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
250	/* For non rpc-based layout drivers, retry-through-MDS */
251	if (!task->tk_ops) {
252		hdr->pnfs_error = -EAGAIN;
253		return;
254	}
255
256	/* Yes, so retry the read at the end of the hdr */
257	hdr->mds_offset += resp->count;
258	argp->offset += resp->count;
259	argp->pgbase += resp->count;
260	argp->count -= resp->count;
261	rpc_restart_call_prepare(task);
262}
263
264static void nfs_readpage_result(struct rpc_task *task,
265				struct nfs_pgio_header *hdr)
266{
267	if (hdr->res.eof) {
268		loff_t bound;
269
270		bound = hdr->args.offset + hdr->res.count;
271		spin_lock(&hdr->lock);
272		if (bound < hdr->io_start + hdr->good_bytes) {
273			set_bit(NFS_IOHDR_EOF, &hdr->flags);
274			clear_bit(NFS_IOHDR_ERROR, &hdr->flags);
275			hdr->good_bytes = bound - hdr->io_start;
276		}
277		spin_unlock(&hdr->lock);
278	} else if (hdr->res.count < hdr->args.count)
279		nfs_readpage_retry(task, hdr);
280}
281
282/*
283 * Read a page over NFS.
284 * We read the page synchronously in the following case:
285 *  -	The error flag is set for this page. This happens only when a
286 *	previous async read operation failed.
287 */
288int nfs_readpage(struct file *file, struct page *page)
289{
290	struct nfs_open_context *ctx;
291	struct inode *inode = page_file_mapping(page)->host;
292	int		error;
293
294	dprintk("NFS: nfs_readpage (%p %ld@%lu)\n",
295		page, PAGE_CACHE_SIZE, page_file_index(page));
296	nfs_inc_stats(inode, NFSIOS_VFSREADPAGE);
297	nfs_add_stats(inode, NFSIOS_READPAGES, 1);
298
299	/*
300	 * Try to flush any pending writes to the file..
301	 *
302	 * NOTE! Because we own the page lock, there cannot
303	 * be any new pending writes generated at this point
304	 * for this page (other pages can be written to).
305	 */
306	error = nfs_wb_page(inode, page);
307	if (error)
308		goto out_unlock;
309	if (PageUptodate(page))
310		goto out_unlock;
311
312	error = -ESTALE;
313	if (NFS_STALE(inode))
314		goto out_unlock;
315
316	if (file == NULL) {
317		error = -EBADF;
318		ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
319		if (ctx == NULL)
320			goto out_unlock;
321	} else
322		ctx = get_nfs_open_context(nfs_file_open_context(file));
323
324	if (!IS_SYNC(inode)) {
325		error = nfs_readpage_from_fscache(ctx, inode, page);
326		if (error == 0)
327			goto out;
328	}
329
330	error = nfs_readpage_async(ctx, inode, page);
331
332out:
333	put_nfs_open_context(ctx);
334	return error;
335out_unlock:
336	unlock_page(page);
337	return error;
338}
339
340struct nfs_readdesc {
341	struct nfs_pageio_descriptor *pgio;
342	struct nfs_open_context *ctx;
343};
344
345static int
346readpage_async_filler(void *data, struct page *page)
347{
348	struct nfs_readdesc *desc = (struct nfs_readdesc *)data;
349	struct nfs_page *new;
350	unsigned int len;
351	int error;
352
353	len = nfs_page_length(page);
354	if (len == 0)
355		return nfs_return_empty_page(page);
356
357	new = nfs_create_request(desc->ctx, page, NULL, 0, len);
358	if (IS_ERR(new))
359		goto out_error;
360
361	if (len < PAGE_CACHE_SIZE)
362		zero_user_segment(page, len, PAGE_CACHE_SIZE);
363	if (!nfs_pageio_add_request(desc->pgio, new)) {
364		error = desc->pgio->pg_error;
365		goto out_unlock;
366	}
367	return 0;
368out_error:
369	error = PTR_ERR(new);
370out_unlock:
371	unlock_page(page);
372	return error;
373}
374
375int nfs_readpages(struct file *filp, struct address_space *mapping,
376		struct list_head *pages, unsigned nr_pages)
377{
378	struct nfs_pageio_descriptor pgio;
379	struct nfs_pgio_mirror *pgm;
380	struct nfs_readdesc desc = {
381		.pgio = &pgio,
382	};
383	struct inode *inode = mapping->host;
384	unsigned long npages;
385	int ret = -ESTALE;
386
387	dprintk("NFS: nfs_readpages (%s/%Lu %d)\n",
388			inode->i_sb->s_id,
389			(unsigned long long)NFS_FILEID(inode),
390			nr_pages);
391	nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
392
393	if (NFS_STALE(inode))
394		goto out;
395
396	if (filp == NULL) {
397		desc.ctx = nfs_find_open_context(inode, NULL, FMODE_READ);
398		if (desc.ctx == NULL)
399			return -EBADF;
400	} else
401		desc.ctx = get_nfs_open_context(nfs_file_open_context(filp));
402
403	/* attempt to read as many of the pages as possible from the cache
404	 * - this returns -ENOBUFS immediately if the cookie is negative
405	 */
406	ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
407					 pages, &nr_pages);
408	if (ret == 0)
409		goto read_complete; /* all pages were read */
410
411	nfs_pageio_init_read(&pgio, inode, false,
412			     &nfs_async_read_completion_ops);
413
414	ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
415	nfs_pageio_complete(&pgio);
416
417	/* It doesn't make sense to do mirrored reads! */
418	WARN_ON_ONCE(pgio.pg_mirror_count != 1);
419
420	pgm = &pgio.pg_mirrors[0];
421	NFS_I(inode)->read_io += pgm->pg_bytes_written;
422	npages = (pgm->pg_bytes_written + PAGE_CACHE_SIZE - 1) >>
423		 PAGE_CACHE_SHIFT;
424	nfs_add_stats(inode, NFSIOS_READPAGES, npages);
425read_complete:
426	put_nfs_open_context(desc.ctx);
427out:
428	return ret;
429}
430
431int __init nfs_init_readpagecache(void)
432{
433	nfs_rdata_cachep = kmem_cache_create("nfs_read_data",
434					     sizeof(struct nfs_pgio_header),
435					     0, SLAB_HWCACHE_ALIGN,
436					     NULL);
437	if (nfs_rdata_cachep == NULL)
438		return -ENOMEM;
439
440	return 0;
441}
442
443void nfs_destroy_readpagecache(void)
444{
445	kmem_cache_destroy(nfs_rdata_cachep);
446}
447
448static const struct nfs_rw_ops nfs_rw_read_ops = {
449	.rw_mode		= FMODE_READ,
450	.rw_alloc_header	= nfs_readhdr_alloc,
451	.rw_free_header		= nfs_readhdr_free,
452	.rw_done		= nfs_readpage_done,
453	.rw_result		= nfs_readpage_result,
454	.rw_initiate		= nfs_initiate_read,
455};
456