NFS: Cleanup of NFS read code
[safe/jmp/linux-2.6] / 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  * 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.
16  */
17
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>
24 #include <linux/mm.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>
31
32 #include <asm/system.h>
33
34 #include "iostat.h"
35
36 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
37
38 static int nfs_pagein_one(struct list_head *, struct inode *);
39 static const struct rpc_call_ops nfs_read_partial_ops;
40 static const struct rpc_call_ops nfs_read_full_ops;
41
42 static kmem_cache_t *nfs_rdata_cachep;
43 mempool_t *nfs_rdata_mempool;
44
45 #define MIN_POOL_READ   (32)
46
47 void nfs_readdata_release(void *data)
48 {
49         nfs_readdata_free(data);
50 }
51
52 static
53 unsigned int nfs_page_length(struct inode *inode, struct page *page)
54 {
55         loff_t i_size = i_size_read(inode);
56         unsigned long idx;
57
58         if (i_size <= 0)
59                 return 0;
60         idx = (i_size - 1) >> PAGE_CACHE_SHIFT;
61         if (page->index > idx)
62                 return 0;
63         if (page->index != idx)
64                 return PAGE_CACHE_SIZE;
65         return 1 + ((i_size - 1) & (PAGE_CACHE_SIZE - 1));
66 }
67
68 static
69 int nfs_return_empty_page(struct page *page)
70 {
71         memclear_highpage_flush(page, 0, PAGE_CACHE_SIZE);
72         SetPageUptodate(page);
73         unlock_page(page);
74         return 0;
75 }
76
77 /*
78  * Read a page synchronously.
79  */
80 static int nfs_readpage_sync(struct nfs_open_context *ctx, struct inode *inode,
81                 struct page *page)
82 {
83         unsigned int    rsize = NFS_SERVER(inode)->rsize;
84         unsigned int    count = PAGE_CACHE_SIZE;
85         int             result;
86         struct nfs_read_data *rdata;
87
88         rdata = nfs_readdata_alloc(1);
89         if (!rdata)
90                 return -ENOMEM;
91
92         memset(rdata, 0, sizeof(*rdata));
93         rdata->flags = (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
94         rdata->cred = ctx->cred;
95         rdata->inode = inode;
96         INIT_LIST_HEAD(&rdata->pages);
97         rdata->args.fh = NFS_FH(inode);
98         rdata->args.context = ctx;
99         rdata->args.pages = &page;
100         rdata->args.pgbase = 0UL;
101         rdata->args.count = rsize;
102         rdata->res.fattr = &rdata->fattr;
103
104         dprintk("NFS: nfs_readpage_sync(%p)\n", page);
105
106         /*
107          * This works now because the socket layer never tries to DMA
108          * into this buffer directly.
109          */
110         do {
111                 if (count < rsize)
112                         rdata->args.count = count;
113                 rdata->res.count = rdata->args.count;
114                 rdata->args.offset = page_offset(page) + rdata->args.pgbase;
115
116                 dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
117                         NFS_SERVER(inode)->hostname,
118                         inode->i_sb->s_id,
119                         (long long)NFS_FILEID(inode),
120                         (unsigned long long)rdata->args.pgbase,
121                         rdata->args.count);
122
123                 lock_kernel();
124                 result = NFS_PROTO(inode)->read(rdata);
125                 unlock_kernel();
126
127                 /*
128                  * Even if we had a partial success we can't mark the page
129                  * cache valid.
130                  */
131                 if (result < 0) {
132                         if (result == -EISDIR)
133                                 result = -EINVAL;
134                         goto io_error;
135                 }
136                 count -= result;
137                 rdata->args.pgbase += result;
138                 nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, result);
139
140                 /* Note: result == 0 should only happen if we're caching
141                  * a write that extends the file and punches a hole.
