Here's strace from both commands right after drop_caches. cp: https://pastebin.com/raw/GH51BjV4 tar: https://pastebin.com/raw/0wgiu7fE
It seems that cp is calling a few extra fadvise and ioctl. On Tue, Mar 15, 2022 at 7:25 PM Juhyung Park <[email protected]> wrote: > > Hi Chao, > > On Tue, Mar 15, 2022 at 5:45 PM Chao Yu <[email protected]> wrote: > > > > Hi Juhyung, > > > > On 2022/3/15 16:37, Juhyung Park wrote: > > > Hi Chao, > > > > > > I actually have never used compression on f2fs yet and this image is > > > from an unmodified firmware taken directly from the OTA server, not > > > made by me. > > > > Alright. > > > > > > > > If you can tell me how to check if the file is compressed or not, I'll > > > run it and report back. > > > > lsattr <target_file_path> |grep c > > > > You can check the flag's mean in manual of chattr. > > Ah, didn't know that lsattr shows compressed state. > It seems that every single file under that system partition is > compressed except ones with file size of < 4096, but the error only > happens on select files. > > > > > > > > > I never had this sort of issue (a lot of multiple readers causing > > > troubles) before with a regular R/W, uncompressed f2fs partition, so > > > I'm guessing it has something to do with RO or compression feature. > > > > Agreed, maybe we can repacking image to RO one or compressed one to > > check which one can cause this issue? > > Is this easy to do standalone? > > AOSP's ext4 image creation was already super complicated to do > out-of-tree with their dedup and minimizing file size (so that the > resulting image is as small as possible) thinggies going on. I think > I've seen EROFS support but I don't know how f2fs system images are > handled under AOSP. > > Also, I moved the environment to my personal laptop. It's running > v5.15.28 with the latest f2fs-stable/linux-5.15.y merged and it still > happens there. > > One weird thing that I've been able to confirm is that it seems that > multiple readers are not the problem: > #!/bin/bash > > set -v > > APK=system/app/some_app/some_app.apk > > echo 3 > /proc/sys/vm/drop_caches > tar -pcf - $APK | cat > /dev/null # Works! > cp $APK /tmp # Works! > > echo 3 > /proc/sys/vm/drop_caches > tar -pcf - $APK | cat > /dev/null # Works! > echo 3 > /proc/sys/vm/drop_caches > cp $APK /tmp # cp: error reading > 'system/app/some_app/some_app.apk': Input/output error > > echo 3 > /proc/sys/vm/drop_caches > cp $APK /tmp # cp: error reading > 'system/app/some_app/some_app.apk': Input/output error > tar -pcf - $APK | cat > /dev/null # tar: > system/app/some_app/some_app.apk: File shrank by 3375397 bytes; > padding with zeros > > echo 3 > /proc/sys/vm/drop_caches > cp $APK /tmp # cp: error reading > 'system/app/some_app/some_app.apk': Input/output error > echo 3 > /proc/sys/vm/drop_caches > tar -pcf - $APK | cat > /dev/null # Works! > > I'm guessing some system-calls that cp use causes f2fs to malfunction. > (GNU tar is smart enough to detect if the output is /dev/null, so it > skips the actual file read when it is piped directly to /dev/null, so > I added `cat` in between.) > > Hope this new info helps. > > Thanks. > > > > > Thanks, > > > > > > > > Thanks. > > > > > > On Tue, Mar 15, 2022 at 5:33 PM Chao Yu <[email protected]> wrote: > > >> > > >> On 2022/3/14 1:52, 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] > > >> > > >> Hi Juhyung, > > >> > > >> Do you enable compression on specified files? if so, do all EIOs come > > >> from compressed > > >> file? > > >> > > >> Thanks, > > >> > > >>> > > >>> The dmesg remains silent. > > >>> > > >>> 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 > > >>> [email protected] > > >>> https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel _______________________________________________ Linux-f2fs-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
