f2fs: speed up f2fs_empty_dir()
Chao Yu
maintainerPATCH 1/1 f2fs: speed up f2fs_empty_dir()
Wei Chen reports a kernel bug as blew:
INFO: task syz-executor.0:29056 blocked for more than 143 seconds.
Not tainted 5.15.0-rc5 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
task:syz-executor.0 state:D stack:14632 pid:29056 ppid: 6574 flags:0x00000004
Call Trace:
__schedule+0x4a1/0x1720
schedule+0x36/0xe0
rwsem_down_write_slowpath+0x322/0x7a0
fscrypt_ioctl_set_policy+0x11f/0x2a0
__f2fs_ioctl+0x1a9f/0x5780
f2fs_ioctl+0x89/0x3a0
__x64_sys_ioctl+0xe8/0x140
do_syscall_64+0x34/0xb0
entry_SYSCALL_64_after_hwframe+0x44/0xaeEric did some investigation on this issue, quoted from reply of Eric:
"Well, the quality of this bug report has a lot to be desired (not on upstream kernel, reproducer is full of totally irrelevant stuff, not sent to the mailing list of the filesystem whose disk image is being fuzzed, etc.). But what is going on is that f2fs_empty_dir() doesn't consider the case of a directory with an extremely large i_size on a malicious disk image.
Specifically, the reproducer mounts an f2fs image with a directory that has an i_size of 14814520042850357248, then calls FS_IOC_SET_ENCRYPTION_POLICY on it.
That results in a call to f2fs_empty_dir() to check whether the directory is empty. f2fs_empty_dir() then iterates through all 3616826182336513 blocks the directory allegedly contains to check whether any contain anything. i_rwsem is held during this, so anything else that tries to take it will hang."
In order to solve this issue, let's use f2fs_get_next_page_offset() to speed up iteration by skipping holes for all below functions:
- f2fs_empty_dir
- f2fs_readdir
- find_in_level
The way why we can speed up iteration was described in 'commit 3cf4574705b4 ("f2fs: introduce get_next_page_offset to speed up SEEK_DATA")'.
Meanwhile, in f2fs_empty_dir(), let's use f2fs_find_data_page() instead f2fs_get_lock_data_page(), due to i_rwsem was held in caller of f2fs_empty_dir(), there shouldn't be any races, so it's fine to not lock dentry page during lookuping dirents in the page.
--- fs/f2fs/data.c | 17 ++++++++++++----- fs/f2fs/dir.c | 34 ++++++++++++++++++++++++---------- fs/f2fs/f2fs.h | 5 +++-- fs/f2fs/gc.c | 4 ++-- 4 files changed, 41 insertions(+), 19 deletions(-)
fs/f2fs/data.c +12 -5
@@ -1206,7 +1206,8 @@ int f2fs_get_block(struct dnode_of_data *dn, pgoff_t index) }struct page *f2fs_get_read_data_page (struct inode *inode,pgoff_t index, -blk_opf_t op_flags,bool for_write) +blk_opf_t op_flags,bool for_write, +pgoff_t *next_pgofs) {struct address_space *mapping = inode->i_mapping;struct dnode_of_data dn;
@@ -1232,12 +1233,17 @@ struct page *f2fs_get_read_data_page(struct inode *inode, pgoff_t index,set_new_dnode (&dn, inode,NULL ,NULL ,0 ); err =f2fs_get_dnode_of_data (&dn, index,LOOKUP_NODE ); -if (err) +if (err) { +if (err == -ENOENT && next_pgofs) + *next_pgofs =f2fs_get_next_page_offset (&dn, index); goto put_err; + }f2fs_put_dnode (&dn);if (unlikely (dn.data_blkaddr ==NULL_ADDR )) { err = -ENOENT ; +if (next_pgofs) + *next_pgofs = index +1 ; goto put_err; } if (dn.data_blkaddr != NEW_ADDR &&
@@ -1281,7 +1287,8 @@ struct page *f2fs_get_read_data_page(struct inode *inode, pgoff_t index,return ERR_PTR (err); } -struct page *f2fs_find_data_page (struct inode *inode,pgoff_t index) +struct page *f2fs_find_data_page (struct inode *inode,pgoff_t index, +pgoff_t *next_pgofs) {struct address_space *mapping = inode->i_mapping;struct page *page;