142                  */
143                 if (rdata->res.eof != 0 || result == 0)
144                         break;
145         } while (count);
146         spin_lock(&inode->i_lock);
147         NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME;
148         spin_unlock(&inode->i_lock);
149
150         if (count)
151                 memclear_highpage_flush(page, rdata->args.pgbase, count);
152         SetPageUptodate(page);
153         if (PageError(page))
154                 ClearPageError(page);
155         result = 0;
156
157 io_error:
158         unlock_page(page);
159         nfs_readdata_free(rdata);
160         return result;
161 }
162
163 static int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
164                 struct page *page)
165 {
166         LIST_HEAD(one_request);
167         struct nfs_page *new;
168         unsigned int len;
169
170         len = nfs_page_length(inode, page);
171         if (len == 0)
172                 return nfs_return_empty_page(page);
173         new = nfs_create_request(ctx, inode, page, 0, len);
174         if (IS_ERR(new)) {
175                 unlock_page(page);
176                 return PTR_ERR(new);
177         }
178         if (len < PAGE_CACHE_SIZE)
179                 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
180
181         nfs_list_add_request(new, &one_request);
182         nfs_pagein_one(&one_request, inode);
183         return 0;
184 }
185
186 static void nfs_readpage_release(struct nfs_page *req)
187 {
188         unlock_page(req->wb_page);
189
190         dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
191                         req->wb_context->dentry->d_inode->i_sb->s_id,
192                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
193                         req->wb_bytes,
194                         (long long)req_offset(req));
195         nfs_clear_request(req);
196         nfs_release_request(req);
197 }
198
199 /*
200  * Set up the NFS read request struct
201  */
202 static void nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
203                 const struct rpc_call_ops *call_ops,
204                 unsigned int count, unsigned int offset)
205 {
206         struct inode            *inode;
207         int flags;
208
209         data->req         = req;
210         data->inode       = inode = req->wb_context->dentry->d_inode;
211         data->cred        = req->wb_context->cred;
212
213         data->args.fh     = NFS_FH(inode);
214         data->args.offset = req_offset(req) + offset;
215         data->args.pgbase = req->wb_pgbase + offset;
216         data->args.pages  = data->pagevec;
217         data->args.count  = count;
218         data->args.context = req->wb_context;
219
220         data->res.fattr   = &data->fattr;
221         data->res.count   = count;
222         data->res.eof     = 0;
223         nfs_fattr_init(&data->fattr);
224
225         /* Set up the initial task struct. */
226         flags = RPC_TASK_ASYNC | (IS_SWAPFILE(inode)? NFS_RPC_SWAPFLAGS : 0);
227         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
228         NFS_PROTO(inode)->read_setup(data);
229
230         data->task.tk_cookie = (unsigned long)inode;
231
232         dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
233                         data->task.tk_pid,
234                         inode->i_sb->s_id,
235                         (long long)NFS_FILEID(inode),
236                         count,
237                         (unsigned long long)data->args.offset);
238 }
239
240 static void
241 nfs_async_read_error(struct list_head *head)
242 {
243         struct nfs_page *req;
244
245         while (!list_empty(head)) {
246                 req = nfs_list_entry(head->next);
247                 nfs_list_remove_request(req);
248                 SetPageError(req->wb_page);
249                 nfs_readpage_release(req);
250         }
251 }
252
253 /*
254  * Start an async read operation
255  */
256 static void nfs_execute_read(struct nfs_read_data *data)
257 {
258         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
259         sigset_t oldset;
260
261         rpc_clnt_sigmask(clnt, &oldset);
262         lock_kernel();
263         rpc_execute(&data->task);
264         unlock_kernel();
265         rpc_clnt_sigunmask(clnt, &oldset);
266 }
267
268 /*
269  * Generate multiple requests to fill a single page.
270  *
271  * We optimize to reduce the number of read operations on the wire.  If we
272  * detect that we're reading a page, or an area of a page, that is past the
273  * end of file, we do not generate NFS read operations but just clear the
274  * parts of the page that would have come back zero from the server anyway.
