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>
29 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
31 static int nfs_pagein_multi(struct inode *, struct list_head *, unsigned int, size_t, int);
32 static int nfs_pagein_one(struct inode *, struct list_head *, unsigned int, size_t, int);
33 static const struct rpc_call_ops nfs_read_partial_ops;
34 static const struct rpc_call_ops nfs_read_full_ops;
36 static struct kmem_cache *nfs_rdata_cachep;
37 static mempool_t *nfs_rdata_mempool;
39 #define MIN_POOL_READ (32)
41 struct nfs_read_data *nfs_readdata_alloc(unsigned int pagecount)
43 struct nfs_read_data *p = mempool_alloc(nfs_rdata_mempool, GFP_NOFS);
46 memset(p, 0, sizeof(*p));
47 INIT_LIST_HEAD(&p->pages);
48 p->npages = pagecount;
49 p->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
50 if (pagecount <= ARRAY_SIZE(p->page_array))
51 p->pagevec = p->page_array;
53 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
55 mempool_free(p, nfs_rdata_mempool);
63 static void nfs_readdata_free(struct nfs_read_data *p)
65 if (p && (p->pagevec != &p->page_array[0]))
67 mempool_free(p, nfs_rdata_mempool);
70 void nfs_readdata_release(void *data)
72 struct nfs_read_data *rdata = data;
74 put_nfs_open_context(rdata->args.context);
75 nfs_readdata_free(rdata);
79 int nfs_return_empty_page(struct page *page)
81 zero_user(page, 0, PAGE_CACHE_SIZE);
82 SetPageUptodate(page);
87 static void nfs_readpage_truncate_uninitialised_page(struct nfs_read_data *data)
89 unsigned int remainder = data->args.count - data->res.count;
90 unsigned int base = data->args.pgbase + data->res.count;
94 if (data->res.eof == 0 || remainder == 0)
97 * Note: "remainder" can never be negative, since we check for
98 * this in the XDR code.
100 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
101 base &= ~PAGE_CACHE_MASK;
102 pglen = PAGE_CACHE_SIZE - base;
104 if (remainder <= pglen) {
105 zero_user(*pages, base, remainder);
108 zero_user(*pages, base, pglen);
111 pglen = PAGE_CACHE_SIZE;
116 int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
119 LIST_HEAD(one_request);
120 struct nfs_page *new;
123 len = nfs_page_length(page);
125 return nfs_return_empty_page(page);
126 new = nfs_create_request(ctx, inode, page, 0, len);
131 if (len < PAGE_CACHE_SIZE)
132 zero_user_segment(page, len, PAGE_CACHE_SIZE);
134 nfs_list_add_request(new, &one_request);
135 if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
136 nfs_pagein_multi(inode, &one_request, 1, len, 0);
138 nfs_pagein_one(inode, &one_request, 1, len, 0);
142 static void nfs_readpage_release(struct nfs_page *req)
144 struct inode *d_inode = req->wb_context->path.dentry->d_inode;
146 if (PageUptodate(req->wb_page))
147 nfs_readpage_to_fscache(d_inode, req->wb_page, 0);
149 unlock_page(req->wb_page);
151 dprintk("NFS: read done (%s/%Ld %d@%Ld)\n",
152 req->wb_context->path.dentry->d_inode->i_sb->s_id,
153 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
155 (long long)req_offset(req));
156 nfs_clear_request(req);
157 nfs_release_request(req);
161 * Set up the NFS read request struct
163 static int nfs_read_rpcsetup(struct nfs_page *req, struct nfs_read_data *data,
164 const struct rpc_call_ops *call_ops,
165 unsigned int count, unsigned int offset)
167 struct inode *inode = req->wb_context->path.dentry->d_inode;
168 int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
169 struct rpc_task *task;
170 struct rpc_message msg = {
171 .rpc_argp = &data->args,
172 .rpc_resp = &data->res,
173 .rpc_cred = req->wb_context->cred,
175 struct rpc_task_setup task_setup_data = {
177 .rpc_client = NFS_CLIENT(inode),
179 .callback_ops = call_ops,
180 .callback_data = data,
181 .workqueue = nfsiod_workqueue,
182 .flags = RPC_TASK_ASYNC | swap_flags,
187 data->cred = msg.rpc_cred;
189 data->args.fh = NFS_FH(inode);
190 data->args.offset = req_offset(req) + offset;
191 data->args.pgbase = req->wb_pgbase + offset;
192 data->args.pages = data->pagevec;
193 data->args.count = count;
194 data->args.context = get_nfs_open_context(req->wb_context);
196 data->res.fattr = &data->fattr;
197 data->res.count = count;
199 nfs_fattr_init(&data->fattr);
201 /* Set up the initial task struct. */
202 NFS_PROTO(inode)->read_setup(data, &msg);
204 dprintk("NFS: %5u initiated read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
207 (long long)NFS_FILEID(inode),
209 (unsigned long long)data->args.offset);
211 task = rpc_run_task(&task_setup_data);
213 return PTR_ERR(task);
219 nfs_async_read_error(struct list_head *head)
221 struct nfs_page *req;
223 while (!list_empty(head)) {
224 req = nfs_list_entry(head->next);
225 nfs_list_remove_request(req);
226 SetPageError(req->wb_page);
227 nfs_readpage_release(req);
232 * Generate multiple requests to fill a single page.
