xref: /openbmc/linux/fs/nfs/pagelist.c (revision fd589a8f)
1 /*
2  * linux/fs/nfs/pagelist.c
3  *
4  * A set of helper functions for managing NFS read and write requests.
5  * The main purpose of these routines is to provide support for the
6  * coalescing of several requests into a single RPC call.
7  *
8  * Copyright 2000, 2001 (c) Trond Myklebust <trond.myklebust@fys.uio.no>
9  *
10  */
11 
12 #include <linux/slab.h>
13 #include <linux/file.h>
14 #include <linux/sched.h>
15 #include <linux/sunrpc/clnt.h>
16 #include <linux/nfs3.h>
17 #include <linux/nfs4.h>
18 #include <linux/nfs_page.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_mount.h>
21 
22 #include "internal.h"
23 
24 static struct kmem_cache *nfs_page_cachep;
25 
26 static inline struct nfs_page *
27 nfs_page_alloc(void)
28 {
29 	struct nfs_page	*p;
30 	p = kmem_cache_alloc(nfs_page_cachep, GFP_KERNEL);
31 	if (p) {
32 		memset(p, 0, sizeof(*p));
33 		INIT_LIST_HEAD(&p->wb_list);
34 	}
35 	return p;
36 }
37 
38 static inline void
39 nfs_page_free(struct nfs_page *p)
40 {
41 	kmem_cache_free(nfs_page_cachep, p);
42 }
43 
44 /**
45  * nfs_create_request - Create an NFS read/write request.
46  * @file: file descriptor to use
47  * @inode: inode to which the request is attached
48  * @page: page to write
49  * @offset: starting offset within the page for the write
50  * @count: number of bytes to read/write
51  *
52  * The page must be locked by the caller. This makes sure we never
53  * create two different requests for the same page.
54  * User should ensure it is safe to sleep in this function.
55  */
56 struct nfs_page *
57 nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
58 		   struct page *page,
59 		   unsigned int offset, unsigned int count)
60 {
61 	struct nfs_page		*req;
62 
63 	for (;;) {
64 		/* try to allocate the request struct */
65 		req = nfs_page_alloc();
66 		if (req != NULL)
67 			break;
68 
69 		if (fatal_signal_pending(current))
70 			return ERR_PTR(-ERESTARTSYS);
71 		yield();
72 	}
73 
74 	/* Initialize the request struct. Initially, we assume a
75 	 * long write-back delay. This will be adjusted in
76 	 * update_nfs_request below if the region is not locked. */
77 	req->wb_page    = page;
78 	atomic_set(&req->wb_complete, 0);
79 	req->wb_index	= page->index;
80 	page_cache_get(page);
81 	BUG_ON(PagePrivate(page));
82 	BUG_ON(!PageLocked(page));
83 	BUG_ON(page->mapping->host != inode);
84 	req->wb_offset  = offset;
85 	req->wb_pgbase	= offset;
86 	req->wb_bytes   = count;
87 	req->wb_context = get_nfs_open_context(ctx);
88 	kref_init(&req->wb_kref);
89 	return req;
90 }
91 
92 /**
93  * nfs_unlock_request - Unlock request and wake up sleepers.
94  * @req:
95  */
96 void nfs_unlock_request(struct nfs_page *req)
97 {
98 	if (!NFS_WBACK_BUSY(req)) {
99 		printk(KERN_ERR "NFS: Invalid unlock attempted\n");
100 		BUG();
101 	}
102 	smp_mb__before_clear_bit();
103 	clear_bit(PG_BUSY, &req->wb_flags);
104 	smp_mb__after_clear_bit();
105 	wake_up_bit(&req->wb_flags, PG_BUSY);
106 	nfs_release_request(req);
107 }
108 
109 /**
110  * nfs_set_page_tag_locked - Tag a request as locked
111  * @req:
112  */
113 int nfs_set_page_tag_locked(struct nfs_page *req)
114 {
115 	struct nfs_inode *nfsi = NFS_I(req->wb_context->path.dentry->d_inode);
116 
117 	if (!nfs_lock_request_dontget(req))
118 		return 0;
119 	if (req->wb_page != NULL)
120 		radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
121 	return 1;
122 }
123 
124 /**
125  * nfs_clear_page_tag_locked - Clear request tag and wake up sleepers
126  */
127 void nfs_clear_page_tag_locked(struct nfs_page *req)
128 {
129 	struct inode *inode = req->wb_context->path.dentry->d_inode;
130 	struct nfs_inode *nfsi = NFS_I(inode);
131 
132 	if (req->wb_page != NULL) {
133 		spin_lock(&inode->i_lock);
134 		radix_tree_tag_clear(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
135 		nfs_unlock_request(req);
136 		spin_unlock(&inode->i_lock);
137 	} else
138 		nfs_unlock_request(req);
139 }
140 
141 /**
142  * nfs_clear_request - Free up all resources allocated to the request
143  * @req:
144  *
145  * Release page resources associated with a write request after it
146  * has completed.
