Here is a better patch that doesn't deadlock (the last one did) and
uses the more proper generic_sync_sb_inodes instead of freeze_bdev.
I used the make-many-files.c program that Marcin dug up somewhere on
the net to test this, making the partition 4 GB, which is a little
too small to hold all the files. This exercises the nospace handling
nicely.
Note: one thing that freeze_bdev does that generic_sync_sb_inodes does
not is prevent new writes during the flush. We will have to think
about how this is to be handled.
Daniel
diff -r a49705ea1c95 user/kernel/commit.c
--- a/user/kernel/commit.c Tue Mar 03 20:43:11 2009 -0800
+++ b/user/kernel/commit.c Wed Mar 04 03:26:33 2009 -0800
@@ -4,6 +4,7 @@
*/
#include "tux3.h"
+#include <linux/writeback.h>
#ifndef trace
#define trace trace_on
@@ -32,6 +33,8 @@
sb->atomgen = from_be_u32(super->atomgen);
sb->freeatom = from_be_u32(super->freeatom);
sb->dictsize = from_be_u64(super->dictsize);
+ sb->minchange = 8; /* total blocks changed by smallest change */
+ sb->margin = 100; // should be tunable?
trace("blocksize %u, blockbits %u, blockmask %08x",
sb->blocksize, sb->blockbits, sb->blockmask);
trace("volblocks %Lu, freeblocks %Lu, nextalloc %Lu",
@@ -236,6 +239,50 @@
return 0;
}
+int reserve_credits(struct sb *sb, unsigned credits)
+{
+ sb->credits += credits;
+ if (sb->margin > 0 && sb->freeblocks < sb->credits + sb->margin) {
+#ifdef __KERNEL__
+ struct block_device *bdev = vfs_sb(sb)->s_bdev;
+ warn(">>> %Lx free, %Lx credits", (L)sb->freeblocks, (L)sb->credits);
+ sb->margin = -sb->margin;
+ generic_sync_sb_inodes(vfs_sb(sb), &(struct writeback_control){
+ .sync_mode = WB_SYNC_ALL,
+ .range_end = LLONG_MAX,
+ .nr_to_write = LONG_MAX });
+ sb->credits = 0;
+ thaw_bdev(bdev, vfs_sb(sb));
+ sb->margin = -sb->margin;
+#endif
+ warn(">>> nospace = %i", sb->freeblocks < sb->credits + sb->margin);
+ if (sb->freeblocks < sb->credits + sb->margin) {
+ return -ENOSPC;
+ }
+ }
+ return 0;
+}
+
+void release_credits(struct sb *sb, unsigned credits)
+{
+ // <try to return unused credits to pool> //
+}
+
+int reserve_begin(struct sb *sb)
+{
+ int err;
+ if ((err = reserve_credits(sb, sb->minchange)))
+ return err;
+ return change_begin(sb);
+}
+
+int reduce_begin(struct sb *sb)
+{
+ int err;
+ if ((err = reserve_credits(sb, 0)))
+ return err;
+ return change_begin(sb);
+}
#ifdef __KERNEL__
static void *useme[] = { clean_buffer, need_delta, stage_delta, commit_delta, useme };
#endif
diff -r a49705ea1c95 user/kernel/dir.c
--- a/user/kernel/dir.c Tue Mar 03 20:43:11 2009 -0800
+++ b/user/kernel/dir.c Wed Mar 04 03:26:33 2009 -0800
@@ -132,7 +132,7 @@
while (entry <= limit) {
if (entry->rec_len == 0) {
brelse(buffer);
- tux_error(dir->i_sb, "zero-length directory entry");
+ warn("[CORRUPTION] Zero length directory record detected!");
return -EIO;
}
name_len = TUX_REC_LEN(entry->name_len);
diff -r a49705ea1c95 user/kernel/filemap.c
--- a/user/kernel/filemap.c Tue Mar 03 20:43:11 2009 -0800
+++ b/user/kernel/filemap.c Wed Mar 04 03:26:33 2009 -0800
