[Date Prev][Date Next]
[Chronological]
[Thread]
[Top]
Re: (ITS#8784) SIGBUS in mdb_page_touch
balaret@gmail.com wrote:
> Full_Name: Sergey Z
> Version: LMDB_0.9.19
> OS: Android
> URL: ftp://ftp.openldap.org/incoming/
> Submission from: (NULL) (2620:119:5001:2242:9215:2763:ff1b:ae35)
>
>
> Hey guys,
>
> We are using LMDB 0.9.19 in our Android project and sometimes we are getting
> SIGBUS in mdb_page_touch(). We can't reproduce this issue on our side but we
> have plenty of crash reports from our users (about 400 daily):
>
> SIGBUS
> libLMDBAndroid.so.mdb_page_touch ( mdb .c :2412)
> libLMDBAndroid.so.mdb_page_search ( mdb .c :5610)
> libLMDBAndroid.so.mdb_freelist_save ( mdb .c :3128)
> libLMDBAndroid.so.mdb_txn_commit ( mdb .c :3606)
>
> This is probably a platform specific issue because 95% of crashes happened on
> Android 7.0.
>
> I would greatly appreciate if you help to shed a light on this - any ideas what
> might goes wrong or what might cause such an issue.
I've encountered this as well. I believe there's a bug in the Android FUSE
filesystem driver. Every time I've analyzed one of these crashes in the
debugger, the relevant addresses are perfectly valid, which leads me to
believe there's a race condition in their page fault handler. I.e., eventually
the handler returns a valid memory page but the application has been killed
before the handler completes. Then, by the time the debugger gets control, all
of memory looks valid.
If you root the device and mount the storage partition directly, bypassing the
Android FUSE filesystem, you'll find that these crashes all disappear - even
if using the same storage device as before. Which again points to a bug in
their FUSE filesystem driver. But I haven't been able to pinpoint the bug in
their FUSE driver source code yet. I suggest you focus your debugging efforts
there.
>
> Thank you,
> Sergey
>
>
--
-- Howard Chu
CTO, Symas Corp. http://www.symas.com
Director, Highland Sun http://highlandsun.com/hyc/
Chief Architect, OpenLDAP http://www.openldap.org/project/