1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.lustre.org.
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 * Copyright (C) 2002, 2003 Cluster File Systems, Inc
23 * this started as an implementation of an io daemon that woke regularly
24 * to force writeback.. the throttling in prepare_write and kupdate's usual
25 * writeback pressure got rid of our thread, but the file name remains.
28 #include <linux/version.h>
29 #include <linux/config.h>
30 #include <linux/module.h>
32 #include <linux/stat.h>
33 #include <linux/sched.h>
34 #include <linux/smp_lock.h>
35 #include <linux/kmod.h>
36 #include <linux/pagemap.h>
38 #include <linux/rbtree.h>
39 #include <linux/seq_file.h>
40 #include <linux/time.h>
42 /* PG_inactive_clean is shorthand for rmap, we want free_high/low here.. */
43 #ifdef PG_inactive_clean
44 #include <linux/mm_inline.h>
47 #define DEBUG_SUBSYSTEM S_LLITE
48 #include <linux/lustre_lite.h>
49 #include "llite_internal.h"
51 #ifndef list_for_each_prev_safe
52 #define list_for_each_prev_safe(pos, n, head) \
53 for (pos = (head)->prev, n = pos->prev; pos != (head); \
54 pos = n, n = pos->prev )
57 extern spinlock_t inode_lock;
60 * check to see if we're racing with truncate and put the page in
61 * the brw_page array. returns 0 if there is more room and 1
62 * if the array is full.
64 static int llwp_consume_page(struct ll_writeback_pages *llwp,
65 struct inode *inode, struct page *page)
67 obd_off off = ((obd_off)page->index) << PAGE_SHIFT;
70 /* we raced with truncate? */
71 if ( off >= inode->i_size ) {
73 rc = ll_clear_dirty_pages(ll_i2obdconn(inode),
74 ll_i2info(inode)->lli_smd,
75 page->index, page->index);
78 CDEBUG(D_CACHE, "offset "LPU64" (index %lu) > i_size %llu\n",
79 off, page->index, inode->i_size);
85 pg = &llwp->pga[llwp->npgs];
87 LASSERT(llwp->npgs <= llwp->max);
91 pg->flag = OBD_BRW_CREATE|OBD_BRW_FROM_GRANT;
92 pg->count = PAGE_CACHE_SIZE;
94 /* catch partial writes for files that end mid-page */
95 if (pg->off + pg->count > inode->i_size)
96 pg->count = inode->i_size & ~PAGE_CACHE_MASK;
99 * matches ptlrpc_bulk_get assert that trickles down
100 * from a 0 page length going through niobuf and into
101 * the buffer regions being posted
103 LASSERT(pg->count >= 0);
105 CDEBUG(D_CACHE, "brw_page %p: off "LPU64" cnt %d, page %p: ind %ld"
106 " i_size: %llu\n", pg, pg->off, pg->count, page,
107 page->index, inode->i_size);
109 return llwp->npgs == llwp->max;
113 * returns the number of pages that it added to the pgs array
115 * this duplicates filemap_fdatasync and gives us an opportunity to grab lots
118 static void ll_get_dirty_pages(struct inode *inode,
119 struct ll_writeback_pages *llwp)
121 struct address_space *mapping = inode->i_mapping;
123 struct list_head *pos, *n;
126 PGCACHE_WRLOCK(mapping);
128 list_for_each_prev_safe(pos, n, &mapping->dirty_pages) {
129 page = list_entry(pos, struct page, list);
131 if (TryLockPage(page))
134 list_del(&page->list);
135 list_add(&page->list, &mapping->locked_pages);
137 if (!PageDirty(page)) {
141 ClearPageDirty(page);
143 if (llwp_consume_page(llwp, inode, page) != 0)
147 PGCACHE_WRUNLOCK(mapping);
151 static void ll_writeback(struct inode *inode, struct obdo *oa,
152 struct ll_writeback_pages *llwp)
154 struct ptlrpc_request_set *set;
158 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),bytes=%u\n",
159 inode->i_ino, inode->i_generation, inode,
160 ((llwp->npgs-1) << PAGE_SHIFT) + llwp->pga[llwp->npgs-1].count);
162 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
163 set = ptlrpc_prep_set();
165 CERROR ("Can't create request set\n");
168 rc = obd_brw_async(OBD_BRW_WRITE, ll_i2obdconn(inode), oa,
169 ll_i2info(inode)->lli_smd, llwp->npgs,
170 llwp->pga, set, NULL);
172 rc = ptlrpc_set_wait(set);
174 obdo_refresh_inode(inode, oa,
175 oa->o_valid & ~OBD_MD_FLSIZE);
176 ptlrpc_set_destroy (set);
179 * b=1038, we need to pass _brw errors up so that writeback
180 * doesn't get stuck in recovery leaving processes stuck in
181 * D waiting for pages
184 CERROR("error from obd_brw_async: rc = %d\n", rc);
185 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
186 LPROC_LL_WB_FAIL, llwp->npgs);
188 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
189 LPROC_LL_WB_OK, (llwp->npgs));
192 for (i = 0 ; i < llwp->npgs ; i++) {
193 struct page *page = llwp->pga[i].pg;
195 CDEBUG(D_CACHE, "finished page %p at index %lu\n", page,
197 LASSERT(PageLocked(page));
199 rc = ll_clear_dirty_pages(ll_i2obdconn(inode),
200 ll_i2info(inode)->lli_smd,
201 page->index, page->index);
204 page_cache_release(page);
210 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
212 #ifndef PG_inactive_clean
213 #ifdef CONFIG_DISCONTIGMEM
214 #error "sorry, we don't support DISCONTIGMEM yet"
218 * __alloc_pages marks a zone as needing balancing if an allocation is
219 * performed when the zone has fewer free pages than its 'low' water
220 * mark. its cleared when try_to_free_pages makes progress.