@@ -511,7 +511,11 @@
loff_t pos, unsigned len, unsigned flags,
struct page **pagep, void **fsdata)
{
+ struct sb *sb = tux_sb(mapping->host->i_sb);
+ int err;
*pagep = NULL;
+ if ((err = reserve_credits(sb, (len >> sb->blockbits) + 10)))
+ return err;
return block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
tux3_da_get_block);
}
@@ -519,8 +523,10 @@
static int tux3_writepage(struct page *page, struct writeback_control *wbc)
{
struct sb *sb = tux_sb(page->mapping->host->i_sb);
- change_begin(sb);
- int err = block_write_full_page(page, tux3_get_block, wbc);
+ int err = change_begin(sb);
+ if (err)
+ return err;
+ err = block_write_full_page(page, tux3_get_block, wbc);
change_end(sb);
return err;
}
diff -r a49705ea1c95 user/kernel/inode.c
--- a/user/kernel/inode.c Tue Mar 03 20:43:11 2009 -0800
+++ b/user/kernel/inode.c Wed Mar 04 03:26:33 2009 -0800
@@ -233,9 +233,10 @@
trace("save inode 0x%Lx", (L)tux_inode(inode)->inum);
struct sb *sb = tux_sb(inode->i_sb);
struct btree *itable = itable_btree(sb);
+ struct cursor *cursor;
int err;
- struct cursor *cursor = alloc_cursor(itable, 1); /* +1 for new depth */
- if (!cursor)
+
+ if (!(cursor = alloc_cursor(itable, 1))) /* +1 for new depth */
return -ENOMEM;
down_write(&cursor->btree->lock);
if ((err = probe(cursor, tux_inode(inode)->inum)))
diff -r a49705ea1c95 user/kernel/namei.c
--- a/user/kernel/namei.c Tue Mar 03 20:43:11 2009 -0800
+++ b/user/kernel/namei.c Wed Mar 04 03:26:33 2009 -0800
@@ -60,13 +60,14 @@
static int tux3_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t rdev)
{
+ struct sb *sb = tux_sb(dir->i_sb);
struct inode *inode;
int err;
if (!huge_valid_dev(rdev))
return -EINVAL;
-
- change_begin(tux_sb(dir->i_sb));
+ if ((err = reserve_begin(sb)))
+ return err;
inode = tux_create_inode(dir, mode, rdev);
err = PTR_ERR(inode);
if (!IS_ERR(inode)) {
@@ -81,7 +82,7 @@
iput(inode);
}
out:
- change_end(tux_sb(dir->i_sb));
+ change_end(sb);
return err;
}
@@ -100,13 +101,14 @@
static int tux3_link(struct dentry *old_dentry, struct inode *dir,
struct dentry *dentry)
{
+ struct sb *sb = tux_sb(dir->i_sb);
struct inode *inode = old_dentry->d_inode;
int err;
if (inode->i_nlink >= TUX_LINK_MAX)
return -EMLINK;
-
- change_begin(tux_sb(inode->i_sb));
+ if ((err = reserve_begin(sb)))
+ return err;
inode->i_ctime = gettime();
inode_inc_link_count(inode);
atomic_inc(&inode->i_count);
@@ -115,17 +117,19 @@
inode_dec_link_count(inode);
iput(inode);
}
- change_end(tux_sb(inode->i_sb));
+ change_end(sb);
return err;
}
static int tux3_symlink(struct inode *dir, struct dentry *dentry,
const char *symname)
{
+ struct sb *sb = tux_sb(dir->i_sb);
struct inode *inode;
int err;
- change_begin(tux_sb(dir->i_sb));
+ if ((err = reserve_begin(sb)))
+ return err;
inode = tux_create_inode(dir, S_IFLNK | S_IRWXUGO, 0);
err = PTR_ERR(inode);
if (!IS_ERR(inode)) {
@@ -139,41 +143,42 @@
iput(inode);
}
out:
- change_end(tux_sb(dir->i_sb));
+ change_end(sb);
return err;
}
static int tux3_unlink(struct inode *dir, struct dentry *dentry)
{
