Whamcloud - gitweb
file ext3-largefile.diff was initially added on branch b_devel.
[fs/lustre-release.git] / lustre / llite / rw.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * Lustre Lite I/O Page Cache
5  *
6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
7  *
8  *   This file is part of Lustre, http://www.lustre.org.
9  *
10  *   Lustre is free software; you can redistribute it and/or
11  *   modify it under the terms of version 2 of the GNU General Public
12  *   License as published by the Free Software Foundation.
13  *
14  *   Lustre is distributed in the hope that it will be useful,
15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  *   GNU General Public License for more details.
18  *
19  *   You should have received a copy of the GNU General Public License
20  *   along with Lustre; if not, write to the Free Software
21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
22  */
23
24 #include <linux/config.h>
25 #include <linux/kernel.h>
26 #include <linux/mm.h>
27 #include <linux/string.h>
28 #include <linux/stat.h>
29 #include <linux/iobuf.h>
30 #include <linux/errno.h>
31 #include <linux/smp_lock.h>
32 #include <linux/unistd.h>
33 #include <linux/version.h>
34 #include <asm/system.h>
35 #include <asm/uaccess.h>
36
37 #include <linux/fs.h>
38 #include <linux/stat.h>
39 #include <asm/uaccess.h>
40 #include <asm/segment.h>
41 #include <linux/mm.h>
42 #include <linux/pagemap.h>
43 #include <linux/smp_lock.h>
44
45 #define DEBUG_SUBSYSTEM S_LLITE
46
47 #include <linux/lustre_mds.h>
48 #include <linux/lustre_lite.h>
49 #include <linux/lustre_lib.h>
50
51 /*
52  * Remove page from dirty list
53  */
54 static void __set_page_clean(struct page *page)
55 {
56         struct address_space *mapping = page->mapping;
57         struct inode *inode;
58
59         if (!mapping)
60                 return;
61
62 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
63         spin_lock(&pagecache_lock);
64 #endif
65
66         list_del(&page->list);
67         list_add(&page->list, &mapping->clean_pages);
68
69         inode = mapping->host;
70         if (list_empty(&mapping->dirty_pages)) {
71                 CDEBUG(D_INODE, "inode clean\n");
72                 inode->i_state &= ~I_DIRTY_PAGES;
73         }
74 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
75         spin_unlock(&pagecache_lock);
76 #endif
77         EXIT;
78 }
79
80 inline void set_page_clean(struct page *page)
81 {
82         if (PageDirty(page)) {
83                 ClearPageDirty(page);
84                 __set_page_clean(page);
85         }
86 }
87
88 /* SYNCHRONOUS I/O to object storage for an inode */
89 static int ll_brw(int cmd, struct inode *inode, struct page *page, int create)
90 {
91         struct ll_inode_info *lli = ll_i2info(inode);
92         struct lov_stripe_md *lsm = lli->lli_smd;
93         struct obd_brw_set *set;
94         struct brw_page pg;
95         int rc;
96         ENTRY;
97
98         set = obd_brw_set_new();
99         if (set == NULL)
100                 RETURN(-ENOMEM);
101
102         pg.pg = page;
103         pg.off = ((obd_off)page->index) << PAGE_SHIFT;
104
105         if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > inode->i_size))
106                 pg.count = inode->i_size % PAGE_SIZE;
107         else
108                 pg.count = PAGE_SIZE;
109
110         CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
111               cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
112               pg.off, pg.off);
113         if (pg.count == 0) {
114                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
115                        LPU64"\n",
116                        inode->i_ino, inode, inode->i_size, page->mapping->host,
117                        page->mapping->host->i_size, page->index, pg.off);
118         }
119
120         pg.flag = create ? OBD_BRW_CREATE : 0;
121
122         set->brw_callback = ll_brw_sync_wait;
123         rc = obd_brw(cmd, ll_i2obdconn(inode), lsm, 1, &pg, set, NULL);
124         if (rc) {
125                 if (rc != -EIO)
126                         CERROR("error from obd_brw: rc = %d\n", rc);
127         } else {
128                 rc = ll_brw_sync_wait(set, CB_PHASE_START);
129                 if (rc)
130                         CERROR("error from callback: rc = %d\n", rc);
131         }
132         obd_brw_set_free(set);
133
134         RETURN(rc);
135 }
136
137 /* returns the page unlocked, but with a reference */
138 static int ll_readpage(struct file *file, struct page *page)
139 {
140         struct inode *inode = page->mapping->host;
141         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
142         int rc = 0;
143         ENTRY;
144
145         if (!PageLocked(page))
146                 LBUG();
147
148         if (inode->i_size <= offset) {
149                 memset(kmap(page), 0, PAGE_SIZE);
150                 kunmap(page);
151                 GOTO(readpage_out, rc);
152         }
153
154         if (PageUptodate(page)) {
155                 CERROR("Explain this please?\n");
156                 GOTO(readpage_out, rc);
157         }
158
159         rc = ll_brw(OBD_BRW_READ, inode, page, 0);
160         EXIT;
161
