Re: [f2fs-dev] EIO returned when reading files from R/O, compressed f2fs image

[Jaegeuk Kim <jaegeuk@kernel.org>]

Juhyung,

On 03/14, Juhyung Park wrote:
> Hi.
> 
> We have a production server storing some Android firmwares over a ZFS
> file-system, and we noticed some issues when extracting firmware files
> that use f2fs for Android system partitions.
> 
> This is a proprietary environment, so I cannot disclose every detail,
> so I hope you understand. I'll try to elaborate as much as I can.
> 
> The server is running Ubuntu 20.04 with Linux v5.15 (recently upgraded
> from v5.13 after noticing RO feature added on v5.14 being required).
> We have a set of scripts extracting Android firmware files. The input
> is typically the OTA zip file and after going through the script, it
> extracts every file and binary image from a given file.
> 
> So that includes extracting super (dynamic partition), ext4 system
> partitions with dedup enabled, and now, f2fs system partitions with RO
> and compression enabled.
> 
> Our script never had to deal with f2fs before as we only started
> seeing f2fs system partitions with recently released devices.
> 
> This is the f2fs mount flag after mounting with `mount -o ro
> system.raw /some/dir`:
> ro,relatime,lazytime,background_gc=on,discard,no_heap,user_xattr,inline_xattr,acl,inline_data,inline_dentry,extent_cache,mode=adaptive,active_logs=2,alloc_mode=reuse,checkpoint_merge,fsync_mode=posix,compress_algorithm=lz4,compress_log_size=2,compress_mode=fs,discard_unit=block
> 
> There are *a lot* of files in Android firmware these days, so we try
> to parallelize parts when we can.
> 
> This is a snippet of the script:
> ```
> #!/bin/bash
> <...>
> RSYNC="rsync -ahAXx --inplace --numeric-ids"
> <...>
> for val in system vendor product odm; do
>   if ! ls images/$val.raw > /dev/null 2>&1; then continue; fi
> 
>   mkdir -p fs
>   cd fs
> 
>   mkdir -p $val.mount tmp_$val
>   mount -o ro ../images/$val.raw $val.mount
> 
>   $RSYNC $val.mount/ "$DEST_PWD/fs/$val/" &
>   echo $! > $val.pid
>   disown
> 
>   cd $val.mount
>   find . -type d -exec mkdir -p "$DEST_PWD/strings/$val/"{} \;
>   find . -type d -exec mkdir -p "../tmp_$val/"{} \;
> 
>   while read file; do strings "$file" > "$DEST_PWD/strings/$val/$file"
> & done < <(find . -type f | grep -v '\.apk\|\.jar\|\.zip')
>   wait
> 
> <...>
> 
>   cd ../
>   rm -rf tmp_$val
>   cd ../
> done
> 
> wait
> <...>
> for val in system vendor product odm; do
>   if ! ls images/$val.raw > /dev/null 2>&1; then continue; fi
>   tail --pid=$(cat fs/$val.pid) -f /dev/null
>   umount fs/$val.mount
>   rmdir fs/$val.mount
>   rm -f images/$val.img images/$val.raw 2>/dev/null
> done
> ```
> 
> The offending part is:
> ```
>   $RSYNC $val.mount/ "$DEST_PWD/fs/$val/" &
>   find . -type d -exec mkdir -p "$DEST_PWD/strings/$val/"{} \;
>   find . -type d -exec mkdir -p "../tmp_$val/"{} \;
>   while read file; do strings "$file" > "$DEST_PWD/strings/$val/$file"
> & done < <(find . -type f | grep -v '\.apk\|\.jar\|\.zip')
>   wait
> ```
> 
> When that part is reached, the script forks thousands of new processes
> and starts reading from f2fs. (We simply decided to rely on Linux's
> task scheduler and didn't bother to limit the number of
> sub-processes.)
> 
> I am able to reliably cause f2fs to return EIO on some files:
> cp: error reading './system/priv-app/some_apk_1/some_apk_1.apk':
> Input/output error
> cp: error reading './system/priv-app/some_apk_2/some_apk_2.apk':
> Input/output error
> cp: error reading './system/priv-app/some_apk_3/some_apk_3.apk':
> Input/output error
> rsync: [sender] read errors mapping
> "/ssd/some_firmware.zip/fs/system.mount/system/priv-app/some_apk_1/some_apk_1.apk":
> Input/output error (5)
> rsync: [sender] read errors mapping
> "/ssd/some_firmware.zip/fs/system.mount/system/priv-app/some_apk_2/some_apk_2.apk":
> Input/output error (5)
> rsync: [sender] read errors mapping
> "/ssd/some_firmware.zip/fs/system.mount/system/priv-app/some_apk_3/some_apk_3.apk":
> Input/output error (5)
> rsync: [sender] read errors mapping
> "/ssd/some_firmware.zip/fs/system.mount/system/priv-app/some_apk_1/some_apk_1.apk":
> Input/output error (5)
> ERROR: system/priv-app/some_apk_1/some_apk_1.apk failed verification
> -- update retained.
> rsync: [sender] read errors mapping
> "/ssd/some_firmware.zip/fs/system.mount/system/priv-app/some_apk_2/some_apk_2.apk":
> Input/output error (5)
> ERROR: system/priv-app/some_apk_2/some_apk_2.apk failed verification
> -- update retained.
> rsync: [sender] read errors mapping
> "/ssd/some_firmware.zip/fs/system.mount/system/priv-app/some_apk_3/some_apk_3.apk":
> Input/output error (5)
> ERROR: system/priv-app/some_apk_3/some_apk_3.apk failed verification
> -- update retained.
> rsync error: some files/attrs were not transferred (see previous
> errors) (code 23) at main.c(1333) [sender=v3.2.3-45-ga28c4558]
> 
> The dmesg remains silent.

Hmm, could you share fsck.f2fs result with the image?

> 
> When I modify the script a little bit and force it to run in a
> single-thread (by removing &), it runs well.
> 
> I was able to confirm that it isn't a memory issue. The server has
> 50G+ of free memory, and the issue is still reliably reproducible when
> I defragment the memory by dropping caches and doing `echo 1 >
> /proc/sys/vm/compact_memory`.
> 
> I wasn't able to test any recent kernels (v5.16 or v5.17) as it's
> unsupported by ZFS. And it being a production server, I am somewhat
> limited in dabbling around the kernel.
> 
> I am planning to test a new kernel with v5.15 +
> f2fs-stable/linux-5.15.y merged. Meanwhile, if this is a new report or
> fixed with newer commits, I'd appreciate a tip.
> 
> Thanks.
> 
> 
> _______________________________________________
> Linux-f2fs-devel mailing list
> Linux-f2fs-devel@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel


_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel