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/module.h>
22
23#include "nfs4_fs.h"
24#include "internal.h"
25#include "iostat.h"
26#include "fscache.h"
27#include "pnfs.h"
28
29#define NFSDBG_FACILITY		NFSDBG_PAGECACHE
30
31static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops;
32static const struct nfs_rw_ops nfs_rw_read_ops;
33
34static struct kmem_cache *nfs_rdata_cachep;
35
36static struct nfs_pgio_header *nfs_readhdr_alloc(void)
37{
38	return kmem_cache_zalloc(nfs_rdata_cachep, GFP_KERNEL);
39}
40
41static void nfs_readhdr_free(struct nfs_pgio_header *rhdr)
42{
43	kmem_cache_free(nfs_rdata_cachep, rhdr);
44}
45
46static
47int nfs_return_empty_page(struct page *page)
48{
49	zero_user(page, 0, PAGE_CACHE_SIZE);
50	SetPageUptodate(page);
51	unlock_page(page);
52	return 0;
53}
54
55void nfs_pageio_init_read(struct nfs_pageio_descriptor *pgio,
56			      struct inode *inode, bool force_mds,
57			      const struct nfs_pgio_completion_ops *compl_ops)
58{
59	struct nfs_server *server = NFS_SERVER(inode);
60	const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
61
62#ifdef CONFIG_NFS_V4_1
63	if (server->pnfs_curr_ld && !force_mds)
64		pg_ops = server->pnfs_curr_ld->pg_read_ops;
65#endif
66	nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_read_ops,
67			server->rsize, 0);
68}
69EXPORT_SYMBOL_GPL(nfs_pageio_init_read);
70
71void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio)
72{
73	struct nfs_pgio_mirror *mirror;
74
75	if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
76		pgio->pg_ops->pg_cleanup(pgio);
77
78	pgio->pg_ops = &nfs_pgio_rw_ops;
79
80	/* read path should never have more than one mirror */
81	WARN_ON_ONCE(pgio->pg_mirror_count != 1);
82
83	mirror = &pgio->pg_mirrors[0];
84	mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->rsize;
85}
86EXPORT_SYMBOL_GPL(nfs_pageio_reset_read_mds);
87
88int nfs_readpage_async(struct nfs_open_context *ctx, struct inode *inode,
89		       struct page *page)
90{
91	struct nfs_page	*new;
92	unsigned int len;
93	struct nfs_pageio_descriptor pgio;
94	struct nfs_pgio_mirror *pgm;
95
96	len = nfs_page_length(page);
97	if (len == 0)
98		return nfs_return_empty_page(page);
99	new = nfs_create_request(ctx, page, NULL, 0, len);
100	if (IS_ERR(new)) {
101		unlock_page(page);
102		return PTR_ERR(new);
103	}
104	if (len < PAGE_CACHE_SIZE)
105		zero_user_segment(page, len, PAGE_CACHE_SIZE);
106
107	nfs_pageio_init_read(&pgio, inode, false,
108			     &nfs_async_read_completion_ops);
109	nfs_pageio_add_request(&pgio, new);
110	nfs_pageio_complete(&pgio);
111
112	/* It doesn't make sense to do mirrored reads! */
113	WARN_ON_ONCE(pgio.pg_mirror_count != 1);
114
115	pgm = &pgio.pg_mirrors[0];
116	NFS_I(inode)->read_io += pgm->pg_bytes_written;
117
118	return 0;
119}
120
121static void nfs_readpage_release(struct nfs_page *req)
122{
123	struct inode *inode = d_inode(req->wb_context->dentry);
124
125	dprintk("NFS: read done (%s/%llu %d@%lld)\n", inode->i_sb->s_id,
126		(unsigned long long)NFS_FILEID(inode), req->wb_bytes,
127		(long long)req_offset(req));
128
129	if (nfs_page_group_sync_on_bit(req, PG_UNLOCKPAGE)) {
130		if (PageUptodate(req->wb_page))
131			nfs_readpage_to_fscache(inode, req->wb_page, 0);
132
133		unlock_page(req->wb_page);
134	}
135	nfs_release_request(req);
136}
137
138static void nfs_page_group_set_uptodate(struct nfs_page *req)
139{
140	if (nfs_page_group_sync_on_bit(req, PG_UPTODATE))
141		SetPageUptodate(req->wb_page);
142}
143
144static void nfs_read_completion(struct nfs_pgio_header *hdr)
145{
146	unsigned long bytes = 0;
147
148	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
149		goto out;
150	while (!list_empty(&hdr->pages)) {
151		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
152		struct page *page = req->wb_page;
153		unsigned long start = req->wb_pgbase;
154		unsigned long end = req->wb_pgbase + req->wb_bytes;
155
156		if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
157			/* note: regions of the page not covered by a
158			 * request are zeroed in nfs_readpage_async /
159			 * readpage_async_filler */
160			if (bytes > hdr->good_bytes) {