162  readpage_out:
163         if (!rc)
164                 SetPageUptodate(page);
165         unlock_page(page);
166         return 0;
167 } /* ll_readpage */
168
169 void ll_truncate(struct inode *inode)
170 {
171         struct obdo oa = {0};
172         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
173         struct lustre_handle lockh = { 0, 0 };
174         int err;
175         ENTRY;
176
177         if (!lsm) {
178                 /* object not yet allocated */
179                 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
180                 return;
181         }
182
183         oa.o_id = lsm->lsm_object_id;
184         oa.o_mode = inode->i_mode;
185         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
186
187         CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n",
188                oa.o_id, inode->i_size);
189
190         err = ll_size_lock(inode, lsm, inode->i_size, LCK_PW, &lockh);
191         if (err) {
192                 CERROR("ll_size_lock failed: %d\n", err);
193                 return;
194         }
195
196         /* truncate == punch from new size to absolute end of file */
197         err = obd_punch(ll_i2obdconn(inode), &oa, lsm, inode->i_size,
198                         OBD_OBJECT_EOF, NULL);
199         if (err)
200                 CERROR("obd_truncate fails (%d) ino %lu\n", err, inode->i_ino);
201         else
202                 obdo_to_inode(inode, &oa, oa.o_valid);
203
204         err = ll_size_unlock(inode, lsm, LCK_PW, &lockh);
205         if (err)
206                 CERROR("ll_size_unlock failed: %d\n", err);
207
208         EXIT;
209         return;
210 } /* ll_truncate */
211
212 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
213
214 static int ll_prepare_write(struct file *file, struct page *page, unsigned from,
215                             unsigned to)
216 {
217         struct inode *inode = page->mapping->host;
218         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
219         int rc = 0;
220         char *addr;
221         ENTRY;
222
223         addr = kmap(page);
224         if (!PageLocked(page))
225                 LBUG();
226
227         if (Page_Uptodate(page))
228                 GOTO(prepare_done, rc);
229
230         /* We're completely overwriting an existing page, so _don't_ set it up
231          * to date until commit_write */
232         if (from == 0 && to == PAGE_SIZE)
233                 RETURN(0);
234
235         /* If are writing to a new page, no need to read old data.  If we
236          * haven't already gotten the file size in ll_file_write() since
237          * we got our extent lock, we need to verify it here before we
238          * overwrite some other node's write (bug 445).
239          */
240         if (inode->i_size <= offset) {
241                 if (!S_ISBLK(inode->i_mode) && !(file->f_flags & O_APPEND)) {
242                         struct ll_file_data *fd = file->private_data;
243                         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
244
245                         rc = ll_file_size(inode, lsm, &fd->fd_osthandle);
246                         if (rc)
247                                 GOTO(prepare_done, rc);
248                 }
249                 if (inode->i_size <= offset) {
250                         memset(addr, 0, PAGE_SIZE);
251                         GOTO(prepare_done, rc=0);
252                 }
253         }
254
255         rc = ll_brw(OBD_BRW_READ, inode, page, 0);
256
257         EXIT;
258  prepare_done:
259         if (!rc)
260                 SetPageUptodate(page);
261         else
262                 kunmap (page);
263         
264         return rc;
265 }
266
267 /* Write a page from kupdated or kswapd.
268  *
269  * We unlock the page even in the face of an error, otherwise dirty
270  * pages could OOM the system if they cannot be written.  Also, there
271  * is nobody to return an error code to from here - the application
272  * may not even be running anymore.
273  *
274  * Returns the page unlocked, but with a reference.
275  */
276 static int ll_writepage(struct page *page) {
277         struct inode *inode = page->mapping->host; int err; ENTRY;
278
279         LASSERT(PageLocked(page));
280
281         /* XXX need to make sure we have LDLM lock on this page */
282         err = ll_brw(OBD_BRW_WRITE, inode, page, 1);
283         if (err)
284                 CERROR("ll_brw failure %d\n", err);
285         else
286                 set_page_clean(page);
287
288         unlock_page(page);
289         RETURN(err);
290 }
291
292
293 /* SYNCHRONOUS I/O to object storage for an inode -- object attr will be updated
294  * too */
295 static int ll_commit_write(struct file *file, struct page *page,
296                            unsigned from, unsigned to)
297 {
298         struct inode *inode = page->mapping->host;
299         struct ll_inode_info *lli = ll_i2info(inode);
300         struct lov_stripe_md *md = lli->lli_smd;
301         struct brw_page pg;
302         struct obd_brw_set *set;
303         int rc, create = 1;
304         loff_t size;
305         ENTRY;
306
307         pg.pg = page;
308         pg.count = to;
309         /* XXX make the starting offset "from" */
