Loading...
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#include "pnfs.h"
24
25static struct kmem_cache *nfs_page_cachep;
26
27static inline struct nfs_page *
28nfs_page_alloc(void)
29{
30 struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_KERNEL);
31 if (p)
32 INIT_LIST_HEAD(&p->wb_list);
33 return p;
34}
35
36static inline void
37nfs_page_free(struct nfs_page *p)
38{
39 kmem_cache_free(nfs_page_cachep, p);
40}
41
42/**
43 * nfs_create_request - Create an NFS read/write request.
44 * @file: file descriptor to use
45 * @inode: inode to which the request is attached
46 * @page: page to write
47 * @offset: starting offset within the page for the write
48 * @count: number of bytes to read/write
49 *
50 * The page must be locked by the caller. This makes sure we never
51 * create two different requests for the same page.
52 * User should ensure it is safe to sleep in this function.
53 */
54struct nfs_page *
55nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
56 struct page *page,
57 unsigned int offset, unsigned int count)
58{
59 struct nfs_page *req;
60
61 /* try to allocate the request struct */
62 req = nfs_page_alloc();
63 if (req == NULL)
64 return ERR_PTR(-ENOMEM);
65
66 /* get lock context early so we can deal with alloc failures */
67 req->wb_lock_context = nfs_get_lock_context(ctx);
68 if (req->wb_lock_context == NULL) {
69 nfs_page_free(req);
70 return ERR_PTR(-ENOMEM);
71 }
72
73 /* Initialize the request struct. Initially, we assume a
74 * long write-back delay. This will be adjusted in
75 * update_nfs_request below if the region is not locked. */
76 req->wb_page = page;
77 atomic_set(&req->wb_complete, 0);
78 req->wb_index = page->index;
79 page_cache_get(page);
80 BUG_ON(PagePrivate(page));
81 BUG_ON(!PageLocked(page));
82 BUG_ON(page->mapping->host != inode);
83 req->wb_offset = offset;
84 req->wb_pgbase = offset;
85 req->wb_bytes = count;
86 req->wb_context = get_nfs_open_context(ctx);
87 kref_init(&req->wb_kref);
88 return req;
89}
90
91/**
92 * nfs_unlock_request - Unlock request and wake up sleepers.
93 * @req:
94 */
95void nfs_unlock_request(struct nfs_page *req)
96{
97 if (!NFS_WBACK_BUSY(req)) {
98 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
99 BUG();
100 }
101 smp_mb__before_clear_bit();
102 clear_bit(PG_BUSY, &req->wb_flags);
103 smp_mb__after_clear_bit();
104 wake_up_bit(&req->wb_flags, PG_BUSY);
105 nfs_release_request(req);
106}
107
108/**
109 * nfs_set_page_tag_locked - Tag a request as locked
110 * @req:
111 */
112int nfs_set_page_tag_locked(struct nfs_page *req)
113{
114 if (!nfs_lock_request_dontget(req))
115 return 0;
116 if (test_bit(PG_MAPPED, &req->wb_flags))
117 radix_tree_tag_set(&NFS_I(req->wb_context->dentry->d_inode)->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
118 return 1;
119}
120
121/**
122 * nfs_clear_page_tag_locked - Clear request tag and wake up sleepers
123 */
124void nfs_clear_page_tag_locked(struct nfs_page *req)
125{
126 if (test_bit(PG_MAPPED, &req->wb_flags)) {
127 struct inode *inode = req->wb_context->dentry->d_inode;
128 struct nfs_inode *nfsi = NFS_I(inode);
129
130 spin_lock(&inode->i_lock);
131 radix_tree_tag_clear(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
132 nfs_unlock_request(req);
133 spin_unlock(&inode->i_lock);
134 } else
135 nfs_unlock_request(req);
136}
137
138/*
139 * nfs_clear_request - Free up all resources allocated to the request
140 * @req:
141 *
142 * Release page and open context resources associated with a read/write
143 * request after it has completed.
