| Message ID | 1523361653-14439-1-git-send-email-ge.changwei@h3c.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Changwei, It looks like a code bug, and 'iocb' should not be freed at this place. Could this BUG reproduced easily? thanks, Jun On 2018/4/10 20:00, Changwei Ge wrote: > When -EIOCBQUEUED returns, it means that aio_complete() will be called > from dio_complete(), which is an asynchronous progress against write_iter. > Generally, IO is a very slow progress than executing instruction, but we > still can't take the risk to access a freed iocb. > > And we do face a BUG crash issue. >>From crash tool, iocb is obviously freed already. > crash> struct -x kiocb ffff881a350f5900 > struct kiocb { > ki_filp = 0xffff881a350f5a80, > ki_pos = 0x0, > ki_complete = 0x0, > private = 0x0, > ki_flags = 0x0 > } > > And the backtrace shows: > ocfs2_file_write_iter+0xcaa/0xd00 [ocfs2] > ? ocfs2_check_range_for_refcount+0x150/0x150 [ocfs2] > aio_run_iocb+0x229/0x2f0 > ? try_to_wake_up+0x380/0x380 > do_io_submit+0x291/0x540 > ? syscall_trace_leave+0xad/0x130 > SyS_io_submit+0x10/0x20 > system_call_fastpath+0x16/0x75 > > Signed-off-by: Changwei Ge <ge.changwei@h3c.com> > --- > fs/ocfs2/file.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c > index 5d1784a..1393ff2 100644 > --- a/fs/ocfs2/file.c > +++ b/fs/ocfs2/file.c > @@ -2343,7 +2343,7 @@ static ssize_t ocfs2_file_write_iter(struct kiocb *iocb, > > written = __generic_file_write_iter(iocb, from); > /* buffered aio wouldn't have proper lock coverage today */ > - BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); > + BUG_ON(written == -EIOCBQUEUED && !direct_io); > > /* > * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io > @@ -2463,7 +2463,7 @@ static ssize_t ocfs2_file_read_iter(struct kiocb *iocb, > trace_generic_file_aio_read_ret(ret); > > /* buffered aio wouldn't have proper lock coverage today */ > - BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); > + BUG_ON(ret == -EIOCBQUEUED && !direct_io); > > /* see ocfs2_file_write_iter */ > if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) { >
Hi Jun, On 2018/4/11 8:52, piaojun wrote: > Hi Changwei, > > It looks like a code bug, and 'iocb' should not be freed at this place. > Could this BUG reproduced easily? Actually, it's not easy to be reproduced since IO is much slower than CPU executing instructions. But the logic here is broken, we'd better fix this. Thanks, Changwei > > thanks, > Jun > > On 2018/4/10 20:00, Changwei Ge wrote: >> When -EIOCBQUEUED returns, it means that aio_complete() will be called >> from dio_complete(), which is an asynchronous progress against write_iter. >> Generally, IO is a very slow progress than executing instruction, but we >> still can't take the risk to access a freed iocb. >> >> And we do face a BUG crash issue. >> >From crash tool, iocb is obviously freed already. >> crash> struct -x kiocb ffff881a350f5900 >> struct kiocb { >> ki_filp = 0xffff881a350f5a80, >> ki_pos = 0x0, >> ki_complete = 0x0, >> private = 0x0, >> ki_flags = 0x0 >> } >> >> And the backtrace shows: >> ocfs2_file_write_iter+0xcaa/0xd00 [ocfs2] >> ? ocfs2_check_range_for_refcount+0x150/0x150 [ocfs2] >> aio_run_iocb+0x229/0x2f0 >> ? try_to_wake_up+0x380/0x380 >> do_io_submit+0x291/0x540 >> ? syscall_trace_leave+0xad/0x130 >> SyS_io_submit+0x10/0x20 >> system_call_fastpath+0x16/0x75 >> >> Signed-off-by: Changwei Ge <ge.changwei@h3c.com> >> --- >> fs/ocfs2/file.c | 4 ++-- >> 1 file changed, 2 insertions(+), 2 deletions(-) >> >> diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c >> index 5d1784a..1393ff2 100644 >> --- a/fs/ocfs2/file.c >> +++ b/fs/ocfs2/file.c >> @@ -2343,7 +2343,7 @@ static ssize_t ocfs2_file_write_iter(struct kiocb *iocb, >> >> written = __generic_file_write_iter(iocb, from); >> /* buffered aio wouldn't have proper lock coverage today */ >> - BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >> + BUG_ON(written == -EIOCBQUEUED && !direct_io); >> >> /* >> * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io >> @@ -2463,7 +2463,7 @@ static ssize_t ocfs2_file_read_iter(struct kiocb *iocb, >> trace_generic_file_aio_read_ret(ret); >> >> /* buffered aio wouldn't have proper lock coverage today */ >> - BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >> + BUG_ON(ret == -EIOCBQUEUED && !direct_io); >> >> /* see ocfs2_file_write_iter */ >> if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) { >> >
Hi Changwei, It seems other codes which try to access 'iocb' will also cause error, right? I think we should find the reason why 'iocb' is freed first. thanks, Jun On 2018/4/11 9:07, Changwei Ge wrote: > Hi Jun, > > On 2018/4/11 8:52, piaojun wrote: >> Hi Changwei, >> >> It looks like a code bug, and 'iocb' should not be freed at this place. >> Could this BUG reproduced easily? > > Actually, it's not easy to be reproduced since IO is much slower than CPU > executing instructions. But the logic here is broken, we'd better fix this. > > Thanks, > Changwei > >> >> thanks, >> Jun >> >> On 2018/4/10 20:00, Changwei Ge wrote: >>> When -EIOCBQUEUED returns, it means that aio_complete() will be called >>> from dio_complete(), which is an asynchronous progress against write_iter. >>> Generally, IO is a very slow progress than executing instruction, but we >>> still can't take the risk to access a freed iocb. >>> >>> And we do face a BUG crash issue. >>> >From crash tool, iocb is obviously freed already. >>> crash> struct -x kiocb ffff881a350f5900 >>> struct kiocb { >>> ki_filp = 0xffff881a350f5a80, >>> ki_pos = 0x0, >>> ki_complete = 0x0, >>> private = 0x0, >>> ki_flags = 0x0 >>> } >>> >>> And the backtrace shows: >>> ocfs2_file_write_iter+0xcaa/0xd00 [ocfs2] >>> ? ocfs2_check_range_for_refcount+0x150/0x150 [ocfs2] >>> aio_run_iocb+0x229/0x2f0 >>> ? try_to_wake_up+0x380/0x380 >>> do_io_submit+0x291/0x540 >>> ? syscall_trace_leave+0xad/0x130 >>> SyS_io_submit+0x10/0x20 >>> system_call_fastpath+0x16/0x75 >>> >>> Signed-off-by: Changwei Ge <ge.changwei@h3c.com> >>> --- >>> fs/ocfs2/file.c | 4 ++-- >>> 1 file changed, 2 insertions(+), 2 deletions(-) >>> >>> diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c >>> index 5d1784a..1393ff2 100644 >>> --- a/fs/ocfs2/file.c >>> +++ b/fs/ocfs2/file.c >>> @@ -2343,7 +2343,7 @@ static ssize_t ocfs2_file_write_iter(struct kiocb *iocb, >>> >>> written = __generic_file_write_iter(iocb, from); >>> /* buffered aio wouldn't have proper lock coverage today */ >>> - BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>> + BUG_ON(written == -EIOCBQUEUED && !direct_io); >>> >>> /* >>> * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io >>> @@ -2463,7 +2463,7 @@ static ssize_t ocfs2_file_read_iter(struct kiocb *iocb, >>> trace_generic_file_aio_read_ret(ret); >>> >>> /* buffered aio wouldn't have proper lock coverage today */ >>> - BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>> + BUG_ON(ret == -EIOCBQUEUED && !direct_io); >>> >>> /* see ocfs2_file_write_iter */ >>> if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) { >>> >> > . >