161				/* nothing in this request was good, so zero
162				 * the full extent of the request */
163				zero_user_segment(page, start, end);
164
165			} else if (hdr->good_bytes - bytes < req->wb_bytes) {
166				/* part of this request has good bytes, but
167				 * not all. zero the bad bytes */
168				start += hdr->good_bytes - bytes;
169				WARN_ON(start < req->wb_pgbase);
170				zero_user_segment(page, start, end);
171			}
172		}
173		bytes += req->wb_bytes;
174		if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
175			if (bytes <= hdr->good_bytes)
176				nfs_page_group_set_uptodate(req);
177		} else
178			nfs_page_group_set_uptodate(req);
179		nfs_list_remove_request(req);
180		nfs_readpage_release(req);
181	}
182out:
183	hdr->release(hdr);
184}
185
186static void nfs_initiate_read(struct nfs_pgio_header *hdr,
187			      struct rpc_message *msg,
188			      const struct nfs_rpc_ops *rpc_ops,
189			      struct rpc_task_setup *task_setup_data, int how)
190{
191	struct inode *inode = hdr->inode;
192	int swap_flags = IS_SWAPFILE(inode) ? NFS_RPC_SWAPFLAGS : 0;
193
194	task_setup_data->flags |= swap_flags;
195	rpc_ops->read_setup(hdr, msg);
196}
197
198static void
199nfs_async_read_error(struct list_head *head)
200{
201	struct nfs_page	*req;
202
203	while (!list_empty(head)) {
204		req = nfs_list_entry(head->next);
205		nfs_list_remove_request(req);
206		nfs_readpage_release(req);
207	}
208}
209
210static const struct nfs_pgio_completion_ops nfs_async_read_completion_ops = {
211	.error_cleanup = nfs_async_read_error,
212	.completion = nfs_read_completion,
213};
214
215/*
216 * This is the callback from RPC telling us whether a reply was
217 * received or some error occurred (timeout or socket shutdown).
218 */
219static int nfs_readpage_done(struct rpc_task *task,
220			     struct nfs_pgio_header *hdr,
221			     struct inode *inode)
222{
223	int status = NFS_PROTO(inode)->read_done(task, hdr);
224	if (status != 0)
225		return status;
226
227	nfs_add_stats(inode, NFSIOS_SERVERREADBYTES, hdr->res.count);
228
229	if (task->tk_status == -ESTALE) {
230		set_bit(NFS_INO_STALE, &NFS_I(inode)->flags);
231		nfs_mark_for_revalidate(inode);
232	}
233	return 0;
234}
235
236static void nfs_readpage_retry(struct rpc_task *task,
237			       struct nfs_pgio_header *hdr)
238{
239	struct nfs_pgio_args *argp = &hdr->args;
240	struct nfs_pgio_res  *resp = &hdr->res;
241
242	/* This is a short read! */
243	nfs_inc_stats(hdr->inode, NFSIOS_SHORTREAD);
244	/* Has the server at least made some progress? */
245	if (resp->count == 0) {
246		nfs_set_pgio_error(hdr, -EIO, argp->offset);
247		return;
248	}
249	/* Yes, so retry the read at the end of the hdr */
250	hdr->mds_offset += resp->count;
251	argp->offset += resp->count;
252	argp->pgbase += resp->count;
253	argp->count -= resp->count;
254	rpc_restart_call_prepare(task);
255}
256
257static void nfs_readpage_result(struct rpc_task *task,
258				struct nfs_pgio_header *hdr)
259{
260	if (hdr->res.eof) {
261		loff_t bound;
262
263		bound = hdr->args.offset + hdr->res.count;
264		spin_lock(&hdr->lock);
265		if (bound < hdr->io_start + hdr->good_bytes) {
266			set_bit(NFS_IOHDR_EOF, &hdr->flags);
267			clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
268			hdr->good_bytes = bound - hdr->io_start;
269		}
270		spin_unlock(&hdr->lock);
271	} else if (hdr->res.count != hdr->args.count)
272		nfs_readpage_retry(task, hdr);
273}
274
275/*
276 * Read a page over NFS.
277 * We read the page synchronously in the following case:
278 *  -	The error flag is set for this page. This happens only when a
279 *	previous async read operation failed.
280 */
281int nfs_readpage(struct file *file, struct page *page)
282{
283	struct nfs_open_context *ctx;
284	struct inode *inode = page_file_mapping(page)->host;
285	int		error;
286
287	dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
288		page, PAGE_CACHE_SIZE, page_file_index(page));
289	nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
290	nfs_add_stats(inode, NFSIOS_READPAGES, 1);
291
292	/*
293	 * Try to flush any pending writes to the file..
