Messages in this thread | | | Subject | Re: [PATCH v6] f2fs: introduce /sys/fs/f2fs/<disk>/fsck_stack node | From | Chao Yu <> | Date | Thu, 19 Aug 2021 16:04:15 +0800 |
| |
On 2021/8/16 15:23, 李扬韬 wrote: >>> HI Chao, >>>>> SBI_NEED_FSCK is an indicator that fsck.f2fs needs to be triggered, >>>>> this flag is set in too many places. For some scenes that are not very >>>>> reproducible, adding stack information will help locate the problem. >>>>> >>>>> Let's record all fsck stack history, I added F2FS_FSCK_STACK_TRACE >>>>> configuration options and sysfs nodes. After opening the configuration >>>>> options and enabling the node, it will start recording. The recorded >>>>> stack information will not be clear, and we can get information form >>>>> kernel log. >>>>> >>>>> Signed-off-by: Yangtao Li <frank.li@vivo.com> >>>>> --- >>>>> Documentation/ABI/testing/sysfs-fs-f2fs | 7 ++++ >>>>> fs/f2fs/Kconfig | 10 ++++++ >>>>> fs/f2fs/f2fs.h | 45 +++++++++++++++++++++++++ >>>>> fs/f2fs/sysfs.c | 27 +++++++++++++++ >>>>> 4 files changed, 89 insertions(+) >>>>> >>>>> diff --git a/Documentation/ABI/testing/sysfs-fs-f2fs b/Documentation/ABI/testing/sysfs-fs-f2fs >>>>> index ef4b9218ae1e..047c398093cf 100644 >>>>> --- a/Documentation/ABI/testing/sysfs-fs-f2fs >>>>> +++ b/Documentation/ABI/testing/sysfs-fs-f2fs >>>>> @@ -493,3 +493,10 @@ Contact: "Chao Yu" <yuchao0@huawei.com> >>>>> Description: When ATGC is on, it controls age threshold to bypass GCing young >>>>> candidates whose age is not beyond the threshold, by default it was >>>>> initialized as 604800 seconds (equals to 7 days). >>>>> + >>>>> +What: /sys/fs/f2fs/<disk>/fsck_stack >>>>> +Date: August 2021 >>>>> +Contact: "Yangtao Li" <frank.li@vivo.com> >>>>> +Description: Controls to enable/disable fsck stack trace, you can get stack >>>>> + information from kernel log. Note that the recorded stack information >>>>> + will not be cleared. >>>> >>>> Again, please don't add this into sysfs. >> >>Oh, I missed to check the details... >> >>> >>> I added this node, part of the idea is to trigger the export of stack information. >>> There is no information transmitted through sysfs here, but the record of the stack is switched on and off. >>> If don't export this information through procfs and sysfs, is there a more appropriate way? >> >>Well, I doubt why we should export stack info via proc/sysfs node or >>sysfs switch. >> >>Those info will always be needed to troubleshoot issues no matter in >>user or eng version of Android, can we just print them directly into >>kernel message... what I concern is we may lost the bug scene due to >>no one can help to trigger dmesg printing via sysfs. >> >>Jaegeuk, thoughts? >> >>> >>>> >>>>> diff --git a/fs/f2fs/Kconfig b/fs/f2fs/Kconfig >>>>> index 7669de7b49ce..f451e567e4a8 100644 >>>>> --- a/fs/f2fs/Kconfig >>>>> +++ b/fs/f2fs/Kconfig >>>>> @@ -135,3 +135,13 @@ config F2FS_FS_LZORLE >>>>> default y >>>>> help >>>>> Support LZO-RLE compress algorithm, if unsure, say Y. >>>>> + >>>>> +config F2FS_FSCK_STACK_TRACE >>>> >>>> I don't think we need another config to wrap this functionality, may be we >>>> can use F2FS_CHECK_FS instead. >>> >>> OK. >>> >>>> >>>>> + bool "F2FS fsck stack information record" >>>>> + depends on F2FS_FS >>>>> + depends on STACKDEPOT >>>>> + default y >>>>> + help >>>>> + Support printing out fsck stack history. With this, you have to >>>>> + turn on "fsck_stack" sysfs node. Then you can get information >>>>> + from kernel log. >>>>> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h >>>>> index ee8eb33e2c25..b2d1d1a5a3fc 100644 >>>>> --- a/fs/f2fs/f2fs.h >>>>> +++ b/fs/f2fs/f2fs.h >>>>> @@ -24,6 +24,8 @@ >>>>> #include <linux/quotaops.h> >>>>> #include <linux/part_stat.h> >>>>> #include <crypto/hash.h> >>>>> +#include <linux/stackdepot.h> >>>>> +#include <linux/stacktrace.h> >>>>> >>>>> #include <linux/fscrypt.h> >>>>> #include <linux/fsverity.h> >>>>> @@ -117,6 +119,8 @@ typedef u32 nid_t; >>>>> >>>>> #define COMPRESS_EXT_NUM 16 >>>>> >>>>> +#define FSCK_STACK_DEPTH 64 >>>> >>>> 16? >>> >>> OK. >>> >>>> >>>>> + >>>>> struct f2fs_mount_info { >>>>> unsigned int opt; >>>>> int write_io_size_bits; /* Write IO size bits */ >>>>> @@ -1748,6 +1752,11 @@ struct f2fs_sb_info { >>>>> unsigned int compress_watermark; /* cache page watermark */ >>>>> atomic_t compress_page_hit; /* cache hit count */ >>>>> #endif >>>>> +#ifdef CONFIG_F2FS_FSCK_STACK_TRACE >>>>> + depot_stack_handle_t *fsck_stack_history; >>>>> + unsigned int fsck_count; >>>>> + bool fsck_stack; >>>> >>>> IMO, all bug_on()s are corner cases, and catching those stacks won't cost >>>> much, so we can just use CONFIG_XXX to enable/disable this feature. >>> >>> F2FS_CHECK_FS ? >>> >>>> >>>>> +#endif >>>>> }; >>>>> >>>>> struct f2fs_private_dio { >>>>> @@ -1954,6 +1963,38 @@ static inline struct address_space *NODE_MAPPING(struct f2fs_sb_info *sbi) >>>>> return sbi->node_inode->i_mapping; >>>>> } >>>>> >>>>> +#ifdef CONFIG_F2FS_FSCK_STACK_TRACE >>>>> +static void fsck_stack_trace(struct f2fs_sb_info *sbi) >>>>> +{ >>>>> + unsigned long entries[FSCK_STACK_DEPTH]; >>>>> + depot_stack_handle_t stack, *new; >>>>> + unsigned int nr_entries; >>>>> + int i; >>>>> + >>>>> + if (!sbi->fsck_stack) >>>>> + return; >>>>> + >>>>> + nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0); >>>>> + nr_entries = filter_irq_stacks(entries, nr_entries); >>>>> + stack = stack_depot_save(entries, nr_entries, GFP_KERNEL); >>>>> + if (!stack) >>>>> + return; >>>>> + >>>>> + /* Try to find an existing entry for this backtrace */ >>>>> + for (i = 0; i < sbi->fsck_count; i++) >>>>> + if (sbi->fsck_stack_history[i] == stack) >>>>> + return; >>>>> + >>>>> + new = krealloc(sbi->fsck_stack_history, (sbi->fsck_count + 1) * >>>>> + sizeof(*sbi->fsck_stack_history), GFP_KERNEL); >>>>> + if (!new) >>>>> + return; >>>>> + >>>>> + sbi->fsck_stack_history = new; >>>>> + sbi->fsck_stack_history[sbi->fsck_count++] = stack; >>>> >>>> It will case memory leak after f2fs module exits. >>> >>> So let's enable this feature when f2fs is not a module and enable F2FS_CHECK_FS. >> >>I mean it needs to free .fsck_stack_history during umount(). > > > > So this is a double-edged sword, and the person who uses it must be aware of this issue. > The release seems unrealistic under the current circumstances. This is also the reason
You mean the handles allocated from stack_depot_save() should not be released, right?
> why I add a switch, instead of enabling it by default after opening the configuration.
Well, I still didn't get why we can not release the memory allocated by krealloc().
Thanks,
> > Yangtao > >
| |