310         pg.off = (((obd_off)page->index) << PAGE_SHIFT);
311         pg.flag = create ? OBD_BRW_CREATE : 0;
312
313         set = obd_brw_set_new();
314         if (set == NULL)
315                 RETURN(-ENOMEM);
316
317         SetPageUptodate(page);
318
319         if (!PageLocked(page))
320                 LBUG();
321
322         CDEBUG(D_INODE, "commit_page writing (off "LPD64"), count %d\n",
323                pg.off, pg.count);
324
325         set->brw_callback = ll_brw_sync_wait;
326         rc = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode), md, 1, &pg, set, NULL);
327         if (rc)
328                 CERROR("error from obd_brw: rc = %d\n", rc);
329         else {
330                 rc = ll_brw_sync_wait(set, CB_PHASE_START);
331                 if (rc)
332                         CERROR("error from callback: rc = %d\n", rc);
333         }
334         obd_brw_set_free(set);
335         kunmap(page);
336
337         size = pg.off + pg.count;
338         /* do NOT truncate when writing in the middle of a file */
339         if (size > inode->i_size)
340                 inode->i_size = size;
341
342         RETURN(rc);
343 } /* ll_commit_write */
344
345
346 static int ll_direct_IO(int rw, struct inode *inode, struct kiobuf *iobuf,
347                         unsigned long blocknr, int blocksize)
348 {
349         obd_count bufs_per_obdo = iobuf->nr_pages;
350         struct ll_inode_info *lli = ll_i2info(inode);
351         struct lov_stripe_md *lsm = lli->lli_smd;
352         struct brw_page *pga;
353         struct obd_brw_set *set;
354         int i, rc = 0;
355         ENTRY;
356
357         if (!lsm || !lsm->lsm_object_id)
358                 RETURN(-ENOMEM);
359
360         if (blocksize != PAGE_SIZE) {
361                 CERROR("direct_IO blocksize != PAGE_SIZE\n");
362                 RETURN(-EINVAL);
363         }
364
365         set = obd_brw_set_new();
366         if (set == NULL)
367                 RETURN(-ENOMEM);
368
369         OBD_ALLOC(pga, sizeof(*pga) * bufs_per_obdo);
370         if (!pga) {
371                 obd_brw_set_free(set);
372                 RETURN(-ENOMEM);
373         }
374
375         /* NB: we can't use iobuf->maplist[i]->index for the offset
376          * instead of "blocknr" because ->index contains garbage.
377          */
378         for (i = 0; i < bufs_per_obdo; i++, blocknr++) {
379                 pga[i].pg = iobuf->maplist[i];
380                 pga[i].count = PAGE_SIZE;
381                 pga[i].off = (obd_off)blocknr << PAGE_SHIFT;
382                 pga[i].flag = OBD_BRW_CREATE;
383         }
384
385         set->brw_callback = ll_brw_sync_wait;
386         rc = obd_brw(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
387                      ll_i2obdconn(inode), lsm, bufs_per_obdo, pga, set, NULL);
388         if (rc)
389                 CERROR("error from obd_brw: rc = %d\n", rc);
390         else {
391                 rc = ll_brw_sync_wait(set, CB_PHASE_START);
392                 if (rc)
393                         CERROR("error from callback: rc = %d\n", rc);
394         }
395         obd_brw_set_free(set);
396         if (rc == 0)
397                 rc = bufs_per_obdo * PAGE_SIZE;
398
399         OBD_FREE(pga, sizeof(*pga) * bufs_per_obdo);
400         RETURN(rc);
401 }
402
403 int ll_flush_inode_pages(struct inode * inode)
404 {
405         obd_count        bufs_per_obdo = 0;
406         obd_size         *count = NULL;
407         obd_off          *offset = NULL;
408         obd_flag         *flags = NULL;
409         int              err = 0;
410
411         ENTRY;
412
413         spin_lock(&pagecache_lock);
414
415         spin_unlock(&pagecache_lock);
416
417
418         OBD_ALLOC(count, sizeof(*count) * bufs_per_obdo);
419         OBD_ALLOC(offset, sizeof(*offset) * bufs_per_obdo);
420         OBD_ALLOC(flags, sizeof(*flags) * bufs_per_obdo);
421         if (!count || !offset || !flags)
422                 GOTO(out, err=-ENOMEM);
423
424 #if 0
425         for (i = 0 ; i < bufs_per_obdo ; i++) {
426                 count[i] = PAGE_SIZE;
427                 offset[i] = ((obd_off)(iobuf->maplist[i])->index) << PAGE_SHIFT;
428                 flags[i] = OBD_BRW_CREATE;
429         }
430
431         err = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode),
432                       ll_i2info(inode)->lli_smd, bufs_per_obdo,
433                       iobuf->maplist, count, offset, flags, NULL, NULL);
434         if (err == 0)
435                 err = bufs_per_obdo * 4096;
436 #endif
437  out:
438         OBD_FREE(flags, sizeof(*flags) * bufs_per_obdo);
439         OBD_FREE(count, sizeof(*count) * bufs_per_obdo);
440         OBD_FREE(offset, sizeof(*offset) * bufs_per_obdo);
441         RETURN(err);
442 }
443
444 #endif
445
446
447 struct address_space_operations ll_aops = {
448         readpage: ll_readpage,
449 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
450         direct_IO: ll_direct_IO,
451         writepage: ll_writepage,
452         sync_page: block_sync_page,
453         prepare_write: ll_prepare_write,
454         commit_write: ll_commit_write,
455         bmap: NULL
456 #endif
457 };