144 */
145static void nfs_clear_request(struct nfs_page *req)
146{
147 struct page *page = req->wb_page;
148 struct nfs_open_context *ctx = req->wb_context;
149 struct nfs_lock_context *l_ctx = req->wb_lock_context;
150
151 if (page != NULL) {
152 page_cache_release(page);
153 req->wb_page = NULL;
154 }
155 if (l_ctx != NULL) {
156 nfs_put_lock_context(l_ctx);
157 req->wb_lock_context = NULL;
158 }
159 if (ctx != NULL) {
160 put_nfs_open_context(ctx);
161 req->wb_context = NULL;
162 }
163}
164
165
166/**
167 * nfs_release_request - Release the count on an NFS read/write request
168 * @req: request to release
169 *
170 * Note: Should never be called with the spinlock held!
171 */
172static void nfs_free_request(struct kref *kref)
173{
174 struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
175
176 /* Release struct file and open context */
177 nfs_clear_request(req);
178 nfs_page_free(req);
179}
180
181void nfs_release_request(struct nfs_page *req)
182{
183 kref_put(&req->wb_kref, nfs_free_request);
184}
185
186static int nfs_wait_bit_uninterruptible(void *word)
187{
188 io_schedule();
189 return 0;
190}
191
192/**
193 * nfs_wait_on_request - Wait for a request to complete.
194 * @req: request to wait upon.
195 *
196 * Interruptible by fatal signals only.
197 * The user is responsible for holding a count on the request.
198 */
199int
200nfs_wait_on_request(struct nfs_page *req)
201{
202 return wait_on_bit(&req->wb_flags, PG_BUSY,
203 nfs_wait_bit_uninterruptible,
204 TASK_UNINTERRUPTIBLE);
205}
206
207bool nfs_generic_pg_test(struct nfs_pageio_descriptor *desc, struct nfs_page *prev, struct nfs_page *req)
208{
209 /*
210 * FIXME: ideally we should be able to coalesce all requests
211 * that are not block boundary aligned, but currently this
212 * is problematic for the case of bsize < PAGE_CACHE_SIZE,
213 * since nfs_flush_multi and nfs_pagein_multi assume you
214 * can have only one struct nfs_page.
215 */
216 if (desc->pg_bsize < PAGE_SIZE)
217 return 0;
218
219 return desc->pg_count + req->wb_bytes <= desc->pg_bsize;
220}
221EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
222
223/**
224 * nfs_pageio_init - initialise a page io descriptor
225 * @desc: pointer to descriptor
226 * @inode: pointer to inode
227 * @doio: pointer to io function
228 * @bsize: io block size
229 * @io_flags: extra parameters for the io function
230 */
231void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
232 struct inode *inode,
233 const struct nfs_pageio_ops *pg_ops,
234 size_t bsize,
235 int io_flags)
236{
237 INIT_LIST_HEAD(&desc->pg_list);
238 desc->pg_bytes_written = 0;
239 desc->pg_count = 0;
240 desc->pg_bsize = bsize;
241 desc->pg_base = 0;
242 desc->pg_moreio = 0;
243 desc->pg_recoalesce = 0;
244 desc->pg_inode = inode;
245 desc->pg_ops = pg_ops;
246 desc->pg_ioflags = io_flags;
247 desc->pg_error = 0;
248 desc->pg_lseg = NULL;
249}
250
251/**
252 * nfs_can_coalesce_requests - test two requests for compatibility
253 * @prev: pointer to nfs_page
254 * @req: pointer to nfs_page
255 *
256 * The nfs_page structures 'prev' and 'req' are compared to ensure that the
257 * page data area they describe is contiguous, and that their RPC
258 * credentials, NFSv4 open state, and lockowners are the same.
259 *
260 * Return 'true' if this is the case, else return 'false'.
261 */
262static bool nfs_can_coalesce_requests(struct nfs_page *prev,
263 struct nfs_page *req,
264 struct nfs_pageio_descriptor *pgio)
265{
266 if (req->wb_context->cred != prev->wb_context->cred)
267 return false;
268 if (req->wb_lock_context->lockowner != prev->wb_lock_context->lockowner)
269 return false;
270 if (req->wb_context->state != prev->wb_context->state)
271 return false;
272 if (req->wb_index != (prev->wb_index + 1))
273 return false;
274 if (req->wb_pgbase != 0)
275 return false;
276 if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
277 return false;
278 return pgio->pg_ops->pg_test(pgio, prev, req);
279}
280
281/**
282 * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
283 * @desc: destination io descriptor
284 * @req: request
285 *
286 * Returns true if the request 'req' was successfully coalesced into the
287 * existing list of pages 'desc'.
288 */
289static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
290 struct nfs_page *req)
291{
292 if (desc->pg_count != 0) {
293 struct nfs_page *prev;
294
295 prev = nfs_list_entry(desc->pg_list.prev);
296 if (!nfs_can_coalesce_requests(prev, req, desc))
297 return 0;
298 } else {
299 if (desc->pg_ops->pg_init)
300 desc->pg_ops->pg_init(desc, req);
301 desc->pg_base = req->wb_pgbase;
302 }
303 nfs_list_remove_request(req);
304 nfs_list_add_request(req, &desc->pg_list);
305 desc->pg_count += req->wb_bytes;
306 return 1;
307}
308
309/*
310 * Helper for nfs_pageio_add_request and nfs_pageio_complete
311 */
312static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
313{
314 if (!list_empty(&desc->pg_list)) {
315 int error = desc->pg_ops->pg_doio(desc);
316 if (error < 0)
317 desc->pg_error = error;
318 else
319 desc->pg_bytes_written += desc->pg_count;
320 }
321 if (list_empty(&desc->pg_list)) {
322 desc->pg_count = 0;
323 desc->pg_base = 0;
324 }
325}
326
327/**
328 * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
329 * @desc: destination io descriptor
330 * @req: request
331 *
332 * Returns true if the request 'req' was successfully coalesced into the
333 * existing list of pages 'desc'.
334 */
335static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
336 struct nfs_page *req)
337{
338 while (!nfs_pageio_do_add_request(desc, req)) {
339 desc->pg_moreio = 1;
340 nfs_pageio_doio(desc);
341 if (desc->pg_error < 0)
342 return 0;
343 desc->pg_moreio = 0;
344 if (desc->pg_recoalesce)
345 return 0;
346 }
347 return 1;
348}
349
350static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
351{
352 LIST_HEAD(head);
353
354 do {
355 list_splice_init(&desc->pg_list, &head);
356 desc->pg_bytes_written -= desc->pg_count;
357 desc->pg_count = 0;
358 desc->pg_base = 0;
359 desc->pg_recoalesce = 0;
360
361 while (!list_empty(&head)) {
362 struct nfs_page *req;
363
364 req = list_first_entry(&head, struct nfs_page, wb_list);
365 nfs_list_remove_request(req);
366 if (__nfs_pageio_add_request(desc, req))
367 continue;
368 if (desc->pg_error < 0)
369 return 0;
370 break;
371 }
372 } while (desc->pg_recoalesce);
373 return 1;
374}
375
376int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
377 struct nfs_page *req)
378{
379 int ret;
380
381 do {
382 ret = __nfs_pageio_add_request(desc, req);
383 if (ret)
384 break;
385 if (desc->pg_error < 0)
386 break;
387 ret = nfs_do_recoalesce(desc);
388 } while (ret);
389 return ret;
390}
391
392/**
393 * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
394 * @desc: pointer to io descriptor
395 */
396void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
397{
398 for (;;) {
399 nfs_pageio_doio(desc);
400 if (!desc->pg_recoalesce)
401 break;
402 if (!nfs_do_recoalesce(desc))
403 break;
404 }
405}
406
407/**
408 * nfs_pageio_cond_complete - Conditional I/O completion
409 * @desc: pointer to io descriptor
410 * @index: page index
411 *
412 * It is important to ensure that processes don't try to take locks
413 * on non-contiguous ranges of pages as that might deadlock. This
414 * function should be called before attempting to wait on a locked
415 * nfs_page. It will complete the I/O if the page index 'index'
416 * is not contiguous with the existing list of pages in 'desc'.
417 */
418void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
419{
420 if (!list_empty(&desc->pg_list)) {
421 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
422 if (index != prev->wb_index + 1)
423 nfs_pageio_complete(desc);
424 }
425}
426
427#define NFS_SCAN_MAXENTRIES 16
428/**
429 * nfs_scan_list - Scan a list for matching requests
430 * @nfsi: NFS inode
431 * @dst: Destination list
432 * @idx_start: lower bound of page->index to scan
433 * @npages: idx_start + npages sets the upper bound to scan.
434 * @tag: tag to scan for
435 *
436 * Moves elements from one of the inode request lists.
437 * If the number of requests is set to 0, the entire address_space
438 * starting at index idx_start, is scanned.
439 * The requests are *not* checked to ensure that they form a contiguous set.
440 * You must be holding the inode's i_lock when calling this function
441 */
442int nfs_scan_list(struct nfs_inode *nfsi,
443 struct list_head *dst, pgoff_t idx_start,
444 unsigned int npages, int tag)
445{
446 struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
447 struct nfs_page *req;
448 pgoff_t idx_end;
449 int found, i;
450 int res;
451 struct list_head *list;
452
453 res = 0;
454 if (npages == 0)
455 idx_end = ~0;
456 else
457 idx_end = idx_start + npages - 1;
458
459 for (;;) {
460 found = radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree,
461 (void **)&pgvec[0], idx_start,
462 NFS_SCAN_MAXENTRIES, tag);
463 if (found <= 0)
464 break;
465 for (i = 0; i < found; i++) {
466 req = pgvec[i];
467 if (req->wb_index > idx_end)
468 goto out;
469 idx_start = req->wb_index + 1;
470 if (nfs_set_page_tag_locked(req)) {
471 kref_get(&req->wb_kref);
472 radix_tree_tag_clear(&nfsi->nfs_page_tree,
473 req->wb_index, tag);
474 list = pnfs_choose_commit_list(req, dst);
475 nfs_list_add_request(req, list);
476 res++;
477 if (res == INT_MAX)
478 goto out;
479 }
480 }
481 /* for latency reduction */
482 cond_resched_lock(&nfsi->vfs_inode.i_lock);
483 }
484out:
485 return res;
486}
487
488int __init nfs_init_nfspagecache(void)
489{
490 nfs_page_cachep = kmem_cache_create("nfs_page",
491 sizeof(struct nfs_page),
492 0, SLAB_HWCACHE_ALIGN,
493 NULL);
494 if (nfs_page_cachep == NULL)
495 return -ENOMEM;
496
497 return 0;
498}
499
500void nfs_destroy_nfspagecache(void)
501{
502 kmem_cache_destroy(nfs_page_cachep);
503}
504
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/nfs.h>
17#include <linux/nfs3.h>
18#include <linux/nfs4.h>
19#include <linux/nfs_page.h>
20#include <linux/nfs_fs.h>
21#include <linux/nfs_mount.h>
22#include <linux/export.h>
23
24#include "internal.h"
25#include "pnfs.h"
26
27static struct kmem_cache *nfs_page_cachep;
28
29bool nfs_pgarray_set(struct nfs_page_array *p, unsigned int pagecount)
30{
31 p->npages = pagecount;
32 if (pagecount <= ARRAY_SIZE(p->page_array))
33 p->pagevec = p->page_array;
34 else {
35 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_KERNEL);
36 if (!p->pagevec)
37 p->npages = 0;
38 }
39 return p->pagevec != NULL;
40}
41
42void nfs_pgheader_init(struct nfs_pageio_descriptor *desc,
43 struct nfs_pgio_header *hdr,
44 void (*release)(struct nfs_pgio_header *hdr))
45{
46 hdr->req = nfs_list_entry(desc->pg_list.next);
47 hdr->inode = desc->pg_inode;
48 hdr->cred = hdr->req->wb_context->cred;
49 hdr->io_start = req_offset(hdr->req);
50 hdr->good_bytes = desc->pg_count;
51 hdr->dreq = desc->pg_dreq;
52 hdr->release = release;
53 hdr->completion_ops = desc->pg_completion_ops;
54 if (hdr->completion_ops->init_hdr)
55 hdr->completion_ops->init_hdr(hdr);
56}
57
58void nfs_set_pgio_error(struct nfs_pgio_header *hdr, int error, loff_t pos)
59{
60 spin_lock(&hdr->lock);
61 if (pos < hdr->io_start + hdr->good_bytes) {
62 set_bit(NFS_IOHDR_ERROR, &hdr->flags);
63 clear_bit(NFS_IOHDR_EOF, &hdr->flags);
64 hdr->good_bytes = pos - hdr->io_start;
65 hdr->error = error;
66 }
67 spin_unlock(&hdr->lock);
68}
69
70static inline struct nfs_page *
71nfs_page_alloc(void)
72{
73 struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_KERNEL);
74 if (p)
75 INIT_LIST_HEAD(&p->wb_list);
76 return p;
77}
78
79static inline void
80nfs_page_free(struct nfs_page *p)
81{
82 kmem_cache_free(nfs_page_cachep, p);
83}
84
85/**
86 * nfs_create_request - Create an NFS read/write request.