275  *
276  * We rely on the cached value of i_size to make this determination; another
277  * client can fill pages on the server past our cached end-of-file, but we
278  * won't see the new data until our attribute cache is updated.  This is more
279  * or less conventional NFS client behavior.
280  */
281 static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
282 {
283         struct nfs_page *req = nfs_list_entry(head->next);
284         struct page *page = req->wb_page;
285         struct nfs_read_data *data;
286         unsigned int rsize = NFS_SERVER(inode)->rsize;
287         unsigned int nbytes, offset;
288         int requests = 0;
289         LIST_HEAD(list);
290
291         nfs_list_remove_request(req);
292
293         nbytes = req->wb_bytes;
294         for(;;) {
295                 data = nfs_readdata_alloc(1);
296                 if (!data)
297                         goto out_bad;
298                 INIT_LIST_HEAD(&data->pages);
299                 list_add(&data->pages, &list);
300                 requests++;
301                 if (nbytes <= rsize)
302                         break;
303                 nbytes -= rsize;
304         }
305         atomic_set(&req->wb_complete, requests);
306
307         ClearPageError(page);
308         offset = 0;
309         nbytes = req->wb_bytes;
310         do {
311                 data = list_entry(list.next, struct nfs_read_data, pages);
312                 list_del_init(&data->pages);
313
314                 data->pagevec[0] = page;
315
316                 if (nbytes > rsize) {
317                         nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
318                                         rsize, offset);
319                         offset += rsize;
320                         nbytes -= rsize;
321                 } else {
322                         nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
323                                         nbytes, offset);
324                         nbytes = 0;
325                 }
326                 nfs_execute_read(data);
327         } while (nbytes != 0);
328
329         return 0;
330
331 out_bad:
332         while (!list_empty(&list)) {
333                 data = list_entry(list.next, struct nfs_read_data, pages);
334                 list_del(&data->pages);
335                 nfs_readdata_free(data);
336         }
337         SetPageError(page);
338         nfs_readpage_release(req);
339         return -ENOMEM;
340 }
341
342 static int nfs_pagein_one(struct list_head *head, struct inode *inode)
343 {
344         struct nfs_page         *req;
345         struct page             **pages;
346         struct nfs_read_data    *data;
347         unsigned int            count;
348
349         if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
350                 return nfs_pagein_multi(head, inode);
351
352         data = nfs_readdata_alloc(NFS_SERVER(inode)->rpages);
353         if (!data)
354                 goto out_bad;
355
356         INIT_LIST_HEAD(&data->pages);
357         pages = data->pagevec;
358         count = 0;
359         while (!list_empty(head)) {
360                 req = nfs_list_entry(head->next);
361                 nfs_list_remove_request(req);
362                 nfs_list_add_request(req, &data->pages);
363                 ClearPageError(req->wb_page);
364                 *pages++ = req->wb_page;
365                 count += req->wb_bytes;
366         }
367         req = nfs_list_entry(data->pages.next);
368
369         nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
370
371         nfs_execute_read(data);
372         return 0;
373 out_bad:
374         nfs_async_read_error(head);
375         return -ENOMEM;
376 }
377
378 static int
379 nfs_pagein_list(struct list_head *head, int rpages)
380 {
381         LIST_HEAD(one_request);
382         struct nfs_page         *req;
383         int                     error = 0;
384         unsigned int            pages = 0;
385
386         while (!list_empty(head)) {
387                 pages += nfs_coalesce_requests(head, &one_request, rpages);
388                 req = nfs_list_entry(one_request.next);
389                 error = nfs_pagein_one(&one_request, req->wb_context->dentry->d_inode);
390                 if (error < 0)
391                         break;
392         }
393         if (error >= 0)
394                 return pages;
395
396         nfs_async_read_error(head);
397         return error;
398 }
399
400 /*
401  * Handle a read reply that fills part of a page.
