1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Lustre Lite I/O Page Cache
6 * Copyright (C) 2002 Cluster File Systems, Inc.
9 #include <linux/config.h>
10 #include <linux/kernel.h>
12 #include <linux/string.h>
13 #include <linux/stat.h>
14 #include <linux/iobuf.h>
15 #include <linux/errno.h>
16 #include <linux/locks.h>
17 #include <linux/unistd.h>
18 #include <linux/version.h>
19 #include <asm/system.h>
20 #include <asm/uaccess.h>
23 #include <linux/stat.h>
24 #include <asm/uaccess.h>
25 #include <asm/segment.h>
27 #include <linux/pagemap.h>
28 #include <linux/smp_lock.h>
30 #define DEBUG_SUBSYSTEM S_LLITE
32 #include <linux/lustre_mds.h>
33 #include <linux/lustre_lite.h>
34 #include <linux/lustre_lib.h>
37 /* SYNCHRONOUS I/O to object storage for an inode */
38 static int ll_brw(int rw, struct inode *inode, struct page *page, int create)
40 struct ll_inode_info *lli = ll_i2info(inode);
41 struct lov_stripe_md *md = lli->lli_smd;
44 struct io_cb_data *cbd = ll_init_cb();
52 pg.off = ((obd_off)page->index) << PAGE_SHIFT;
53 pg.flag = create ? OBD_BRW_CREATE : 0;
55 err = obd_brw(rw, ll_i2obdconn(inode), md, 1, &pg, ll_sync_io_cb, cbd);
60 /* returns the page unlocked, but with a reference */
61 static int ll_readpage(struct file *file, struct page *page)
63 struct inode *inode = page->mapping->host;
64 obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
68 if (!PageLocked(page))
71 if (inode->i_size <= offset) {
72 memset(kmap(page), 0, PAGE_SIZE);
74 GOTO(readpage_out, rc);
77 if (Page_Uptodate(page)) {
78 CERROR("Explain this please?\n");
79 GOTO(readpage_out, rc);
82 rc = ll_brw(OBD_BRW_READ, inode, page, 0);
87 SetPageUptodate(page);
93 static int ll_prepare_write(struct file *file, struct page *page, unsigned from,
96 struct inode *inode = page->mapping->host;
97 obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
103 if (!PageLocked(page))
106 if (Page_Uptodate(page))
107 GOTO(prepare_done, rc);
109 /* We're completely overwriting an existing page, so _don't_ set it up
110 * to date until commit_write */
111 if (from == 0 && to == PAGE_SIZE)
114 /* We are writing to a new page, no need to read old data */
115 if (inode->i_size <= offset) {
116 memset(addr, 0, PAGE_SIZE);
120 /* prepare write should not read what lies beyond the end of
122 rc = ll_brw(OBD_BRW_READ, inode, page, 0);
127 SetPageUptodate(page);
132 /* returns the page unlocked, but with a reference */
133 static int ll_writepage(struct page *page)
135 struct inode *inode = page->mapping->host;
139 if (!PageLocked(page))
142 err = ll_brw(OBD_BRW_WRITE, inode, page, 1);
144 //SetPageUptodate(page);
145 set_page_clean(page);
147 CERROR("ll_brw failure %d\n", err);
154 /* SYNCHRONOUS I/O to object storage for an inode -- object attr will be updated
156 static int ll_commit_write(struct file *file, struct page *page,
157 unsigned from, unsigned to)
160 struct inode *inode = page->mapping->host;
161 struct ll_inode_info *lli = ll_i2info(inode);
162 struct lov_stripe_md *md = lli->lli_smd;
166 struct io_cb_data *cbd = ll_init_cb();
170 pg.off = (((obd_off)page->index) << PAGE_SHIFT);
171 pg.flag = create ? OBD_BRW_CREATE : 0;
177 SetPageUptodate(page);
179 if (!PageLocked(page))
182 CDEBUG(D_INODE, "commit_page writing (at %d) to %d, count %Ld\n",
183 from, to, (unsigned long long)pg.count);
185 err = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode), md,
186 1, &pg, ll_sync_io_cb, cbd);
189 iattr.ia_size = pg.off + pg.count;
190 if (iattr.ia_size > inode->i_size) {
191 /* do NOT truncate when writing in the middle of a file */
192 inode->i_size = iattr.ia_size;
193 iattr.ia_valid = ATTR_SIZE;
195 err = ll_inode_setattr(inode, &iattr, 0);
197 CERROR("failed - %d.\n", err);
203 } /* ll_commit_write */
205 void ll_truncate(struct inode *inode)
207 struct obdo oa = {0};
208 struct lov_stripe_md *md = ll_i2info(inode)->lli_smd;
209 struct lustre_handle *lockhs = NULL;
214 /* object not yet allocated */
215 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
219 oa.o_id = md->lmd_object_id;
220 oa.o_size = inode->i_size;
222 CDEBUG(D_INFO, "calling punch for %ld (all bytes after %Ld)\n",
223 (long)oa.o_id, (unsigned long long)oa.o_size);
225 err = ll_size_lock(inode, md, oa.o_size, LCK_PW, &lockhs);
227 CERROR("ll_size_lock failed: %d\n", err);
228 /* FIXME: What to do here? It's too late to back out... */
232 oa.o_valid = OBD_MD_FLID;
233 /* truncate == punch to/from start from/to end:
234 set end to -1 for that. */
235 err = obd_punch(ll_i2obdconn(inode), &oa, md, inode->i_size,
238 CERROR("obd_truncate fails (%d)\n", err);
240 /* This is done for us at the OST and MDS, but the
241 * updated timestamps are not sent back to us.