147  */
148 void nfs_clear_request(struct nfs_page *req)
149 {
150 	struct page *page = req->wb_page;
151 	if (page != NULL) {
152 		page_cache_release(page);
153 		req->wb_page = NULL;
154 	}
155 }
156 
157 
158 /**
159  * nfs_release_request - Release the count on an NFS read/write request
160  * @req: request to release
161  *
162  * Note: Should never be called with the spinlock held!
163  */
164 static void nfs_free_request(struct kref *kref)
165 {
166 	struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
167 
168 	/* Release struct file or cached credential */
169 	nfs_clear_request(req);
170 	put_nfs_open_context(req->wb_context);
171 	nfs_page_free(req);
172 }
173 
174 void nfs_release_request(struct nfs_page *req)
175 {
176 	kref_put(&req->wb_kref, nfs_free_request);
177 }
178 
179 /**
180  * nfs_wait_on_request - Wait for a request to complete.
181  * @req: request to wait upon.
182  *
183  * Interruptible by fatal signals only.
184  * The user is responsible for holding a count on the request.
185  */
186 int
187 nfs_wait_on_request(struct nfs_page *req)
188 {
189 	int ret = 0;
190 
191 	if (!test_bit(PG_BUSY, &req->wb_flags))
192 		goto out;
193 	ret = out_of_line_wait_on_bit(&req->wb_flags, PG_BUSY,
194 			nfs_wait_bit_killable, TASK_KILLABLE);
195 out:
196 	return ret;
197 }
198 
199 /**
200  * nfs_pageio_init - initialise a page io descriptor
201  * @desc: pointer to descriptor
202  * @inode: pointer to inode
203  * @doio: pointer to io function
204  * @bsize: io block size
205  * @io_flags: extra parameters for the io function
206  */
207 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
208 		     struct inode *inode,
209 		     int (*doio)(struct inode *, struct list_head *, unsigned int, size_t, int),
210 		     size_t bsize,
211 		     int io_flags)
212 {
213 	INIT_LIST_HEAD(&desc->pg_list);
214 	desc->pg_bytes_written = 0;
215 	desc->pg_count = 0;
216 	desc->pg_bsize = bsize;
217 	desc->pg_base = 0;
218 	desc->pg_inode = inode;
219 	desc->pg_doio = doio;
220 	desc->pg_ioflags = io_flags;
221 	desc->pg_error = 0;
222 }
223 
224 /**
225  * nfs_can_coalesce_requests - test two requests for compatibility
226  * @prev: pointer to nfs_page
227  * @req: pointer to nfs_page
228  *
229  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
230  * page data area they describe is contiguous, and that their RPC
231  * credentials, NFSv4 open state, and lockowners are the same.
232  *
233  * Return 'true' if this is the case, else return 'false'.
234  */
235 static int nfs_can_coalesce_requests(struct nfs_page *prev,
236 				     struct nfs_page *req)
237 {
238 	if (req->wb_context->cred != prev->wb_context->cred)
239 		return 0;
240 	if (req->wb_context->lockowner != prev->wb_context->lockowner)
241 		return 0;
242 	if (req->wb_context->state != prev->wb_context->state)
243 		return 0;
244 	if (req->wb_index != (prev->wb_index + 1))
245 		return 0;
246 	if (req->wb_pgbase != 0)
247 		return 0;
248 	if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
249 		return 0;
250 	return 1;
251 }
252 
253 /**
254  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
255  * @desc: destination io descriptor
256  * @req: request
257  *
258  * Returns true if the request 'req' was successfully coalesced into the
259  * existing list of pages 'desc'.
260  */
261 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
262 				     struct nfs_page *req)
263 {
264 	size_t newlen = req->wb_bytes;
265 
266 	if (desc->pg_count != 0) {
267 		struct nfs_page *prev;
268 
269 		/*
270 		 * FIXME: ideally we should be able to coalesce all requests
271 		 * that are not block boundary aligned, but currently this
272 		 * is problematic for the case of bsize < PAGE_CACHE_SIZE,
273 		 * since nfs_flush_multi and nfs_pagein_multi assume you
274 		 * can have only one struct nfs_page.
275 		 */
276 		if (desc->pg_bsize < PAGE_SIZE)
277 			return 0;
278 		newlen += desc->pg_count;
279 		if (newlen > desc->pg_bsize)
280 			return 0;
281 		prev = nfs_list_entry(desc->pg_list.prev);
282 		if (!nfs_can_coalesce_requests(prev, req))
283 			return 0;
284 	} else
285 		desc->pg_base = req->wb_pgbase;
286 	nfs_list_remove_request(req);
287 	nfs_list_add_request(req, &desc->pg_list);
288 	desc->pg_count = newlen;
289 	return 1;
290 }
291 
292 /*
293  * Helper for nfs_pageio_add_request and nfs_pageio_complete
294  */
295 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
296 {
297 	if (!list_empty(&desc->pg_list)) {
298 		int error = desc->pg_doio(desc->pg_inode,
299 					  &desc->pg_list,
300 					  nfs_page_array_len(desc->pg_base,
301 							     desc->pg_count),
302 					  desc->pg_count,
303 					  desc->pg_ioflags);
304 		if (error < 0)
305 			desc->pg_error = error;
306 		else
307 			desc->pg_bytes_written += desc->pg_count;
308 	}
309 	if (list_empty(&desc->pg_list)) {
310 		desc->pg_count = 0;
311 		desc->pg_base = 0;
312 	}
313 }
314 
315 /**
316  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
317  * @desc: destination io descriptor
318  * @req: request
319  *
320  * Returns true if the request 'req' was successfully coalesced into the
321  * existing list of pages 'desc'.
322  */
323 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
324 			   struct nfs_page *req)
325 {
326 	while (!nfs_pageio_do_add_request(desc, req)) {
327 		nfs_pageio_doio(desc);
328 		if (desc->pg_error < 0)
329 			return 0;
330 	}
331 	return 1;
332 }
333 
334 /**
335  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
336  * @desc: pointer to io descriptor
337  */
338 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
339 {
340 	nfs_pageio_doio(desc);
341 }
342 
343 /**
344  * nfs_pageio_cond_complete - Conditional I/O completion
345  * @desc: pointer to io descriptor
346  * @index: page index
347  *
348  * It is important to ensure that processes don't try to take locks
349  * on non-contiguous ranges of pages as that might deadlock. This
350  * function should be called before attempting to wait on a locked
351  * nfs_page. It will complete the I/O if the page index 'index'
352  * is not contiguous with the existing list of pages in 'desc'.
353  */
354 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
355 {
356 	if (!list_empty(&desc->pg_list)) {
357 		struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
358 		if (index != prev->wb_index + 1)
359 			nfs_pageio_doio(desc);
360 	}
361 }
362 
363 #define NFS_SCAN_MAXENTRIES 16
364 /**
365  * nfs_scan_list - Scan a list for matching requests
366  * @nfsi: NFS inode
367  * @dst: Destination list
368  * @idx_start: lower bound of page->index to scan
369  * @npages: idx_start + npages sets the upper bound to scan.
370  * @tag: tag to scan for
371  *
372  * Moves elements from one of the inode request lists.
373  * If the number of requests is set to 0, the entire address_space
374  * starting at index idx_start, is scanned.
375  * The requests are *not* checked to ensure that they form a contiguous set.
376  * You must be holding the inode's i_lock when calling this function
377  */
378 int nfs_scan_list(struct nfs_inode *nfsi,
379 		struct list_head *dst, pgoff_t idx_start,
380 		unsigned int npages, int tag)
381 {
382 	struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
383 	struct nfs_page *req;
384 	pgoff_t idx_end;
385 	int found, i;
386 	int res;
387 
388 	res = 0;
389 	if (npages == 0)
390 		idx_end = ~0;
391 	else
392 		idx_end = idx_start + npages - 1;
393 
394 	for (;;) {
395 		found = radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree,
396 				(void **)&pgvec[0], idx_start,
397 				NFS_SCAN_MAXENTRIES, tag);
398 		if (found <= 0)
399 			break;
400 		for (i = 0; i < found; i++) {
401 			req = pgvec[i];
402 			if (req->wb_index > idx_end)
403 				goto out;
404 			idx_start = req->wb_index + 1;
405 			if (nfs_set_page_tag_locked(req)) {
406 				kref_get(&req->wb_kref);
407 				nfs_list_remove_request(req);
408 				radix_tree_tag_clear(&nfsi->nfs_page_tree,
409 						req->wb_index, tag);
410 				nfs_list_add_request(req, dst);
411 				res++;
412 				if (res == INT_MAX)
413 					goto out;
414 			}
415 		}
416 		/* for latency reduction */
417 		cond_resched_lock(&nfsi->vfs_inode.i_lock);
418 	}
419 out:
420 	return res;
421 }
422 
423 int __init nfs_init_nfspagecache(void)
424 {
425 	nfs_page_cachep = kmem_cache_create("nfs_page",
426 					    sizeof(struct nfs_page),
427 					    0, SLAB_HWCACHE_ALIGN,
428 					    NULL);
429 	if (nfs_page_cachep == NULL)
430 		return -ENOMEM;
431 
432 	return 0;
433 }
434 
435 void nfs_destroy_nfspagecache(void)
436 {
437 	kmem_cache_destroy(nfs_page_cachep);
438 }
439 
440