402  */
403 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
404 {
405         struct nfs_read_data *data = calldata;
406         struct nfs_page *req = data->req;
407         struct page *page = req->wb_page;
408  
409         if (nfs_readpage_result(task, data) != 0)
410                 return;
411         if (task->tk_status >= 0) {
412                 unsigned int request = data->args.count;
413                 unsigned int result = data->res.count;
414
415                 if (result < request) {
416                         memclear_highpage_flush(page,
417                                                 data->args.pgbase + result,
418                                                 request - result);
419                 }
420         } else
421                 SetPageError(page);
422
423         if (atomic_dec_and_test(&req->wb_complete)) {
424                 if (!PageError(page))
425                         SetPageUptodate(page);
426                 nfs_readpage_release(req);
427         }
428 }
429
430 static const struct rpc_call_ops nfs_read_partial_ops = {
431         .rpc_call_done = nfs_readpage_result_partial,
432         .rpc_release = nfs_readdata_release,
433 };
434
435 /*
436  * This is the callback from RPC telling us whether a reply was
437  * received or some error occurred (timeout or socket shutdown).
438  */
439 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
440 {
441         struct nfs_read_data *data = calldata;
442         unsigned int count = data->res.count;
443
444         if (nfs_readpage_result(task, data) != 0)
445                 return;
446         while (!list_empty(&data->pages)) {
447                 struct nfs_page *req = nfs_list_entry(data->pages.next);
448                 struct page *page = req->wb_page;
449                 nfs_list_remove_request(req);
450
451                 if (task->tk_status >= 0) {
452                         if (count < PAGE_CACHE_SIZE) {
453                                 if (count < req->wb_bytes)
454                                         memclear_highpage_flush(page,
455                                                         req->wb_pgbase + count,
456                                                         req->wb_bytes - count);
457                                 count = 0;
458                         } else
459                                 count -= PAGE_CACHE_SIZE;
460                         SetPageUptodate(page);
461                 } else
462                         SetPageError(page);
463                 nfs_readpage_release(req);
464         }
465 }
466
467 static const struct rpc_call_ops nfs_read_full_ops = {
468         .rpc_call_done = nfs_readpage_result_full,
469         .rpc_release = nfs_readdata_release,
470 };
471
472 /*
473  * This is the callback from RPC telling us whether a reply was
474  * received or some error occurred (timeout or socket shutdown).
475  */
476 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
477 {
478         struct nfs_readargs *argp = &data->args;
479         struct nfs_readres *resp = &data->res;
480         int status;
481
482         dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
483                 task->tk_pid, task->tk_status);
484
485         status = NFS_PROTO(data->inode)->read_done(task, data);
486         if (status != 0)
487                 return status;
488
489         nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, resp->count);
490
491         /* Is this a short read? */
492         if (task->tk_status >= 0 && resp->count < argp->count && !resp->eof) {
493                 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
494                 /* Has the server at least made some progress? */
495                 if (resp->count != 0) {
496                         /* Yes, so retry the read at the end of the data */
497                         argp->offset += resp->count;
498                         argp->pgbase += resp->count;
499                         argp->count -= resp->count;
500                         rpc_restart_call(task);
501                         return -EAGAIN;
502                 }
503                 task->tk_status = -EIO;
504         }
505         spin_lock(&data->inode->i_lock);
506         NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME;
507         spin_unlock(&data->inode->i_lock);
508         return 0;
509 }
510
511 /*
512  * Read a page over NFS.
513  * We read the page synchronously in the following case:
514  *  -   The error flag is set for this page. This happens only when a
515  *      previous async read operation failed.
516  */
517 int nfs_readpage(struct file *file, struct page *page)
518 {
519         struct nfs_open_context *ctx;
520         struct inode *inode = page->mapping->host;
521         int             error;
522
523         dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
524                 page, PAGE_CACHE_SIZE, page->index);
525         nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
526         nfs_add_stats(inode, NFSIOS_READPAGES, 1);
527
528         /*
529          * Try to flush any pending writes to the file..