244 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
246 err = ll_size_unlock(inode, md, LCK_PW, lockhs);
248 CERROR("ll_size_unlock failed: %d\n", err);
254 int ll_direct_IO(int rw, struct inode *inode, struct kiobuf *iobuf,
255 unsigned long blocknr, int blocksize)
257 obd_count bufs_per_obdo = iobuf->nr_pages;
258 struct ll_inode_info *lli = ll_i2info(inode);
259 struct lov_stripe_md *md = lli->lli_smd;
260 struct brw_page *pga;
263 struct io_cb_data *cbd = ll_init_cb();
269 if (blocksize != PAGE_SIZE) {
270 CERROR("direct_IO blocksize != PAGE_SIZE\n");
274 OBD_ALLOC(pga, sizeof(*pga) * bufs_per_obdo);
276 GOTO(out, rc = -ENOMEM);
278 /* NB: we can't use iobuf->maplist[i]->index for the offset
279 * instead of "blocknr" because ->index contains garbage.
281 for (i = 0; i < bufs_per_obdo; i++, blocknr++) {
282 pga[i].pg = iobuf->maplist[i];
283 pga[i].count = PAGE_SIZE;
284 pga[i].off = (obd_off)blocknr << PAGE_SHIFT;
285 pga[i].flag = OBD_BRW_CREATE;
288 if (!md || !md->lmd_object_id)
289 GOTO(out, rc = -ENOMEM);
291 rc = obd_brw(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
292 ll_i2obdconn(inode), md, bufs_per_obdo, pga,
295 rc = bufs_per_obdo * PAGE_SIZE;
298 OBD_FREE(pga, sizeof(*pga) * bufs_per_obdo);
303 int ll_flush_inode_pages(struct inode * inode)
305 obd_count bufs_per_obdo = 0;
306 obd_size *count = NULL;
307 obd_off *offset = NULL;
308 obd_flag *flags = NULL;
313 spin_lock(&pagecache_lock);
315 spin_unlock(&pagecache_lock);
318 OBD_ALLOC(count, sizeof(*count) * bufs_per_obdo);
319 OBD_ALLOC(offset, sizeof(*offset) * bufs_per_obdo);
320 OBD_ALLOC(flags, sizeof(*flags) * bufs_per_obdo);
321 if (!count || !offset || !flags)
322 GOTO(out, err=-ENOMEM);
325 for (i = 0 ; i < bufs_per_obdo ; i++) {
326 count[i] = PAGE_SIZE;
327 offset[i] = ((obd_off)(iobuf->maplist[i])->index) << PAGE_SHIFT;
328 flags[i] = OBD_BRW_CREATE;
331 err = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode),
332 ll_i2info(inode)->lli_smd, bufs_per_obdo,
333 iobuf->maplist, count, offset, flags, NULL, NULL);
335 err = bufs_per_obdo * 4096;
338 OBD_FREE(flags, sizeof(*flags) * bufs_per_obdo);
339 OBD_FREE(count, sizeof(*count) * bufs_per_obdo);
340 OBD_FREE(offset, sizeof(*offset) * bufs_per_obdo);
346 struct address_space_operations ll_aops = {
347 readpage: ll_readpage,
348 writepage: ll_writepage,
349 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,4,17))
350 direct_IO: ll_direct_IO,
352 sync_page: block_sync_page,
353 prepare_write: ll_prepare_write,
354 commit_write: ll_commit_write,