87 * @ctx: open context to use
88 * @inode: inode to which the request is attached
89 * @page: page to write
90 * @offset: starting offset within the page for the write
91 * @count: number of bytes to read/write
92 *
93 * The page must be locked by the caller. This makes sure we never
94 * create two different requests for the same page.
95 * User should ensure it is safe to sleep in this function.
96 */
97struct nfs_page *
98nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
99 struct page *page,
100 unsigned int offset, unsigned int count)
101{
102 struct nfs_page *req;
103
104 /* try to allocate the request struct */
105 req = nfs_page_alloc();
106 if (req == NULL)
107 return ERR_PTR(-ENOMEM);
108
109 /* get lock context early so we can deal with alloc failures */
110 req->wb_lock_context = nfs_get_lock_context(ctx);
111 if (req->wb_lock_context == NULL) {
112 nfs_page_free(req);
113 return ERR_PTR(-ENOMEM);
114 }
115
116 /* Initialize the request struct. Initially, we assume a
117 * long write-back delay. This will be adjusted in
118 * update_nfs_request below if the region is not locked. */
119 req->wb_page = page;
120 req->wb_index = page->index;
121 page_cache_get(page);
122 req->wb_offset = offset;
123 req->wb_pgbase = offset;
124 req->wb_bytes = count;
125 req->wb_context = get_nfs_open_context(ctx);
126 kref_init(&req->wb_kref);
127 return req;
128}
129
130/**
131 * nfs_unlock_request - Unlock request and wake up sleepers.
132 * @req:
133 */
134void nfs_unlock_request(struct nfs_page *req)
135{
136 if (!NFS_WBACK_BUSY(req)) {
137 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
138 BUG();
139 }
140 smp_mb__before_clear_bit();
141 clear_bit(PG_BUSY, &req->wb_flags);
142 smp_mb__after_clear_bit();
143 wake_up_bit(&req->wb_flags, PG_BUSY);
144}
145
146/**
147 * nfs_unlock_and_release_request - Unlock request and release the nfs_page
148 * @req:
149 */
150void nfs_unlock_and_release_request(struct nfs_page *req)
151{
152 nfs_unlock_request(req);
153 nfs_release_request(req);
154}
155
156/*
157 * nfs_clear_request - Free up all resources allocated to the request
158 * @req:
159 *
160 * Release page and open context resources associated with a read/write
161 * request after it has completed.
162 */
163static void nfs_clear_request(struct nfs_page *req)
164{
165 struct page *page = req->wb_page;
166 struct nfs_open_context *ctx = req->wb_context;
167 struct nfs_lock_context *l_ctx = req->wb_lock_context;
168
169 if (page != NULL) {
170 page_cache_release(page);
171 req->wb_page = NULL;
172 }
173 if (l_ctx != NULL) {
174 nfs_put_lock_context(l_ctx);
175 req->wb_lock_context = NULL;
176 }
177 if (ctx != NULL) {
178 put_nfs_open_context(ctx);
179 req->wb_context = NULL;
180 }
181}
182
183
184/**
185 * nfs_release_request - Release the count on an NFS read/write request
186 * @req: request to release
187 *
188 * Note: Should never be called with the spinlock held!