530          *
531          * NOTE! Because we own the page lock, there cannot
532          * be any new pending writes generated at this point
533          * for this page (other pages can be written to).
534          */
535         error = nfs_wb_page(inode, page);
536         if (error)
537                 goto out_error;
538
539         if (file == NULL) {
540                 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
541                 if (ctx == NULL)
542                         return -EBADF;
543         } else
544                 ctx = get_nfs_open_context((struct nfs_open_context *)
545                                 file->private_data);
546         if (!IS_SYNC(inode)) {
547                 error = nfs_readpage_async(ctx, inode, page);
548                 goto out;
549         }
550
551         error = nfs_readpage_sync(ctx, inode, page);
552         if (error < 0 && IS_SWAPFILE(inode))
553                 printk("Aiee.. nfs swap-in of page failed!\n");
554 out:
555         put_nfs_open_context(ctx);
556         return error;
557
558 out_error:
559         unlock_page(page);
560         return error;
561 }
562
563 struct nfs_readdesc {
564         struct list_head *head;
565         struct nfs_open_context *ctx;
566 };
567
568 static int
569 readpage_async_filler(void *data, struct page *page)
570 {
571         struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
572         struct inode *inode = page->mapping->host;
573         struct nfs_page *new;
574         unsigned int len;
575
576         nfs_wb_page(inode, page);
577         len = nfs_page_length(inode, page);
578         if (len == 0)
579                 return nfs_return_empty_page(page);
580         new = nfs_create_request(desc->ctx, inode, page, 0, len);
581         if (IS_ERR(new)) {
582                         SetPageError(page);
583                         unlock_page(page);
584                         return PTR_ERR(new);
585         }
586         if (len < PAGE_CACHE_SIZE)
587                 memclear_highpage_flush(page, len, PAGE_CACHE_SIZE - len);
588         nfs_list_add_request(new, desc->head);
589         return 0;
590 }
591
592 int nfs_readpages(struct file *filp, struct address_space *mapping,
593                 struct list_head *pages, unsigned nr_pages)
594 {
595         LIST_HEAD(head);
596         struct nfs_readdesc desc = {
597                 .head           = &head,
598         };
599         struct inode *inode = mapping->host;
600         struct nfs_server *server = NFS_SERVER(inode);
601         int ret;
602
603         dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
604                         inode->i_sb->s_id,
605                         (long long)NFS_FILEID(inode),
606                         nr_pages);
607         nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
608
609         if (filp == NULL) {
610                 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
611                 if (desc.ctx == NULL)
612                         return -EBADF;
613         } else
614                 desc.ctx = get_nfs_open_context((struct nfs_open_context *)
615                                 filp->private_data);
616         ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
617         if (!list_empty(&head)) {
618                 int err = nfs_pagein_list(&head, server->rpages);
619                 if (!ret)
620                         nfs_add_stats(inode, NFSIOS_READPAGES, err);
621                         ret = err;
622         }
623         put_nfs_open_context(desc.ctx);
624         return ret;
625 }
626
627 int nfs_init_readpagecache(void)
628 {
629         nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
630                                              sizeof(struct nfs_read_data),
631                                              0, SLAB_HWCACHE_ALIGN,
632                                              NULL, NULL);
633         if (nfs_rdata_cachep == NULL)
634                 return -ENOMEM;
635
636         nfs_rdata_mempool = mempool_create(MIN_POOL_READ,
637                                            mempool_alloc_slab,
638                                            mempool_free_slab,
639                                            nfs_rdata_cachep);
640         if (nfs_rdata_mempool == NULL)
641                 return -ENOMEM;
642
643         return 0;
644 }
645
646 void nfs_destroy_readpagecache(void)
647 {
648         mempool_destroy(nfs_rdata_mempool);
649         if (kmem_cache_destroy(nfs_rdata_cachep))
650                 printk(KERN_INFO "nfs_read_data: not all structures were freed\n");
651 }