Hi Jun, On 2018/4/11 9:52, piaojun wrote: > Hi Changwei, > > It seems other codes which try to access 'iocb' will also cause error, > right? I think we should find the reason why 'iocb' is freed first. Which code snippet do you mean? Actually, I have checked most of other parts in write_iter() and read_iter(). I don't see any risk accessing freed iocb yet. The root cause is clear for this issue. iocb is freed by aio_complete(). You can refer to path: dio_complete aio_complete kiocb_free Thanks, Changwei > > thanks, > Jun > > On 2018/4/11 9:07, Changwei Ge wrote: >> Hi Jun, >> >> On 2018/4/11 8:52, piaojun wrote: >>> Hi Changwei, >>> >>> It looks like a code bug, and 'iocb' should not be freed at this place. >>> Could this BUG reproduced easily? >> >> Actually, it's not easy to be reproduced since IO is much slower than CPU >> executing instructions. But the logic here is broken, we'd better fix this. >> >> Thanks, >> Changwei >> >>> >>> thanks, >>> Jun >>> >>> On 2018/4/10 20:00, Changwei Ge wrote: >>>> When -EIOCBQUEUED returns, it means that aio_complete() will be called >>>> from dio_complete(), which is an asynchronous progress against write_iter. >>>> Generally, IO is a very slow progress than executing instruction, but we >>>> still can't take the risk to access a freed iocb. >>>> >>>> And we do face a BUG crash issue. >>>> >From crash tool, iocb is obviously freed already. >>>> crash> struct -x kiocb ffff881a350f5900 >>>> struct kiocb { >>>> ki_filp = 0xffff881a350f5a80, >>>> ki_pos = 0x0, >>>> ki_complete = 0x0, >>>> private = 0x0, >>>> ki_flags = 0x0 >>>> } >>>> >>>> And the backtrace shows: >>>> ocfs2_file_write_iter+0xcaa/0xd00 [ocfs2] >>>> ? ocfs2_check_range_for_refcount+0x150/0x150 [ocfs2] >>>> aio_run_iocb+0x229/0x2f0 >>>> ? try_to_wake_up+0x380/0x380 >>>> do_io_submit+0x291/0x540 >>>> ? syscall_trace_leave+0xad/0x130 >>>> SyS_io_submit+0x10/0x20 >>>> system_call_fastpath+0x16/0x75 >>>> >>>> Signed-off-by: Changwei Ge <ge.changwei@h3c.com> >>>> --- >>>> fs/ocfs2/file.c | 4 ++-- >>>> 1 file changed, 2 insertions(+), 2 deletions(-) >>>> >>>> diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c >>>> index 5d1784a..1393ff2 100644 >>>> --- a/fs/ocfs2/file.c >>>> +++ b/fs/ocfs2/file.c >>>> @@ -2343,7 +2343,7 @@ static ssize_t ocfs2_file_write_iter(struct kiocb *iocb, >>>> >>>> written = __generic_file_write_iter(iocb, from); >>>> /* buffered aio wouldn't have proper lock coverage today */ >>>> - BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>>> + BUG_ON(written == -EIOCBQUEUED && !direct_io); >>>> >>>> /* >>>> * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io >>>> @@ -2463,7 +2463,7 @@ static ssize_t ocfs2_file_read_iter(struct kiocb *iocb, >>>> trace_generic_file_aio_read_ret(ret); >>>> >>>> /* buffered aio wouldn't have proper lock coverage today */ >>>> - BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>>> + BUG_ON(ret == -EIOCBQUEUED && !direct_io); >>>> >>>> /* see ocfs2_file_write_iter */ >>>> if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) { >>>> >>> >> . >> >
Hi Changwei, The code change just works around the problem, but theoretically the IOCB object should not be freed before which is handled. Anyway, if we can find the root cause behind via some way (e.g. inject delay in some place), the result is more perfect. Thanks Gang >>> > Hi Jun, > > On 2018/4/11 8:52, piaojun wrote: >> Hi Changwei, >> >> It looks like a code bug, and 'iocb' should not be freed at this place. >> Could this BUG reproduced easily? > > Actually, it's not easy to be reproduced since IO is much slower than CPU > executing instructions. But the logic here is broken, we'd better fix this. > > Thanks, > Changwei > >> >> thanks, >> Jun >> >> On 2018/4/10 20:00, Changwei Ge wrote: >>> When -EIOCBQUEUED returns, it means that aio_complete() will be called >>> from dio_complete(), which is an asynchronous progress against