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>
36 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
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
;
42 static kmem_cache_t
*nfs_rdata_cachep
;
43 static mempool_t
*nfs_rdata_mempool
;
45 #define MIN_POOL_READ (32)
47 struct nfs_read_data
*nfs_readdata_alloc(unsigned int pagecount
)
49 struct nfs_read_data
*p
= mempool_alloc(nfs_rdata_mempool
, SLAB_NOFS
);
52 memset(p
, 0, sizeof(*p
));
53 INIT_LIST_HEAD(&p
->pages
);
54 if (pagecount
<= ARRAY_SIZE(p
->page_array
))
55 p
->pagevec
= p
->page_array
;
57 p
->pagevec
= kcalloc(pagecount
, sizeof(struct page
*), GFP_NOFS
);
59 mempool_free(p
, nfs_rdata_mempool
);
67 void nfs_readdata_free(struct nfs_read_data
*p
)
69 if (p
&& (p
->pagevec
!= &p
->page_array
[0]))
71 mempool_free(p
, nfs_rdata_mempool
);
74 void nfs_readdata_release(void *data
)
76 nfs_readdata_free(data
);
80 unsigned int nfs_page_length(struct inode
*inode
, struct page
*page
)
82 loff_t i_size
= i_size_read(inode
);
87 idx
= (i_size
- 1) >> PAGE_CACHE_SHIFT
;
88 if (page
->index
> idx
)
90 if (page
->index
!= idx
)
91 return PAGE_CACHE_SIZE
;
92 return 1 + ((i_size
- 1) & (PAGE_CACHE_SIZE
- 1));
96 int nfs_return_empty_page(struct page
*page
)
98 memclear_highpage_flush(page
, 0, PAGE_CACHE_SIZE
);
99 SetPageUptodate(page
);
104 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data
*data
)
106 unsigned int remainder
= data
->args
.count
- data
->res
.count
;
107 unsigned int base
= data
->args
.pgbase
+ data
->res
.count
;
111 if (data
->res
.eof
== 0 || remainder
== 0)
114 * Note: "remainder" can never be negative, since we check for
115 * this in the XDR code.
117 pages
= &data
->args
.pages
[base
>> PAGE_CACHE_SHIFT
];
118 base
&= ~PAGE_CACHE_MASK
;
119 pglen
= PAGE_CACHE_SIZE
- base
;
120 if (pglen
< remainder
)
121 memclear_highpage_flush(*pages
, base
, pglen
);
123 memclear_highpage_flush(*pages
, base
, remainder
);
127 * Read a page synchronously.
129 static int nfs_readpage_sync(struct nfs_open_context
*ctx
, struct inode
*inode
,
132 unsigned int rsize
= NFS_SERVER(inode
)->rsize
;
133 unsigned int count
= PAGE_CACHE_SIZE
;
135 struct nfs_read_data
*rdata
;
137 rdata
= nfs_readdata_alloc(1);
141 memset(rdata
, 0, sizeof(*rdata
));
142 rdata
->flags
= (IS_SWAPFILE(inode
)? NFS_RPC_SWAPFLAGS
: 0);
143 rdata
->cred
= ctx
->cred
;
144 rdata
->inode
= inode
;
145 INIT_LIST_HEAD(&rdata
->pages
);
146 rdata
->args
.fh
= NFS_FH(inode
);
147 rdata
->args
.context
= ctx
;
148 rdata
->args
.pages
= &page
;
149 rdata
->args
.pgbase
= 0UL;
150 rdata
->args
.count
= rsize
;
151 rdata
->res
.fattr
= &rdata
->fattr
;
153 dprintk("NFS: nfs_readpage_sync(%p)\n", page
);
156 * This works now because the socket layer never tries to DMA
157 * into this buffer directly.