222 static int zones_need_balancing(void)
228 for ( pgdat = pgdat_list ; pgdat != NULL ; pgdat = pgdat->node_next ) {
229 for ( i = pgdat->nr_zones-1 ; i >= 0 ; i-- ) {
230 zone = &pgdat->node_zones[i];
232 if ( zone->need_balance )
239 /* 2.4 doesn't give us a way to find out how many pages we have
240 * cached 'cause we're not using buffer_heads. we are very
241 * conservative here and flush the superblock of all dirty data
242 * when the vm (rmap or stock) thinks that it is running low
243 * and kswapd would have done work. kupdated isn't good enough
244 * because writers (dbench) can dirty _very quickly_, and we
245 * allocate under writepage..
247 * 2.5 gets this right, see the {inc,dec}_page_state(nr_dirty, )
249 static int should_writeback(void)
251 #ifdef PG_inactive_clean
252 if (free_high(ALL_ZONES) > 0 || free_low(ANY_ZONE) > 0)
254 if (zones_need_balancing())
260 static int ll_alloc_brw(struct inode *inode, struct ll_writeback_pages *llwp)
262 memset(llwp, 0, sizeof(struct ll_writeback_pages));
264 llwp->max = inode->i_blksize >> PAGE_CACHE_SHIFT;
265 if (llwp->max == 0) {
266 CERROR("forcing llwp->max to 1. blksize: %lu\n",
270 llwp->pga = kmalloc(llwp->max * sizeof(*llwp->pga), GFP_ATOMIC);
271 if (llwp->pga == NULL)
276 int ll_check_dirty(struct super_block *sb)
278 unsigned long old_flags; /* hack? */
285 if (!should_writeback())
288 old_flags = current->flags;
289 current->flags |= PF_MEMALLOC;
291 spin_lock(&inode_lock);
294 * first we try and write back dirty pages from dirty inodes
295 * until the VM thinkgs we're ok again..
298 struct ll_writeback_pages llwp;
299 struct list_head *pos;
303 list_for_each_prev(pos, &sb->s_dirty) {
304 inode = list_entry(pos, struct inode, i_list);
306 if (!(inode->i_state & I_DIRTY_PAGES)) {
316 /* duplicate __sync_one, *sigh* */
317 list_del(&inode->i_list);
318 list_add(&inode->i_list, &inode->i_sb->s_locked_inodes);
319 inode->i_state |= I_LOCK;
320 inode->i_state &= ~I_DIRTY_PAGES;
322 spin_unlock(&inode_lock);
324 rc = ll_alloc_brw(inode, &llwp);
330 ll_get_dirty_pages(inode, &llwp);
333 ll_i2info(inode)->lli_smd->lsm_object_id;
334 oa.o_valid = OBD_MD_FLID;
335 obdo_from_inode(&oa, inode,
336 OBD_MD_FLTYPE | OBD_MD_FLATIME|
337 OBD_MD_FLMTIME| OBD_MD_FLCTIME);
338 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
339 LPROC_LL_WB_PRESSURE,
341 ll_writeback(inode, &oa, &llwp);
345 } while (llwp.npgs && should_writeback());
347 spin_lock(&inode_lock);
349 if (!list_empty(&inode->i_mapping->dirty_pages))
350 inode->i_state |= I_DIRTY_PAGES;
352 inode->i_state &= ~I_LOCK;
354 * we are sneaky and leave the inode on the dirty list,
355 * even though it might not still be..
357 if (!(inode->i_state & I_FREEING)) {
358 list_del(&inode->i_list);
359 list_add(&inode->i_list, &inode->i_sb->s_dirty);
361 wake_up(&inode->i_wait);
363 } while (making_progress && should_writeback());
366 * and if that didn't work, we sleep on any data that might
367 * be under writeback..
369 while (should_writeback()) {
370 if (list_empty(&sb->s_locked_inodes))
373 inode = list_entry(sb->s_locked_inodes.next, struct inode,
376 atomic_inc(&inode->i_count); /* XXX hack? */
377 spin_unlock(&inode_lock);
378 wait_event(inode->i_wait, !(inode->i_state & I_LOCK));
380 spin_lock(&inode_lock);
383 spin_unlock(&inode_lock);
386 current->flags = old_flags;
390 #endif /* linux 2.5 */
392 int ll_batch_writepage(struct inode *inode, struct obdo *oa, struct page *page)
394 unsigned long old_flags; /* hack? */
395 struct ll_writeback_pages llwp;
399 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
400 old_flags = current->flags;
401 current->flags |= PF_MEMALLOC;
402 rc = ll_alloc_brw(inode, &llwp);
404 GOTO(restore_flags, rc);
406 if (llwp_consume_page(&llwp, inode, page) == 0)
407 ll_get_dirty_pages(inode, &llwp);
410 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
411 LPROC_LL_WB_WRITEPAGE, llwp.npgs);
412 ll_writeback(inode, oa, &llwp);
417 current->flags = old_flags;