294	 *
295	 * NOTE! Because we own the page lock, there cannot
296	 * be any new pending writes generated at this point
297	 * for this page (other pages can be written to).
298	 */
299	error = nfs_wb_page(inode, page);
300	if (error)
301		goto out_unlock;
302	if (PageUptodate(page))
303		goto out_unlock;
304
305	error = -ESTALE;
306	if (NFS_STALE(inode))
307		goto out_unlock;
308
309	if (file == NULL) {
310		error = -EBADF;
311		ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
312		if (ctx == NULL)
313			goto out_unlock;
314	} else
315		ctx = get_nfs_open_context(nfs_file_open_context(file));
316
317	if (!IS_SYNC(inode)) {
318		error = nfs_readpage_from_fscache(ctx, inode, page);
319		if (error == 0)
320			goto out;
321	}
322
323	error = nfs_readpage_async(ctx, inode, page);
324
325out:
326	put_nfs_open_context(ctx);
327	return error;
328out_unlock:
329	unlock_page(page);
330	return error;
331}
332
333struct nfs_readdesc {
334	struct nfs_pageio_descriptor *pgio;
335	struct nfs_open_context *ctx;
336};
337
338static int
339readpage_async_filler(void *data, struct page *page)
340{
341	struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
342	struct nfs_page *new;
343	unsigned int len;
344	int error;
345
346	len = nfs_page_length(page);
347	if (len == 0)
348		return nfs_return_empty_page(page);
349
350	new = nfs_create_request(desc->ctx, page, NULL, 0, len);
351	if (IS_ERR(new))
352		goto out_error;
353
354	if (len < PAGE_CACHE_SIZE)
355		zero_user_segment(page, len, PAGE_CACHE_SIZE);
356	if (!nfs_pageio_add_request(desc->pgio, new)) {
357		error = desc->pgio->pg_error;
358		goto out_unlock;
359	}
360	return 0;
361out_error:
362	error = PTR_ERR(new);
363out_unlock:
364	unlock_page(page);
365	return error;
366}
367
368int nfs_readpages(struct file *filp, struct address_space *mapping,
369		struct list_head *pages, unsigned nr_pages)
370{
371	struct nfs_pageio_descriptor pgio;
372	struct nfs_pgio_mirror *pgm;
373	struct nfs_readdesc desc = {
374		.pgio = &pgio,
375	};
376	struct inode *inode = mapping->host;
377	unsigned long npages;
378	int ret = -ESTALE;
379
380	dprintk("NFS: nfs_readpages (%s/%Lu %d)\n",
381			inode->i_sb->s_id,
382			(unsigned long long)NFS_FILEID(inode),
383			nr_pages);
384	nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
385
386	if (NFS_STALE(inode))
387		goto out;
388
389	if (filp == NULL) {
390		desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
391		if (desc.ctx == NULL)
392			return -EBADF;
393	} else
394		desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
395
396	/* attempt to read as many of the pages as possible from the cache
397	 * - this returns -ENOBUFS immediately if the cookie is negative
398	 */
399	ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
400					 pages, &nr_pages);
401	if (ret == 0)
402		goto read_complete; /* all pages were read */
403
404	nfs_pageio_init_read(&pgio, inode, false,
405			     &nfs_async_read_completion_ops);
406
407	ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
408	nfs_pageio_complete(&pgio);
409
410	/* It doesn't make sense to do mirrored reads! */
411	WARN_ON_ONCE(pgio.pg_mirror_count != 1);
412
413	pgm = &pgio.pg_mirrors[0];
414	NFS_I(inode)->read_io += pgm->pg_bytes_written;
415	npages = (pgm->pg_bytes_written + PAGE_CACHE_SIZE - 1) >>
416		 PAGE_CACHE_SHIFT;
417	nfs_add_stats(inode, NFSIOS_READPAGES, npages);
418read_complete:
419	put_nfs_open_context(desc.ctx);
420out:
421	return ret;
422}
423
424int __init nfs_init_readpagecache(void)
425{
426	nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
427					     sizeof(struct nfs_pgio_header),
428					     0, SLAB_HWCACHE_ALIGN,
429					     NULL);
430	if (nfs_rdata_cachep == NULL)
431		return -ENOMEM;
432
433	return 0;
434}
435
436void nfs_destroy_readpagecache(void)
437{
438	kmem_cache_destroy(nfs_rdata_cachep);
439}
440
441static const struct nfs_rw_ops nfs_rw_read_ops = {
442	.rw_mode		= FMODE_READ,
443	.rw_alloc_header	= nfs_readhdr_alloc,
444	.rw_free_header		= nfs_readhdr_free,
445	.rw_done		= nfs_readpage_done,
446	.rw_result		= nfs_readpage_result,
447	.rw_initiate		= nfs_initiate_read,
448};
449