161 rdata
->args
.count
= count
;
162 rdata
->res
.count
= rdata
->args
.count
;
163 rdata
->args
.offset
= page_offset(page
) + rdata
->args
.pgbase
;
165 dprintk("NFS: nfs_proc_read(%s, (%s/%Ld), %Lu, %u)\n",
166 NFS_SERVER(inode
)->hostname
,
168 (long long)NFS_FILEID(inode
),
169 (unsigned long long)rdata
->args
.pgbase
,
173 result
= NFS_PROTO(inode
)->read(rdata
);
177 * Even if we had a partial success we can't mark the page
181 if (result
== -EISDIR
)
186 rdata
->args
.pgbase
+= result
;
187 nfs_add_stats(inode
, NFSIOS_SERVERREADBYTES
, result
);
189 /* Note: result == 0 should only happen if we're caching
190 * a write that extends the file and punches a hole.
192 if (rdata
->res
.eof
!= 0 || result
== 0)
195 spin_lock(&inode
->i_lock
);
196 NFS_I(inode
)->cache_validity
|= NFS_INO_INVALID_ATIME
;
197 spin_unlock(&inode
->i_lock
);
199 nfs_readpage_truncate_uninitialised_page(rdata
);
200 if (rdata
->res
.eof
|| rdata
->res
.count
== rdata
->args
.count
)
201 SetPageUptodate(page
);
206 nfs_readdata_free(rdata
);
210 static int nfs_readpage_async(struct nfs_open_context
*ctx
, struct inode
*inode
,
213 LIST_HEAD(one_request
);
214 struct nfs_page
*new;
217 len
= nfs_page_length(inode
, page
);
219 return nfs_return_empty_page(page
);
220 new = nfs_create_request(ctx
, inode
, page
, 0, len
);
225 if (len
< PAGE_CACHE_SIZE
)
226 memclear_highpage_flush(page
, len
, PAGE_CACHE_SIZE
- len
);
228 nfs_list_add_request(new, &one_request
);
229 nfs_pagein_one(&one_request
, inode
);
233 static void nfs_readpage_release(struct nfs_page
*req
)
235 unlock_page(req
->wb_page
);
237 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
238 req
->wb_context
->dentry
->d_inode
->i_sb
->s_id
,
239 (long long)NFS_FILEID(req
->wb_context
->dentry
->d_inode
),
241 (long long)req_offset(req
));
242 nfs_clear_request(req
);
243 nfs_release_request(req
);
247 * Set up the NFS read request struct
249 static void nfs_read_rpcsetup(struct nfs_page
*req
, struct nfs_read_data
*data
,
250 const struct rpc_call_ops
*call_ops
,
251 unsigned int count
, unsigned int offset
)
257 data
->inode
= inode
= req
->wb_context
->dentry
->d_inode
;
258 data
->cred
= req
->wb_context
->cred
;
260 data
->args
.fh
= NFS_FH(inode
);
261 data
->args
.offset
= req_offset(req
) + offset
;
262 data
->args
.pgbase
= req
->wb_pgbase
+ offset
;
263 data
->args
.pages
= data
->pagevec
;
264 data
->args
.count
= count
;
265 data
->args
.context
= req
->wb_context
;
267 data
->res
.fattr
= &data
->fattr
;
268 data
->res
.count
= count
;
270 nfs_fattr_init(&data
->fattr
);
272 /* Set up the initial task struct. */
273 flags
= RPC_TASK_ASYNC
| (IS_SWAPFILE(inode
)? NFS_RPC_SWAPFLAGS
: 0);
274 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), flags
, call_ops
, data
);
275 NFS_PROTO(inode
)->read_setup(data
);
277 data
->task
.tk_cookie
= (unsigned long)inode
;
279 dprintk("NFS: %4d initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
282 (long long)NFS_FILEID(inode
),
284 (unsigned long long)data
->args
.offset
);
288 nfs_async_read_error(struct list_head
*head
)
290 struct nfs_page
*req
;
292 while (!list_empty(head
)) {
293 req
= nfs_list_entry(head
->next
);
294 nfs_list_remove_request(req
);
295 SetPageError(req
->wb_page
);
296 nfs_readpage_release(req
);
301 * Start an async read operation
303 static void nfs_execute_read(struct nfs_read_data
*data
)
305 struct rpc_clnt
*clnt
= NFS_CLIENT(data
->inode
);
308 rpc_clnt_sigmask(clnt
, &oldset
);
310 rpc_execute(&data
->task
);
312 rpc_clnt_sigunmask(clnt
, &oldset
);
316 * Generate multiple requests to fill a single page.
318 * We optimize to reduce the number of read operations on the wire. If we
319 * detect that we're reading a page, or an area of a page, that is past the
320 * end of file, we do not generate NFS read operations but just clear the
321 * parts of the page that would have come back zero from the server anyway.
323 * We rely on the cached value of i_size to make this determination; another
324 * client can fill pages on the server past our cached end-of-file, but we
325 * won't see the new data until our attribute cache is updated. This is more
326 * or less conventional NFS client behavior.
328 static int nfs_pagein_multi(struct list_head
*head
, struct inode
*inode
)
330 struct nfs_page
*req
= nfs_list_entry(head
->next
);
331 struct page
*page
= req
->wb_page
;
332 struct nfs_read_data
*data
;
333 unsigned int rsize
= NFS_SERVER(inode
)->rsize
;
334 unsigned int nbytes
, offset
;
338 nfs_list_remove_request(req
);
340 nbytes
= req
->wb_bytes
;
342 data
= nfs_readdata_alloc(1);
345 INIT_LIST_HEAD(&data
->pages
);
346 list_add(&data
->pages
, &list
);
352 atomic_set(&req
->wb_complete
, requests
);
354 ClearPageError(page
);
356 nbytes
= req
->wb_bytes
;
358 data
= list_entry(list
.next
, struct nfs_read_data
, pages
);
359 list_del_init(&data
->pages
);
361 data
->pagevec
[0] = page
;
363 if (nbytes
> rsize
) {
364 nfs_read_rpcsetup(req
, data
, &nfs_read_partial_ops
,
369 nfs_read_rpcsetup(req
, data
, &nfs_read_partial_ops
,
373 nfs_execute_read(data
);
374 } while (nbytes
!= 0);
379 while (!list_empty(&list
)) {
380 data
= list_entry(list
.next
, struct nfs_read_data
, pages
);
381 list_del(&data
->pages
);
382 nfs_readdata_free(data
);
385 nfs_readpage_release(req
);
389 static int nfs_pagein_one(struct list_head
*head
, struct inode
*inode
)
391 struct nfs_page
*req
;
393 struct nfs_read_data
*data
;
396 if (NFS_SERVER(inode
)->rsize
< PAGE_CACHE_SIZE
)
397 return nfs_pagein_multi(head
, inode
);
399 data
= nfs_readdata_alloc(NFS_SERVER(inode
)->rpages
);
403 INIT_LIST_HEAD(&data
->pages
);
404 pages
= data
->pagevec
;
406 while (!list_empty(head
)) {
407 req
= nfs_list_entry(head
->next
);
408 nfs_list_remove_request(req
);
409 nfs_list_add_request(req
, &data
->pages
);
410 ClearPageError(req
->wb_page
);
411 *pages
++ = req
->wb_page
;
412 count
+= req
->wb_bytes
;
414 req
= nfs_list_entry(data
->pages
.next
);
416 nfs_read_rpcsetup(req
, data
, &nfs_read_full_ops
, count
, 0);
418 nfs_execute_read(data
);
421 nfs_async_read_error(head
);
426 nfs_pagein_list(struct list_head
*head
, int rpages
)
428 LIST_HEAD(one_request
);
429 struct nfs_page
*req
;
431 unsigned int pages
= 0;
433 while (!list_empty(head
)) {
434 pages
+= nfs_coalesce_requests(head
, &one_request
, rpages
);
435 req
= nfs_list_entry(one_request
.next
);
436 error
= nfs_pagein_one(&one_request
, req
->wb_context
->dentry
->d_inode
);
443 nfs_async_read_error(head
);
448 * Handle a read reply that fills part of a page.
