Whamcloud - gitweb
merge b_md onto HEAD. as best as I can remember:
[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, 2002 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         pg.flag = create ? OBD_BRW_CREATE : 0;
111
112         set->brw_callback = ll_brw_sync_wait;
113         rc = obd_brw(cmd, ll_i2obdconn(inode), lsm, 1, &pg, set);
114         if (rc)
115                 CERROR("error from obd_brw: rc = %d\n", rc);
116         else {
117                 rc = ll_brw_sync_wait(set, CB_PHASE_START);
118                 if (rc)
119                         CERROR("error from callback: rc = %d\n", rc);
120         }
121         obd_brw_set_free(set);
122
123         RETURN(rc);
124 }
125
126 /* returns the page unlocked, but with a reference */
127 static int ll_readpage(struct file *file, struct page *page)
128 {
129         struct inode *inode = page->mapping->host;
130         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
131         int rc = 0;
132         ENTRY;
133
134         if (!PageLocked(page))
135                 LBUG();
136
137         if (inode->i_size <= offset) {
138                 memset(kmap(page), 0, PAGE_SIZE);
139                 kunmap(page);
140                 GOTO(readpage_out, rc);
141         }
142
143         if (PageUptodate(page)) {
144                 CERROR("Explain this please?\n");
145                 GOTO(readpage_out, rc);
146         }
147
148         rc = ll_brw(OBD_BRW_READ, inode, page, 0);
149         EXIT;
150
151  readpage_out:
152         if (!rc)
153                 SetPageUptodate(page);
154         unlock_page(page);
155         return 0;
156 } /* ll_readpage */
157
158 void ll_truncate(struct inode *inode)
159 {
160         struct obdo oa = {0};
161         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
162         struct lustre_handle *lockhs = NULL;
163         int err;
164         ENTRY;
165
166         if (!lsm) {
167                 /* object not yet allocated */
168                 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
169                 return;
170         }
171
172         oa.o_id = lsm->lsm_object_id;
173         oa.o_mode = inode->i_mode;
174         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
175
176         CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after "LPD64")\n",
177                oa.o_id, inode->i_size);
178
179         err = ll_size_lock(inode, lsm, inode->i_size, LCK_PW, &lockhs);
180         if (err) {
181                 CERROR("ll_size_lock failed: %d\n", err);
182                 /* FIXME: What to do here?  It's too late to back out... */
183                 LBUG();
184         }
185
186         /* truncate == punch from new size to absolute end of file */
187         err = obd_punch(ll_i2obdconn(inode), &oa, lsm, inode->i_size,
188                         OBD_OBJECT_EOF);
189         if (err) { 
190                 LBUG();
191                 CERROR("obd_truncate fails (%d) ino %ld\n", err,
192                        inode->i_ino);
193         } else
194                 obdo_to_inode(inode, &oa, oa.o_valid);
195
196         err = ll_size_unlock(inode, lsm, LCK_PW, lockhs);
197         if (err)
198                 CERROR("ll_size_unlock failed: %d\n", err);
199
200         EXIT;
201         return;
202 } /* ll_truncate */
203
204 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
205
206 static int ll_prepare_write(struct file *file, struct page *page, unsigned from,
207                             unsigned to)
208 {
209         struct inode *inode = page->mapping->host;
210         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
211         int rc = 0;
212         char *addr;
213         ENTRY;
214
215         addr = kmap(page);
216         if (!PageLocked(page))
217                 LBUG();
218
219         if (Page_Uptodate(page))
220                 GOTO(prepare_done, rc);
221
222         /* We're completely overwriting an existing page, so _don't_ set it up
223          * to date until commit_write */
224         if (from == 0 && to == PAGE_SIZE)
225                 RETURN(0);
226
227         /* We are writing to a new page, no need to read old data */
228         if (inode->i_size <= offset) {
229                 memset(addr, 0, PAGE_SIZE);
230                 GOTO(prepare_done, rc=0);
231         }
232
233         rc = ll_brw(OBD_BRW_READ, inode, page, 0);
234
235         EXIT;
236  prepare_done:
237         if (!rc)
238                 SetPageUptodate(page);
239
240         return rc;
241 }
242
243 /* returns the page unlocked, but with a reference */
244 static int ll_writepage(struct page *page)
245 {
246         struct inode *inode = page->mapping->host;
247         int err;
248         ENTRY;
249
250         if (!PageLocked(page))
251                 LBUG();
252
253         err = ll_brw(OBD_BRW_WRITE, inode, page, 1);
254         if ( !err ) {
255                 //SetPageUptodate(page);
256                 set_page_clean(page);
257         } else {
258                 CERROR("ll_brw failure %d\n", err);
259         }
260         unlock_page(page);
261         RETURN(err);
262 }
263
264
265 /* SYNCHRONOUS I/O to object storage for an inode -- object attr will be updated
266  * too */
267 static int ll_commit_write(struct file *file, struct page *page,
268                            unsigned from, unsigned to)
269 {
270         struct inode *inode = page->mapping->host;
271         struct ll_inode_info *lli = ll_i2info(inode);