@@ -1291,7 +1298,7 @@ struct page *f2fs_find_data_page(struct inode *inode, pgoff_t index)return page;f2fs_put_page (page,0 ); - page =f2fs_get_read_data_page (inode, index,0 , false); + page =f2fs_get_read_data_page (inode, index,0 , false, next_pgofs);if (IS_ERR (page))return page;
@@ -1317,7 +1324,7 @@ struct page *f2fs_get_lock_data_page(struct inode *inode, pgoff_t index,struct address_space *mapping = inode->i_mapping;struct page *page; repeat: - page =f2fs_get_read_data_page (inode, index,0 , for_write); + page =f2fs_get_read_data_page (inode, index,0 , for_write,NULL );if (IS_ERR (page))return page;
fs/f2fs/dir.c +24 -10
@@ -340,6 +340,7 @@ static struct f2fs_dir_entry *find_in_level(struct inode *dir,unsigned int bidx, end_block;struct page *dentry_page;struct f2fs_dir_entry *de =NULL ; +pgoff_t next_pgofs;bool room = false;int max_slots;
@@ -350,12 +351,13 @@ static struct f2fs_dir_entry *find_in_level(struct inode *dir, le32_to_cpu(fname->hash) % nbucket); end_block = bidx + nblock; -for (; bidx < end_block; bidx++) { +while (bidx < end_block) {/* no need to allocate new dentry pages to all the indices */ - dentry_page =f2fs_find_data_page (dir, bidx); + dentry_page =f2fs_find_data_page (dir, bidx, &next_pgofs);if (IS_ERR (dentry_page)) {if (PTR_ERR (dentry_page) == -ENOENT ) { room = true; + bidx = next_pgofs;continue ; }else { *res_page = dentry_page;
@@ -376,6 +378,8 @@ static struct f2fs_dir_entry *find_in_level(struct inode *dir,if (max_slots >= s) room = true;f2fs_put_page (dentry_page,0 ); + + bidx++; }if (!de && room &&F2FS_I (dir)->chash != fname->hash) {
@@ -956,7 +960,7 @@ void f2fs_delete_entry(struct f2fs_dir_entry *dentry, struct page *page,bool f2fs_empty_dir (struct inode *dir) { -unsigned long bidx; +unsigned long bidx =0 ;struct page *dentry_page;unsigned int bit_pos;struct f2fs_dentry_block *dentry_blk;
@@ -965,13 +969,17 @@ bool f2fs_empty_dir(struct inode *dir)if (f2fs_has_inline_dentry (dir))return f2fs_empty_inline_dir (dir); -for (bidx =0 ; bidx < nblock; bidx++) { - dentry_page =f2fs_get_lock_data_page (dir, bidx, false); +while (bidx < nblock) { +pgoff_t next_pgofs; + + dentry_page =f2fs_find_data_page (dir, bidx, &next_pgofs);if (IS_ERR (dentry_page)) { -if (PTR_ERR (dentry_page) == -ENOENT ) +if (PTR_ERR (dentry_page) == -ENOENT ) { + bidx = next_pgofs;continue ; -else + }else {return false; + } } dentry_blk =page_address (dentry_page);
@@ -983,10 +991,12 @@ bool f2fs_empty_dir(struct inode *dir)NR_DENTRY_IN_BLOCK , bit_pos); -f2fs_put_page (dentry_page,1 ); +f2fs_put_page (dentry_page,0 );if (bit_pos <NR_DENTRY_IN_BLOCK )return false; + + bidx++; }return true; }
@@ -1104,7 +1114,8 @@ static int f2fs_readdir(struct file *file, struct dir_context *ctx) goto out_free; } -for (; n < npages; n++, ctx->pos = n *NR_DENTRY_IN_BLOCK ) { +for (; n < npages; ctx->pos = n *NR_DENTRY_IN_BLOCK ) { +pgoff_t next_pgofs;/* allow readdir() to be interrupted */ if (fatal_signal_pending (current)) {
@@ -1118,11 +1129,12 @@ static int f2fs_readdir(struct file *file, struct dir_context *ctx)page_cache_sync_readahead (inode->i_mapping, ra, file, n,min (npages - n, (pgoff_t )MAX_DIR_RA_PAGES )); - dentry_page =f2fs_find_data_page (inode, n); + dentry_page =f2fs_find_data_page (inode, n, &next_pgofs);if (IS_ERR (dentry_page)) { err =PTR_ERR (dentry_page);if (err == -ENOENT ) { err =0 ; + n = next_pgofs;continue ; }else { goto out_free;
@@ -1141,6 +1153,8 @@ static int f2fs_readdir(struct file *file, struct dir_context *ctx) }f2fs_put_page (dentry_page,0 ); + + n++; } out_free:fscrypt_fname_free_buffer (&fstr);