450 static void nfs_readpage_result_partial(struct rpc_task
*task
, void *calldata
)
452 struct nfs_read_data
*data
= calldata
;
453 struct nfs_page
*req
= data
->req
;
454 struct page
*page
= req
->wb_page
;
456 if (likely(task
->tk_status
>= 0))
457 nfs_readpage_truncate_uninitialised_page(data
);
460 if (nfs_readpage_result(task
, data
) != 0)
462 if (atomic_dec_and_test(&req
->wb_complete
)) {
463 if (!PageError(page
))
464 SetPageUptodate(page
);
465 nfs_readpage_release(req
);
469 static const struct rpc_call_ops nfs_read_partial_ops
= {
470 .rpc_call_done
= nfs_readpage_result_partial
,
471 .rpc_release
= nfs_readdata_release
,
474 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data
*data
)
476 unsigned int count
= data
->res
.count
;
477 unsigned int base
= data
->args
.pgbase
;
480 if (unlikely(count
== 0))
482 pages
= &data
->args
.pages
[base
>> PAGE_CACHE_SHIFT
];
483 base
&= ~PAGE_CACHE_MASK
;
485 for (;count
>= PAGE_CACHE_SIZE
; count
-= PAGE_CACHE_SIZE
, pages
++)
486 SetPageUptodate(*pages
);
488 * Was this an eof or a short read? If the latter, don't mark the page
491 if (count
> 0 && (data
->res
.eof
|| data
->args
.count
== data
->res
.count
))
492 SetPageUptodate(*pages
);
495 static void nfs_readpage_set_pages_error(struct nfs_read_data
*data
)
497 unsigned int count
= data
->args
.count
;
498 unsigned int base
= data
->args
.pgbase
;
501 pages
= &data
->args
.pages
[base
>> PAGE_CACHE_SHIFT
];
502 base
&= ~PAGE_CACHE_MASK
;
504 for (;count
>= PAGE_CACHE_SIZE
; count
-= PAGE_CACHE_SIZE
, pages
++)
505 SetPageError(*pages
);
509 * This is the callback from RPC telling us whether a reply was
510 * received or some error occurred (timeout or socket shutdown).
512 static void nfs_readpage_result_full(struct rpc_task
*task
, void *calldata
)
514 struct nfs_read_data
*data
= calldata
;
517 * Note: nfs_readpage_result may change the values of
518 * data->args. In the multi-page case, we therefore need
519 * to ensure that we call the next nfs_readpage_set_page_uptodate()
520 * first in the multi-page case.
522 if (likely(task
->tk_status
>= 0)) {
523 nfs_readpage_truncate_uninitialised_page(data
);
524 nfs_readpage_set_pages_uptodate(data
);
526 nfs_readpage_set_pages_error(data
);
527 if (nfs_readpage_result(task
, data
) != 0)
529 while (!list_empty(&data
->pages
)) {
530 struct nfs_page
*req
= nfs_list_entry(data
->pages
.next
);
532 nfs_list_remove_request(req
);
533 nfs_readpage_release(req
);
537 static const struct rpc_call_ops nfs_read_full_ops
= {
538 .rpc_call_done
= nfs_readpage_result_full
,
539 .rpc_release
= nfs_readdata_release
,
543 * This is the callback from RPC telling us whether a reply was
544 * received or some error occurred (timeout or socket shutdown).
546 int nfs_readpage_result(struct rpc_task
*task
, struct nfs_read_data
*data
)
548 struct nfs_readargs
*argp
= &data
->args
;
549 struct nfs_readres
*resp
= &data
->res
;
552 dprintk("NFS: %4d nfs_readpage_result, (status %d)\n",
553 task
->tk_pid
, task
->tk_status
);
555 status
= NFS_PROTO(data
->inode
)->read_done(task
, data
);
559 nfs_add_stats(data
->inode
, NFSIOS_SERVERREADBYTES
, resp
->count
);
561 /* Is this a short read? */
562 if (task
->tk_status
>= 0 && resp
->count
< argp
->count
&& !resp
->eof
) {
563 nfs_inc_stats(data
->inode
, NFSIOS_SHORTREAD
);
564 /* Has the server at least made some progress? */
565 if (resp
->count
!= 0) {
566 /* Yes, so retry the read at the end of the data */
567 argp
->offset
+= resp
->count
;
568 argp
->pgbase
+= resp
->count
;
569 argp
->count
-= resp
->count
;
570 rpc_restart_call(task
);
573 task
->tk_status
= -EIO
;
575 spin_lock(&data
->inode
->i_lock
);
576 NFS_I(data
->inode
)->cache_validity
|= NFS_INO_INVALID_ATIME
;
577 spin_unlock(&data
->inode
->i_lock
);
582 * Read a page over NFS.
583 * We read the page synchronously in the following case:
584 * - The error flag is set for this page. This happens only when a
585 * previous async read operation failed.
587 int nfs_readpage(struct file
*file
, struct page
*page
)
589 struct nfs_open_context
*ctx
;
590 struct inode
*inode
= page
->mapping
->host
;
593 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
594 page
, PAGE_CACHE_SIZE
, page
->index
);
595 nfs_inc_stats(inode
, NFSIOS_VFSREADPAGE
);
596 nfs_add_stats(inode
, NFSIOS_READPAGES
, 1);
599 * Try to flush any pending writes to the file..
601 * NOTE! Because we own the page lock, there cannot
602 * be any new pending writes generated at this point
603 * for this page (other pages can be written to).
605 error
= nfs_wb_page(inode
, page
);
610 ctx
= nfs_find_open_context(inode
, NULL
, FMODE_READ
);
614 ctx
= get_nfs_open_context((struct nfs_open_context
*)
616 if (!IS_SYNC(inode
)) {
617 error
= nfs_readpage_async(ctx
, inode
, page
);
621 error
= nfs_readpage_sync(ctx
, inode
, page
);
622 if (error
< 0 && IS_SWAPFILE(inode
))
623 printk("Aiee.. nfs swap-in of page failed!\n");
625 put_nfs_open_context(ctx
);
633 struct nfs_readdesc
{
634 struct list_head
*head
;
635 struct nfs_open_context
*ctx
;
639 readpage_async_filler(void *data
, struct page
*page
)
641 struct nfs_readdesc
*desc
= (struct nfs_readdesc
*)data
;
642 struct inode
*inode
= page
->mapping
->host
;
643 struct nfs_page
*new;
646 nfs_wb_page(inode
, page
);
647 len
= nfs_page_length(inode
, page
);
649 return nfs_return_empty_page(page
);
650 new = nfs_create_request(desc
->ctx
, inode
, page
, 0, len
);
656 if (len
< PAGE_CACHE_SIZE
)
657 memclear_highpage_flush(page
, len
, PAGE_CACHE_SIZE
- len
);
658 nfs_list_add_request(new, desc
->head
);
662 int nfs_readpages(struct file
*filp
, struct address_space
*mapping
,
663 struct list_head
*pages
, unsigned nr_pages
)
666 struct nfs_readdesc desc
= {
669 struct inode
*inode
= mapping
->host
;
670 struct nfs_server
*server
= NFS_SERVER(inode
);
673 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
675 (long long)NFS_FILEID(inode
),
677 nfs_inc_stats(inode
, NFSIOS_VFSREADPAGES
);
680 desc
.ctx
= nfs_find_open_context(inode
, NULL
, FMODE_READ
);
681 if (desc
.ctx
== NULL
)
684 desc
.ctx
= get_nfs_open_context((struct nfs_open_context
*)
686 ret
= read_cache_pages(mapping
, pages
, readpage_async_filler
, &desc
);
687 if (!list_empty(&head
)) {
688 int err
= nfs_pagein_list(&head
, server
->rpages
);
690 nfs_add_stats(inode
, NFSIOS_READPAGES
, err
);
693 put_nfs_open_context(desc
.ctx
);
697 int __init
nfs_init_readpagecache(void)
699 nfs_rdata_cachep
= kmem_cache_create("nfs_read_data",
700 sizeof(struct nfs_read_data
),
701 0, SLAB_HWCACHE_ALIGN
,
703 if (nfs_rdata_cachep
== NULL
)
706 nfs_rdata_mempool
= mempool_create_slab_pool(MIN_POOL_READ
,
708 if (nfs_rdata_mempool
== NULL
)
714 void __exit
nfs_destroy_readpagecache(void)
716 mempool_destroy(nfs_rdata_mempool
);
717 if (kmem_cache_destroy(nfs_rdata_cachep
))
718 printk(KERN_INFO
"nfs_read_data: not all structures were freed\n");