctx == NULL)
592                         return -EBADF;
593         } else
594                 desc.ctx = get_nfs_open_context((struct nfs_open_context *)
595                                 filp->private_data);
596         ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
597         if (!list_empty(&head)) {
598                 int err = nfs_pagein_list(&head, server->rpages);
599                 if (!ret)
600                         nfs_add_stats(inode, NFSIOS_READPAGES, err);
601                         ret = err;
602         }
603         put_nfs_open_context(desc.ctx);
604 out:
605         return ret;
606 }
607
608 int __init nfs_init_readpagecache(void)
609 {
610         nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
611                                              sizeof(struct nfs_read_data),
612                                              0, SLAB_HWCACHE_ALIGN,
613                                              NULL, NULL);
614         if (nfs_rdata_cachep == NULL)
615                 return -ENOMEM;
616
617         nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
618                                                      nfs_rdata_cachep);
619         if (nfs_rdata_mempool == NULL)
620                 return -ENOMEM;
621
622         return 0;
623 }
624
625 void nfs_destroy_readpagecache(void)
626 {
627         mempool_destroy(nfs_rdata_mempool);
628         kmem_cache_destroy(nfs_rdata_cachep);
629 }