fs/f2fs/f2fs.h +3 -2
@@ -3820,8 +3820,9 @@ int f2fs_reserve_new_block(struct dnode_of_data *dn);int f2fs_get_block (struct dnode_of_data *dn,pgoff_t index);int f2fs_reserve_block (struct dnode_of_data *dn,pgoff_t index);struct page *f2fs_get_read_data_page (struct inode *inode,pgoff_t index, -blk_opf_t op_flags,bool for_write); -struct page *f2fs_find_data_page (struct inode *inode,pgoff_t index); +blk_opf_t op_flags,bool for_write,pgoff_t *next_pgofs); +struct page *f2fs_find_data_page (struct inode *inode,pgoff_t index, +pgoff_t *next_pgofs);struct page *f2fs_get_lock_data_page (struct inode *inode,pgoff_t index,bool for_write); struct page *f2fs_get_new_data_page(struct inode *inode,
fs/f2fs/gc.c +2 -2
@@ -1562,8 +1562,8 @@ static int gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,continue ; } - data_page =f2fs_get_read_data_page (inode, - start_bidx,REQ_RAHEAD , true); + data_page =f2fs_get_read_data_page (inode, start_bidx, +REQ_RAHEAD , true,NULL );f2fs_up_write (&F2FS_I (inode)->i_gc_rwsem[WRITE ]);if (IS_ERR (data_page)) {iput (inode);
Signature · 1 line
2.36.1
Quoted 55 lines · Chao Yu
Wei Chen reports a kernel bug as blew:
INFO: task syz-executor.0:29056 blocked for more than 143 seconds. Not tainted 5.15.0-rc5 #1 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. task:syz-executor.0 state:D stack:14632 pid:29056 ppid: 6574 flags:0x00000004 Call Trace: __schedule+0x4a1/0x1720 schedule+0x36/0xe0 rwsem_down_write_slowpath+0x322/0x7a0 fscrypt_ioctl_set_policy+0x11f/0x2a0 __f2fs_ioctl+0x1a9f/0x5780 f2fs_ioctl+0x89/0x3a0 __x64_sys_ioctl+0xe8/0x140 do_syscall_64+0x34/0xb0 entry_SYSCALL_64_after_hwframe+0x44/0xaeEric did some investigation on this issue, quoted from reply of Eric:
"Well, the quality of this bug report has a lot to be desired (not on upstream kernel, reproducer is full of totally irrelevant stuff, not sent to the mailing list of the filesystem whose disk image is being fuzzed, etc.). But what is going on is that
f2fs_empty_dir()doesn't consider the case of a directory with an extremely largei_sizeon a malicious disk image.Specifically, the reproducer mounts an f2fs image with a directory that has an
i_sizeof 14814520042850357248, then callsFS_IOC_SET_ENCRYPTION_POLICYon it.That results in a call to
f2fs_empty_dir()to check whether the directory is empty.f2fs_empty_dir()then iterates through all 3616826182336513 blocks the directory allegedly contains to check whether any contain anything.i_rwsemis held during this, so anything else that tries to take it will hang."In order to solve this issue, let's use
f2fs_get_next_page_offset()to speed up iteration by skipping holes for all below functions:
-f2fs_empty_dir
-f2fs_readdir
-find_in_levelThe way why we can speed up iteration was described in 'commit
3cf4574705b4("f2fs: introduceget_next_page_offsetto speed upSEEK_DATA")'.Meanwhile, in
f2fs_empty_dir(), let's usef2fs_find_data_page()insteadf2fs_get_lock_data_page(), due toi_rwsemwas held in caller off2fs_empty_dir(), there shouldn't be any races, so it's fine to not lock dentry page during lookuping dirents in the page.Reported-by Wei Chen <[email protected]>Cc Eric Biggers <[email protected]>Signed-off-by Chao Yu <[email protected]>
Quoted 6 lines
--- fs/f2fs/data.c | 17 ++++++++++++----- fs/f2fs/dir.c | 34 ++++++++++++++++++++++++---------- fs/f2fs/f2fs.h | 5 +++-- fs/f2fs/gc.c | 4 ++-- 4 files changed, 41 insertions(+), 19 deletions(-)
Thanks. I'm not an expert on all the details, but this patch looks good to me.