189 */
190static void nfs_free_request(struct kref *kref)
191{
192 struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
193
194 /* Release struct file and open context */
195 nfs_clear_request(req);
196 nfs_page_free(req);
197}
198
199void nfs_release_request(struct nfs_page *req)
200{
201 kref_put(&req->wb_kref, nfs_free_request);
202}
203
204static int nfs_wait_bit_uninterruptible(void *word)
205{
206 io_schedule();
207 return 0;
208}
209
210/**
211 * nfs_wait_on_request - Wait for a request to complete.
212 * @req: request to wait upon.
213 *
214 * Interruptible by fatal signals only.
215 * The user is responsible for holding a count on the request.
216 */
217int
218nfs_wait_on_request(struct nfs_page *req)
219{
220 return wait_on_bit(&req->wb_flags, PG_BUSY,
221 nfs_wait_bit_uninterruptible,
222 TASK_UNINTERRUPTIBLE);
223}
224
225bool nfs_generic_pg_test(struct nfs_pageio_descriptor *desc, struct nfs_page *prev, struct nfs_page *req)
226{
227 /*
228 * FIXME: ideally we should be able to coalesce all requests
229 * that are not block boundary aligned, but currently this
230 * is problematic for the case of bsize < PAGE_CACHE_SIZE,
231 * since nfs_flush_multi and nfs_pagein_multi assume you
232 * can have only one struct nfs_page.
233 */
234 if (desc->pg_bsize < PAGE_SIZE)
235 return 0;
236
237 return desc->pg_count + req->wb_bytes <= desc->pg_bsize;
238}
239EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
240
241/**
242 * nfs_pageio_init - initialise a page io descriptor
243 * @desc: pointer to descriptor
244 * @inode: pointer to inode
245 * @doio: pointer to io function
246 * @bsize: io block size
247 * @io_flags: extra parameters for the io function
248 */
249void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
250 struct inode *inode,
251 const struct nfs_pageio_ops *pg_ops,
252 const struct nfs_pgio_completion_ops *compl_ops,
253 size_t bsize,
254 int io_flags)
255{
256 INIT_LIST_HEAD(&desc->pg_list);
257 desc->pg_bytes_written = 0;
258 desc->pg_count = 0;
259 desc->pg_bsize = bsize;
260 desc->pg_base = 0;
261 desc->pg_moreio = 0;
262 desc->pg_recoalesce = 0;
263 desc->pg_inode = inode;
264 desc->pg_ops = pg_ops;
265 desc->pg_completion_ops = compl_ops;
266 desc->pg_ioflags = io_flags;
267 desc->pg_error = 0;
268 desc->pg_lseg = NULL;
269 desc->pg_dreq = NULL;
270}
271
272/**
273 * nfs_can_coalesce_requests - test two requests for compatibility
274 * @prev: pointer to nfs_page
275 * @req: pointer to nfs_page
276 *
277 * The nfs_page structures 'prev' and 'req' are compared to ensure that the
278 * page data area they describe is contiguous, and that their RPC
279 * credentials, NFSv4 open state, and lockowners are the same.
280 *
281 * Return 'true' if this is the case, else return 'false'.
282 */
283static bool nfs_can_coalesce_requests(struct nfs_page *prev,
284 struct nfs_page *req,
285 struct nfs_pageio_descriptor *pgio)
286{
287 if (req->wb_context->cred != prev->wb_context->cred)
288 return false;
289 if (req->wb_lock_context->lockowner != prev->wb_lock_context->lockowner)
290 return false;
291 if (req->wb_context->state != prev->wb_context->state)
292 return false;
293 if (req->wb_pgbase != 0)
294 return false;
295 if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
296 return false;
297 if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
298 return false;
299 return pgio->pg_ops->pg_test(pgio, prev, req);
300}
301
302/**
303 * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
304 * @desc: destination io descriptor
305 * @req: request
306 *
307 * Returns true if the request 'req' was successfully coalesced into the
308 * existing list of pages 'desc'.
