| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <d35d5ad7-5622-fbf5-5853-e541f8c26670@huawei.com>
Date: Thu, 1 Aug 2019 15:08:39 +0800
From: Chao Yu <yuchao0@...wei.com>
To: Jaegeuk Kim <jaegeuk@...nel.org>
CC: <linux-f2fs-devel@...ts.sourceforge.net>,
<linux-kernel@...r.kernel.org>, <chao@...nel.org>
Subject: Re: [PATCH v2] f2fs: separate NOCoW and pinfile semantics
On 2019/8/1 12:14, Jaegeuk Kim wrote:
> On 07/31, Chao Yu wrote:
>> On 2019/7/31 2:02, Jaegeuk Kim wrote:
>>> On 07/29, Chao Yu wrote:
>>>> On 2019/7/29 13:57, Jaegeuk Kim wrote:
>>>>> On 07/23, Chao Yu wrote:
>>>>>> On 2019/7/23 10:36, Jaegeuk Kim wrote:
>>>>>>> On 07/19, Chao Yu wrote:
>>>>>>>> Pinning a file is heavy, because skipping pinned files make GC
>>>>>>>> running with heavy load or no effect.
>>>>>>>
>>>>>>> Pinned file is a part of NOCOW files, so I don't think we can simply drop it
>>>>>>> for backward compatibility.
>>>>>>
>>>>>> Yes,
>>>>>>
>>>>>> But what I concerned is that pin file is too heavy, so in order to satisfy below
>>>>>> demand, how about introducing pin_file_2 flag to triggering IPU only during
>>>>>> flush/writeback.
>>>>>
>>>>> That can be done by cold files?
>>>>
>>>> Then it may inherit property of cold type file, e.g. a) goes into cold area; b)
>>>> update with very low frequency.
>>>>
>>>> Actually pin_file_2 could be used by db-wal/log file, which are updated
>>>> frequently, and should go to hot/warm area, it does not match above two property.
>>>
>>> How about considering another name like "IPU-only mode"?
>>>
>>> fallocate write Flag GC
>>> Pin_file: preallocate IPU FS_NOCOW_FL Not allowed
>>> IPU_file: Not preallocate IPU N/A Default by temperature
>>
>> One question, do we need preallocate physical block address for IPU_file as
>> Pin_file? since it can enhance db file's sequential read performance, not sure,
>> db can handle random data in preallocated blocks.
>
> db file will do atomic writes, which can not be used with this. -wal may be able
Now WAL mode were set by default in Android, so most of db file are -wal type now.
> to preallocate blocks, but it can eat disk space unnecessarily.
I meant .db-wal file rather than .db.
Yes, that's ext4 style, that would bring better performance due to less holes in
block distribution.
I don't think we need to worry about space issue for db-wal file. I tracked
.db-wal file's update before:
- there are very frequently truncation and deletion, that means the preallocated
blocks won't exist for long time.
- and also there are very frequently append writes, I suspect there almost very
few preallocate block are not written.
- total db-wal file number is less.
>
>>
>> Other behaviors looks good to me. :)
>>
>> I plan to use last bit in inode.i_inline to store this flag.
>
> Why not using i_flag like FS_NOCOW_FL?
Oops, as you listed in last email, I can see you don't want to break
FS_NOCOW_FL's semantics for backward compatibility.
Flag
IPU_file N/A
If we plan to use FS_NOCOW_FL, that's what this patch has already did, you can
merge it directly... :P
>
>>
>>> Cold_file: Not preallocate IPU N/A Move in cold area
>>> Hot_file: Not preallocate IPU/OPU N/A Move in hot area
>>
>> Should hot file be gced to hot area? That would mix new hot data with old 'hot'
>> data which actually become cold.
>
> But, user explicitly specified this is hot.
With current implementation, GC will migrate data from hot/warm/cold area to
cold area.
Thanks,
>
>>
>> Thanks,
>>
>>>
>>>>
>>>> Thank,
>>>>
>>>>>
>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>> So that this patch propose to separate nocow and pinfile semantics:
>>>>>>>> - NOCoW flag can only be set on regular file.
>>>>>>>> - NOCoW file will only trigger IPU at common writeback/flush.
>>>>>>>> - NOCow file will do OPU during GC.
>>>>>>>>
>>>>>>>> For the demand of 1) avoid fragment of file's physical block and
>>>>>>>> 2) userspace don't care about file's specific physical address,
>>>>>>>> tagging file as NOCoW will be cheaper than pinned one.
>>>>>>
>>>>>> ^^^
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>>>>
>>>>>>>> Signed-off-by: Chao Yu <yuchao0@...wei.com>
>>>>>>>> ---
>>>>>>>> v2:
>>>>>>>> - rebase code to fix compile error.
>>>>>>>> fs/f2fs/data.c | 3 ++-
>>>>>>>> fs/f2fs/f2fs.h | 1 +
>>>>>>>> fs/f2fs/file.c | 22 +++++++++++++++++++---
>>>>>>>> 3 files changed, 22 insertions(+), 4 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
>>>>>>>> index a2a28bb269bf..15fb8954c363 100644
>>>>>>>> --- a/fs/f2fs/data.c
>>>>>>>> +++ b/fs/f2fs/data.c
>>>>>>>> @@ -1884,7 +1884,8 @@ static inline bool check_inplace_update_policy(struct inode *inode,
>>>>>>>>
>>>>>>>> bool f2fs_should_update_inplace(struct inode *inode, struct f2fs_io_info *fio)
>>>>>>>> {
>>>>>>>> - if (f2fs_is_pinned_file(inode))
>>>>>>>> + if (f2fs_is_pinned_file(inode) ||
>>>>>>>> + F2FS_I(inode)->i_flags & F2FS_NOCOW_FL)
>>>>>>>> return true;
>>>>>>>>
>>>>>>>> /* if this is cold file, we should overwrite to avoid fragmentation */
>>>>>>>> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
>>>>>>>> index 596ab3e1dd7b..f6c5a3d2e659 100644
>>>>>>>> --- a/fs/f2fs/f2fs.h
>>>>>>>> +++ b/fs/f2fs/f2fs.h
>>>>>>>> @@ -2374,6 +2374,7 @@ static inline void f2fs_change_bit(unsigned int nr, char *addr)
>>>>>>>> #define F2FS_NOATIME_FL 0x00000080 /* do not update atime */
>>>>>>>> #define F2FS_INDEX_FL 0x00001000 /* hash-indexed directory */
>>>>>>>> #define F2FS_DIRSYNC_FL 0x00010000 /* dirsync behaviour (directories only) */
>>>>>>>> +#define F2FS_NOCOW_FL 0x00800000 /* Do not cow file */
>>>>>>>> #define F2FS_PROJINHERIT_FL 0x20000000 /* Create with parents projid */
>>>>>>>>
>>>>>>>> /* Flags that should be inherited by new inodes from their parent. */
>>>>>>>> diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
>>>>>>>> index 7ca545874060..ae0fec54cac6 100644
>>>>>>>> --- a/fs/f2fs/file.c
>>>>>>>> +++ b/fs/f2fs/file.c
>>>>>>>> @@ -1692,6 +1692,7 @@ static const struct {
>>>>>>>> { F2FS_NOATIME_FL, FS_NOATIME_FL },
>>>>>>>> { F2FS_INDEX_FL, FS_INDEX_FL },
>>>>>>>> { F2FS_DIRSYNC_FL, FS_DIRSYNC_FL },
>>>>>>>> + { F2FS_NOCOW_FL, FS_NOCOW_FL },
>>>>>>>> { F2FS_PROJINHERIT_FL, FS_PROJINHERIT_FL },
>>>>>>>> };
>>>>>>>>
>>>>>>>> @@ -1715,7 +1716,8 @@ static const struct {
>>>>>>>> FS_NODUMP_FL | \
>>>>>>>> FS_NOATIME_FL | \
>>>>>>>> FS_DIRSYNC_FL | \
>>>>>>>> - FS_PROJINHERIT_FL)
>>>>>>>> + FS_PROJINHERIT_FL | \
>>>>>>>> + FS_NOCOW_FL)
>>>>>>>>
>>>>>>>> /* Convert f2fs on-disk i_flags to FS_IOC_{GET,SET}FLAGS flags */
>>>>>>>> static inline u32 f2fs_iflags_to_fsflags(u32 iflags)
>>>>>>>> @@ -1753,8 +1755,6 @@ static int f2fs_ioc_getflags(struct file *filp, unsigned long arg)
>>>>>>>> fsflags |= FS_ENCRYPT_FL;
>>>>>>>> if (f2fs_has_inline_data(inode) || f2fs_has_inline_dentry(inode))
>>>>>>>> fsflags |= FS_INLINE_DATA_FL;
>>>>>>>> - if (is_inode_flag_set(inode, FI_PIN_FILE))
>>>>>>>> - fsflags |= FS_NOCOW_FL;
>>>>>>>>
>>>>>>>> fsflags &= F2FS_GETTABLE_FS_FL;
>>>>>>>>
>>>>>>>> @@ -1794,6 +1794,22 @@ static int f2fs_ioc_setflags(struct file *filp, unsigned long arg)
>>>>>>>> if (ret)
>>>>>>>> goto out;
>>>>>>>>
>>>>>>>> + if ((fsflags ^ old_fsflags) & FS_NOCOW_FL) {
>>>>>>>> + if (!S_ISREG(inode->i_mode)) {
>>>>>>>> + ret = -EINVAL;
>>>>>>>> + goto out;
>>>>>>>> + }
>>>>>>>> +
>>>>>>>> + if (f2fs_should_update_outplace(inode, NULL)) {
>>>>>>>> + ret = -EINVAL;
>>>>>>>> + goto out;
>>>>>>>> + }
>>>>>>>> +
>>>>>>>> + ret = f2fs_convert_inline_inode(inode);
>>>>>>>> + if (ret)
>>>>>>>> + goto out;
>>>>>>>> + }
>>>>>>>> +
>>>>>>>> ret = f2fs_setflags_common(inode, iflags,
>>>>>>>> f2fs_fsflags_to_iflags(F2FS_SETTABLE_FS_FL));
>>>>>>>> out:
>>>>>>>> --
>>>>>>>> 2.18.0.rc1
>>>>>>> .
>>>>>>>
>>>>> .
>>>>>
>>> .
>>>
> .
>
Powered by blists - more mailing lists