272         struct lov_stripe_md *md = lli->lli_smd;
273         struct brw_page pg;
274         struct obd_brw_set *set;
275         int rc, create = 1;
276         loff_t size;
277         ENTRY;
278
279         pg.pg = page;
280         pg.count = to;
281         pg.off = (((obd_off)page->index) << PAGE_SHIFT);
282         pg.flag = create ? OBD_BRW_CREATE : 0;
283
284         set = obd_brw_set_new();
285         if (set == NULL)
286                 RETURN(-ENOMEM);
287
288         SetPageUptodate(page);
289
290         if (!PageLocked(page))
291                 LBUG();
292
293         CDEBUG(D_INODE, "commit_page writing (off "LPD64"), count "LPD64"\n",
294                pg.off, pg.count);
295
296         set->brw_callback = ll_brw_sync_wait;
297         rc = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode), md, 1, &pg, set);
298         if (rc)
299                 CERROR("error from obd_brw: rc = %d\n", rc);
300         else {
301                 rc = ll_brw_sync_wait(set, CB_PHASE_START);
302                 if (rc)
303                         CERROR("error from callback: rc = %d\n", rc);
304         }
305         obd_brw_set_free(set);
306         kunmap(page);
307
308         size = pg.off + pg.count;
309         /* do NOT truncate when writing in the middle of a file */
310         if (size > inode->i_size)
311                 inode->i_size = size;
312
313         RETURN(rc);
314 } /* ll_commit_write */
315
316
317 static int ll_direct_IO(int rw, struct inode *inode, struct kiobuf *iobuf,
318                         unsigned long blocknr, int blocksize)
319 {
320         obd_count bufs_per_obdo = iobuf->nr_pages;
321         struct ll_inode_info *lli = ll_i2info(inode);
322         struct lov_stripe_md *lsm = lli->lli_smd;
323         struct brw_page *pga;
324         struct obd_brw_set *set;
325         int i, rc = 0;
326         ENTRY;
327
328         if (!lsm || !lsm->lsm_object_id)
329                 RETURN(-ENOMEM);
330
331         if (blocksize != PAGE_SIZE) {
332                 CERROR("direct_IO blocksize != PAGE_SIZE\n");
333                 RETURN(-EINVAL);
334         }
335
336         set = obd_brw_set_new();
337         if (set == NULL)
338                 RETURN(-ENOMEM);
339
340         OBD_ALLOC(pga, sizeof(*pga) * bufs_per_obdo);
341         if (!pga) {
342                 obd_brw_set_free(set);
343                 RETURN(-ENOMEM);
344         }
345
346         /* NB: we can't use iobuf->maplist[i]->index for the offset
347          * instead of "blocknr" because ->index contains garbage.
348          */
349         for (i = 0; i < bufs_per_obdo; i++, blocknr++) {
350                 pga[i].pg = iobuf->maplist[i];
351                 pga[i].count = PAGE_SIZE;
352                 pga[i].off = (obd_off)blocknr << PAGE_SHIFT;
353                 pga[i].flag = OBD_BRW_CREATE;
354         }
355
356         set->brw_callback = ll_brw_sync_wait;
357         rc = obd_brw(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
358                      ll_i2obdconn(inode), lsm, bufs_per_obdo, pga, set);
359         if (rc)
360                 CERROR("error from obd_brw: rc = %d\n", rc);
361         else {
362                 rc = ll_brw_sync_wait(set, CB_PHASE_START);
363                 if (rc)
364                         CERROR("error from callback: rc = %d\n", rc);
365         }
366         obd_brw_set_free(set);
367         if (rc == 0)
368                 rc = bufs_per_obdo * PAGE_SIZE;
369
370         OBD_FREE(pga, sizeof(*pga) * bufs_per_obdo);
371         RETURN(rc);
372 }
373
374 int ll_flush_inode_pages(struct inode * inode)
375 {
376         obd_count        bufs_per_obdo = 0;
377         obd_size         *count = NULL;
378         obd_off          *offset = NULL;
379         obd_flag         *flags = NULL;
380         int              err = 0;
381
382         ENTRY;
383
384         spin_lock(&pagecache_lock);
385
386         spin_unlock(&pagecache_lock);
387
388
389         OBD_ALLOC(count, sizeof(*count) * bufs_per_obdo);
390         OBD_ALLOC(offset, sizeof(*offset) * bufs_per_obdo);
391         OBD_ALLOC(flags, sizeof(*flags) * bufs_per_obdo);
392         if (!count || !offset || !flags)
393                 GOTO(out, err=-ENOMEM);
394
395 #if 0
396         for (i = 0 ; i < bufs_per_obdo ; i++) {
397                 count[i] = PAGE_SIZE;
398                 offset[i] = ((obd_off)(iobuf->maplist[i])->index) << PAGE_SHIFT;
399                 flags[i] = OBD_BRW_CREATE;
400         }
401
402         err = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode),
403                       ll_i2info(inode)->lli_smd, bufs_per_obdo,
404                       iobuf->maplist, count, offset, flags, NULL, NULL);
405         if (err == 0)
406                 err = bufs_per_obdo * 4096;
407 #endif
408  out:
409         OBD_FREE(flags, sizeof(*flags) * bufs_per_obdo);
410         OBD_FREE(count, sizeof(*count) * bufs_per_obdo);
411         OBD_FREE(offset, sizeof(*offset) * bufs_per_obdo);
412         RETURN(err);
413 }
414
415 #endif
416
417
418 struct address_space_operations ll_aops = {
419         readpage: ll_readpage,
420 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
421         direct_IO: ll_direct_IO,
422         writepage: ll_writepage,
423         sync_page: block_sync_page,
424         prepare_write: ll_prepare_write,
425         commit_write: ll_commit_write,
426         bmap: NULL
427 #endif
428 };