Given that it optimizes lookups and readdirs too, a better title for the patch might be something like "f2fs: optimize iteration over sparse directories".
- Eric
Chao Yu
maintainerQuoted 63 lines · Eric Biggers
Quoted 55 lines · Chao Yu
Wei Chen reports a kernel bug as blew:
INFO: task syz-executor.0:29056 blocked for more than 143 seconds. Not tainted 5.15.0-rc5 #1 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. task:syz-executor.0 state:D stack:14632 pid:29056 ppid: 6574 flags:0x00000004 Call Trace: __schedule+0x4a1/0x1720 schedule+0x36/0xe0 rwsem_down_write_slowpath+0x322/0x7a0 fscrypt_ioctl_set_policy+0x11f/0x2a0 __f2fs_ioctl+0x1a9f/0x5780 f2fs_ioctl+0x89/0x3a0 __x64_sys_ioctl+0xe8/0x140 do_syscall_64+0x34/0xb0 entry_SYSCALL_64_after_hwframe+0x44/0xaeEric did some investigation on this issue, quoted from reply of Eric:
"Well, the quality of this bug report has a lot to be desired (not on upstream kernel, reproducer is full of totally irrelevant stuff, not sent to the mailing list of the filesystem whose disk image is being fuzzed, etc.). But what is going on is that
f2fs_empty_dir()doesn't consider the case of a directory with an extremely largei_sizeon a malicious disk image.Specifically, the reproducer mounts an f2fs image with a directory that has an
i_sizeof 14814520042850357248, then callsFS_IOC_SET_ENCRYPTION_POLICYon it.That results in a call to
f2fs_empty_dir()to check whether the directory is empty.f2fs_empty_dir()then iterates through all 3616826182336513 blocks the directory allegedly contains to check whether any contain anything.i_rwsemis held during this, so anything else that tries to take it will hang."In order to solve this issue, let's use
f2fs_get_next_page_offset()to speed up iteration by skipping holes for all below functions:
-f2fs_empty_dir
-f2fs_readdir
-find_in_levelThe way why we can speed up iteration was described in 'commit
3cf4574705b4("f2fs: introduceget_next_page_offsetto speed upSEEK_DATA")'.Meanwhile, in
f2fs_empty_dir(), let's usef2fs_find_data_page()insteadf2fs_get_lock_data_page(), due toi_rwsemwas held in caller off2fs_empty_dir(), there shouldn't be any races, so it's fine to not lock dentry page during lookuping dirents in the page.Reported-by Wei Chen <[email protected]>Cc Eric Biggers <[email protected]>Signed-off-by Chao Yu <[email protected]>Quoted 6 lines
--- fs/f2fs/data.c | 17 ++++++++++++----- fs/f2fs/dir.c | 34 ++++++++++++++++++++++++---------- fs/f2fs/f2fs.h | 5 +++-- fs/f2fs/gc.c | 4 ++-- 4 files changed, 41 insertions(+), 19 deletions(-)
Quoted 4 lines
Thanks. I'm not an expert on all the details, but this patch looks good to me.
Given that it optimizes lookups and readdirs too, a better title for the patch might be something like "f2fs: optimize iteration over sparse directories".
Yes, thanks for your suggestion, will update in v2.
Thanks,
Quoted 2 lines
- Eric