309 */
310static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
311 struct nfs_page *req)
312{
313 if (desc->pg_count != 0) {
314 struct nfs_page *prev;
315
316 prev = nfs_list_entry(desc->pg_list.prev);
317 if (!nfs_can_coalesce_requests(prev, req, desc))
318 return 0;
319 } else {
320 if (desc->pg_ops->pg_init)
321 desc->pg_ops->pg_init(desc, req);
322 desc->pg_base = req->wb_pgbase;
323 }
324 nfs_list_remove_request(req);
325 nfs_list_add_request(req, &desc->pg_list);
326 desc->pg_count += req->wb_bytes;
327 return 1;
328}
329
330/*
331 * Helper for nfs_pageio_add_request and nfs_pageio_complete
332 */
333static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
334{
335 if (!list_empty(&desc->pg_list)) {
336 int error = desc->pg_ops->pg_doio(desc);
337 if (error < 0)
338 desc->pg_error = error;
339 else
340 desc->pg_bytes_written += desc->pg_count;
341 }
342 if (list_empty(&desc->pg_list)) {
343 desc->pg_count = 0;
344 desc->pg_base = 0;
345 }
346}
347
348/**
349 * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
350 * @desc: destination io descriptor
351 * @req: request
352 *
353 * Returns true if the request 'req' was successfully coalesced into the
354 * existing list of pages 'desc'.
355 */
356static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
357 struct nfs_page *req)
358{
359 while (!nfs_pageio_do_add_request(desc, req)) {
360 desc->pg_moreio = 1;
361 nfs_pageio_doio(desc);
362 if (desc->pg_error < 0)
363 return 0;
364 desc->pg_moreio = 0;
365 if (desc->pg_recoalesce)
366 return 0;
367 }
368 return 1;
369}
370
371static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
372{
373 LIST_HEAD(head);
374
375 do {
376 list_splice_init(&desc->pg_list, &head);
377 desc->pg_bytes_written -= desc->pg_count;
378 desc->pg_count = 0;
379 desc->pg_base = 0;
380 desc->pg_recoalesce = 0;
381
382 while (!list_empty(&head)) {
383 struct nfs_page *req;
384
385 req = list_first_entry(&head, struct nfs_page, wb_list);
386 nfs_list_remove_request(req);
387 if (__nfs_pageio_add_request(desc, req))
388 continue;
389 if (desc->pg_error < 0)
390 return 0;
391 break;
392 }
393 } while (desc->pg_recoalesce);
394 return 1;
395}
396
397int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
398 struct nfs_page *req)
399{
400 int ret;
401
402 do {
403 ret = __nfs_pageio_add_request(desc, req);
404 if (ret)
405 break;
406 if (desc->pg_error < 0)
407 break;
408 ret = nfs_do_recoalesce(desc);
409 } while (ret);
410 return ret;
411}
412
413/**
414 * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
415 * @desc: pointer to io descriptor
416 */
417void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
418{
419 for (;;) {
420 nfs_pageio_doio(desc);
421 if (!desc->pg_recoalesce)
422 break;
423 if (!nfs_do_recoalesce(desc))
424 break;
425 }
426}
427
428/**
429 * nfs_pageio_cond_complete - Conditional I/O completion
430 * @desc: pointer to io descriptor
431 * @index: page index
432 *
433 * It is important to ensure that processes don't try to take locks
434 * on non-contiguous ranges of pages as that might deadlock. This
435 * function should be called before attempting to wait on a locked
436 * nfs_page. It will complete the I/O if the page index 'index'
437 * is not contiguous with the existing list of pages in 'desc'.
438 */
439void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
440{
441 if (!list_empty(&desc->pg_list)) {
442 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
443 if (index != prev->wb_index + 1)
444 nfs_pageio_complete(desc);
445 }
446}
447
448int __init nfs_init_nfspagecache(void)
449{
450 nfs_page_cachep = kmem_cache_create("nfs_page",
451 sizeof(struct nfs_page),
452 0, SLAB_HWCACHE_ALIGN,
453 NULL);
454 if (nfs_page_cachep == NULL)
455 return -ENOMEM;
456
457 return 0;
458}
459
460void nfs_destroy_nfspagecache(void)
461{
462 kmem_cache_destroy(nfs_page_cachep);
463}
464