234 * We optimize to reduce the number of read operations on the wire. If we
235 * detect that we're reading a page, or an area of a page, that is past the
236 * end of file, we do not generate NFS read operations but just clear the
237 * parts of the page that would have come back zero from the server anyway.
239 * We rely on the cached value of i_size to make this determination; another
240 * client can fill pages on the server past our cached end-of-file, but we
241 * won't see the new data until our attribute cache is updated. This is more
242 * or less conventional NFS client behavior.
244 static int nfs_pagein_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags)
246 struct nfs_page *req = nfs_list_entry(head->next);
247 struct page *page = req->wb_page;
248 struct nfs_read_data *data;
249 size_t rsize = NFS_SERVER(inode)->rsize, nbytes;
255 nfs_list_remove_request(req);
259 size_t len = min(nbytes,rsize);
261 data = nfs_readdata_alloc(1);
264 list_add(&data->pages, &list);
267 } while(nbytes != 0);
268 atomic_set(&req->wb_complete, requests);
270 ClearPageError(page);
276 data = list_entry(list.next, struct nfs_read_data, pages);
277 list_del_init(&data->pages);
279 data->pagevec[0] = page;
283 ret2 = nfs_read_rpcsetup(req, data, &nfs_read_partial_ops,
289 } while (nbytes != 0);
294 while (!list_empty(&list)) {
295 data = list_entry(list.next, struct nfs_read_data, pages);
296 list_del(&data->pages);
297 nfs_readdata_free(data);
300 nfs_readpage_release(req);
304 static int nfs_pagein_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int flags)
306 struct nfs_page *req;
308 struct nfs_read_data *data;
311 data = nfs_readdata_alloc(npages);
315 pages = data->pagevec;
316 while (!list_empty(head)) {
317 req = nfs_list_entry(head->next);
318 nfs_list_remove_request(req);
319 nfs_list_add_request(req, &data->pages);
320 ClearPageError(req->wb_page);
321 *pages++ = req->wb_page;
323 req = nfs_list_entry(data->pages.next);
325 return nfs_read_rpcsetup(req, data, &nfs_read_full_ops, count, 0);
327 nfs_async_read_error(head);
332 * This is the callback from RPC telling us whether a reply was
333 * received or some error occurred (timeout or socket shutdown).
335 int nfs_readpage_result(struct rpc_task *task, struct nfs_read_data *data)
339 dprintk("NFS: %s: %5u, (status %d)\n", __func__, task->tk_pid,
342 status = NFS_PROTO(data->inode)->read_done(task, data);
346 nfs_add_stats(data->inode, NFSIOS_SERVERREADBYTES, data->res.count);
348 if (task->tk_status == -ESTALE) {
349 set_bit(NFS_INO_STALE, &NFS_I(data->inode)->flags);
350 nfs_mark_for_revalidate(data->inode);
355 static void nfs_readpage_retry(struct rpc_task *task, struct nfs_read_data *data)
357 struct nfs_readargs *argp = &data->args;
358 struct nfs_readres *resp = &data->res;
360 if (resp->eof || resp->count == argp->count)
363 /* This is a short read! */
364 nfs_inc_stats(data->inode, NFSIOS_SHORTREAD);
365 /* Has the server at least made some progress? */
366 if (resp->count == 0)
369 /* Yes, so retry the read at the end of the data */
370 argp->offset += resp->count;
371 argp->pgbase += resp->count;
372 argp->count -= resp->count;
373 rpc_restart_call(task);
377 * Handle a read reply that fills part of a page.
379 static void nfs_readpage_result_partial(struct rpc_task *task, void *calldata)
381 struct nfs_read_data *data = calldata;
383 if (nfs_readpage_result(task, data) != 0)
385 if (task->tk_status < 0)
388 nfs_readpage_truncate_uninitialised_page(data);
389 nfs_readpage_retry(task, data);
392 static void nfs_readpage_release_partial(void *calldata)
394 struct nfs_read_data *data = calldata;
395 struct nfs_page *req = data->req;
396 struct page *page = req->wb_page;
397 int status = data->task.tk_status;
402 if (atomic_dec_and_test(&req->wb_complete)) {
403 if (!PageError(page))
404 SetPageUptodate(page);
405 nfs_readpage_release(req);
407 nfs_readdata_release(calldata);
410 static const struct rpc_call_ops nfs_read_partial_ops = {
411 .rpc_call_done = nfs_readpage_result_partial,
412 .rpc_release = nfs_readpage_release_partial,
415 static void nfs_readpage_set_pages_uptodate(struct nfs_read_data *data)
417 unsigned int count = data->res.count;
418 unsigned int base = data->args.pgbase;
422 count = data->args.count;
423 if (unlikely(count == 0))
425 pages = &data->args.pages[base >> PAGE_CACHE_SHIFT];
426 base &= ~PAGE_CACHE_MASK;
428 for (;count >= PAGE_CACHE_SIZE; count -= PAGE_CACHE_SIZE, pages++)
429 SetPageUptodate(*pages);
432 /* Was this a short read? */
433 if (data->res.eof || data->res.count == data->args.count)
434 SetPageUptodate(*pages);
438 * This is the callback from RPC telling us whether a reply was
439 * received or some error occurred (timeout or socket shutdown).