+ struct sb *sb = tux_sb(dir->i_sb);
struct inode *inode = dentry->d_inode;
- change_begin(tux_sb(inode->i_sb));
- int err = tux_del_dirent(dir, dentry);
- if (!err) {
+ int err;
+
+ if ((err = reduce_begin(sb)))
+ return err;
+ if (!(tux_del_dirent(dir, dentry))) {
inode->i_ctime = dir->i_ctime;
inode_dec_link_count(inode);
}
- change_end(tux_sb(inode->i_sb));
+ change_end(sb);
return err;
}
static int tux3_rmdir(struct inode *dir, struct dentry *dentry)
{
+ struct sb *sb = tux_sb(dir->i_sb);
struct inode *inode = dentry->d_inode;
int err;
- err = tux_dir_is_empty(inode);
- if (!err) {
- change_begin(tux_sb(inode->i_sb));
- err = tux_del_dirent(dir, dentry);
- if (!err) {
- inode->i_ctime = dir->i_ctime;
- inode->i_size = 0;
- clear_nlink(inode);
- mark_inode_dirty(inode);
- inode_dec_link_count(dir);
- }
- change_end(tux_sb(inode->i_sb));
+ if ((err = reduce_begin(sb)))
+ return err;
+ if (!(tux_del_dirent(dir, dentry))) {
+ inode->i_ctime = dir->i_ctime;
+ inode->i_size = 0;
+ clear_nlink(inode);
+ mark_inode_dirty(inode);
+ inode_dec_link_count(dir);
}
+ change_end(sb);
return err;
}
@@ -184,6 +189,7 @@
struct inode *new_inode = new_dentry->d_inode;
struct buffer_head *old_buffer, *new_buffer;
tux_dirent *old_entry, *new_entry;
+ struct sb *sb = tux_sb(old_dir->i_sb);
int err, new_subdir = 0;
old_entry = tux_find_entry(old_dir, old_dentry->d_name.name,
@@ -194,7 +200,8 @@
/* FIXME: is this needed? */
BUG_ON(from_be_u64(old_entry->inum) != tux_inode(old_inode)->inum);
- change_begin(tux_sb(old_inode->i_sb));
+ if ((err = reserve_begin(sb)))
+ return err;
if (new_inode) {
int old_is_dir = S_ISDIR(old_inode->i_mode);
if (old_is_dir) {
@@ -244,11 +251,11 @@
if (!err && new_subdir)
inode_dec_link_count(old_dir);
- change_end(tux_sb(old_inode->i_sb));
+ change_end(sb);
return err;
error:
- change_end(tux_sb(old_inode->i_sb));
+ change_end(sb);
brelse(old_buffer);
return err;
}
diff -r a49705ea1c95 user/kernel/tux3.h
--- a/user/kernel/tux3.h Tue Mar 03 20:43:11 2009 -0800
+++ b/user/kernel/tux3.h Wed Mar 04 03:26:33 2009 -0800
@@ -337,6 +337,9 @@
struct list_head pinned; /* dirty metadata not flushed per delta */
struct list_head commit; /* dirty metadata flushed per delta */
struct list_head dirty_inodes; /* dirty inodes list */
+ unsigned credits; /* Blocks reserved for inflight operations */
+ unsigned minchange; /* total blocks changed by smallest change */
+ int margin; /* ENOSPC when freeblocks less than this */
#ifdef __KERNEL__
struct super_block *vfs_sb; /* Generic kernel superblock */
#else
@@ -900,6 +903,9 @@
int load_itable(struct sb *sb);
int change_begin(struct sb *sb);
int change_end(struct sb *sb);
+int reserve_credits(struct sb *sb, unsigned credits);
+int reserve_begin(struct sb *sb);
+int reduce_begin(struct sb *sb);
/* temporary hack for buffer */
struct buffer_head *blockread(struct address_space *mapping, block_t iblock);
_______________________________________________
Tux3 mailing list
Tux3@tux3.org
http://mailman.tux3.org/cgi-bin/mailman/listinfo/tux3