441 static void nfs_readpage_result_full(struct rpc_task *task, void *calldata)
443 struct nfs_read_data *data = calldata;
445 if (nfs_readpage_result(task, data) != 0)
447 if (task->tk_status < 0)
450 * Note: nfs_readpage_retry may change the values of
451 * data->args. In the multi-page case, we therefore need
452 * to ensure that we call nfs_readpage_set_pages_uptodate()
455 nfs_readpage_truncate_uninitialised_page(data);
456 nfs_readpage_set_pages_uptodate(data);
457 nfs_readpage_retry(task, data);
460 static void nfs_readpage_release_full(void *calldata)
462 struct nfs_read_data *data = calldata;
464 while (!list_empty(&data->pages)) {
465 struct nfs_page *req = nfs_list_entry(data->pages.next);
467 nfs_list_remove_request(req);
468 nfs_readpage_release(req);
470 nfs_readdata_release(calldata);
473 static const struct rpc_call_ops nfs_read_full_ops = {
474 .rpc_call_done = nfs_readpage_result_full,
475 .rpc_release = nfs_readpage_release_full,
479 * Read a page over NFS.
480 * We read the page synchronously in the following case:
481 * - The error flag is set for this page. This happens only when a
482 * previous async read operation failed.
484 int nfs_readpage(struct file *file, struct page *page)
486 struct nfs_open_context *ctx;
487 struct inode *inode = page->mapping->host;
490 dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
491 page, PAGE_CACHE_SIZE, page->index);
492 nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
493 nfs_add_stats(inode, NFSIOS_READPAGES, 1);
496 * Try to flush any pending writes to the file..
498 * NOTE! Because we own the page lock, there cannot
499 * be any new pending writes generated at this point
500 * for this page (other pages can be written to).
502 error = nfs_wb_page(inode, page);
505 if (PageUptodate(page))
509 if (NFS_STALE(inode))
514 ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
518 ctx = get_nfs_open_context(nfs_file_open_context(file));
520 if (!IS_SYNC(inode)) {
521 error = nfs_readpage_from_fscache(ctx, inode, page);
526 error = nfs_readpage_async(ctx, inode, page);
529 put_nfs_open_context(ctx);
536 struct nfs_readdesc {
537 struct nfs_pageio_descriptor *pgio;
538 struct nfs_open_context *ctx;
542 readpage_async_filler(void *data, struct page *page)
544 struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
545 struct inode *inode = page->mapping->host;
546 struct nfs_page *new;
550 len = nfs_page_length(page);
552 return nfs_return_empty_page(page);
554 new = nfs_create_request(desc->ctx, inode, page, 0, len);
558 if (len < PAGE_CACHE_SIZE)
559 zero_user_segment(page, len, PAGE_CACHE_SIZE);
560 if (!nfs_pageio_add_request(desc->pgio, new)) {
561 error = desc->pgio->pg_error;
566 error = PTR_ERR(new);
573 int nfs_readpages(struct file *filp, struct address_space *mapping,
574 struct list_head *pages, unsigned nr_pages)
576 struct nfs_pageio_descriptor pgio;
577 struct nfs_readdesc desc = {
580 struct inode *inode = mapping->host;
581 struct nfs_server *server = NFS_SERVER(inode);
582 size_t rsize = server->rsize;
583 unsigned long npages;
586 dprintk("NFS: nfs_readpages (%s/%Ld %d)\n",
588 (long long)NFS_FILEID(inode),
590 nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
592 if (NFS_STALE(inode))
596 desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
597 if (desc.ctx == NULL)
600 desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
602 /* attempt to read as many of the pages as possible from the cache
603 * - this returns -ENOBUFS immediately if the cookie is negative
605 ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
608 goto read_complete; /* all pages were read */
610 if (rsize < PAGE_CACHE_SIZE)
611 nfs_pageio_init(&pgio, inode, nfs_pagein_multi, rsize, 0);
613 nfs_pageio_init(&pgio, inode, nfs_pagein_one, rsize, 0);
615 ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
617 nfs_pageio_complete(&pgio);
618 npages = (pgio.pg_bytes_written + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
619 nfs_add_stats(inode, NFSIOS_READPAGES, npages);
621 put_nfs_open_context(desc.ctx);
626 int __init nfs_init_readpagecache(void)
628 nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
629 sizeof(struct nfs_read_data),
630 0, SLAB_HWCACHE_ALIGN,
632 if (nfs_rdata_cachep == NULL)
635 nfs_rdata_mempool = mempool_create_slab_pool(MIN_POOL_READ,
637 if (nfs_rdata_mempool == NULL)
643 void nfs_destroy_readpagecache(void)
645 mempool_destroy(nfs_rdata_mempool);
646 kmem_cache_destroy(nfs_rdata_cachep);