write_iter. >>> Generally, IO is a very slow progress than executing instruction, but we >>> still can't take the risk to access a freed iocb. >>> >>> And we do face a BUG crash issue. >>> >From crash tool, iocb is obviously freed already. >>> crash> struct -x kiocb ffff881a350f5900 >>> struct kiocb { >>> ki_filp = 0xffff881a350f5a80, >>> ki_pos = 0x0, >>> ki_complete = 0x0, >>> private = 0x0, >>> ki_flags = 0x0 >>> } >>> >>> And the backtrace shows: >>> ocfs2_file_write_iter+0xcaa/0xd00 [ocfs2] >>> ? ocfs2_check_range_for_refcount+0x150/0x150 [ocfs2] >>> aio_run_iocb+0x229/0x2f0 >>> ? try_to_wake_up+0x380/0x380 >>> do_io_submit+0x291/0x540 >>> ? syscall_trace_leave+0xad/0x130 >>> SyS_io_submit+0x10/0x20 >>> system_call_fastpath+0x16/0x75 >>> >>> Signed-off-by: Changwei Ge <ge.changwei@h3c.com> >>> --- >>> fs/ocfs2/file.c | 4 ++-- >>> 1 file changed, 2 insertions(+), 2 deletions(-) >>> >>> diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c >>> index 5d1784a..1393ff2 100644 >>> --- a/fs/ocfs2/file.c >>> +++ b/fs/ocfs2/file.c >>> @@ -2343,7 +2343,7 @@ static ssize_t ocfs2_file_write_iter(struct kiocb > *iocb, >>> >>> written = __generic_file_write_iter(iocb, from); >>> /* buffered aio wouldn't have proper lock coverage today */ >>> - BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>> + BUG_ON(written == -EIOCBQUEUED && !direct_io); >>> >>> /* >>> * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io >>> @@ -2463,7 +2463,7 @@ static ssize_t ocfs2_file_read_iter(struct kiocb *iocb, >>> trace_generic_file_aio_read_ret(ret); >>> >>> /* buffered aio wouldn't have proper lock coverage today */ >>> - BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>> + BUG_ON(ret == -EIOCBQUEUED && !direct_io); >>> >>> /* see ocfs2_file_write_iter */ >>> if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) { >>> >> > > _______________________________________________ > Ocfs2-devel mailing list > Ocfs2-devel@oss.oracle.com > https://oss.oracle.com/mailman/listinfo/ocfs2-devel
Hi Gang, I don't think this patch is a workaround trick. We do face the risk using freed iocb although it is actually indeed hard to encounter, it still exists. So I propose to fix it making ocfs2 more reliable. Moreover, this patch has been kept in -mm tree for one month. Can anyone help review it with ack or nack? So I can do some improvement for it. :-) Thanks, Changwei On 04/11/2018 10:51 AM, Gang He wrote: > Hi Changwei, > > The code change just works around the problem, but theoretically the IOCB object should not be freed before which is handled. > Anyway, if we can find the root cause behind via some way (e.g. inject delay in some place), the result is more perfect. > > > Thanks > Gang > > >> Hi Jun, >> >> On 2018/4/11 8:52, piaojun wrote: >>> Hi Changwei, >>> >>> It looks like a code bug, and 'iocb' should not be freed at this place. >>> Could this BUG reproduced easily? >> Actually, it's not easy to be reproduced since IO is much slower than CPU >> executing instructions. But the logic here is broken, we'd better fix this. >> >> Thanks, >> Changwei >> >>> thanks, >>> Jun >>> >>> On 2018/4/10 20:00, Changwei Ge wrote: >>>> When -EIOCBQUEUED returns, it means that aio_complete() will be called >>>> from dio_complete(), which is an asynchronous progress against write_iter. >>>> Generally, IO is a very slow progress than executing instruction, but we >>>> still can't take the risk to access a freed iocb. >>>> >>>> And we do face a BUG crash issue. >>>> >From crash tool, iocb is obviously freed already. >>>> crash> struct -x kiocb ffff881a350f5900 >>>> struct kiocb { >>>> ki_filp = 0xffff881a350f5a80, >>>> ki_pos = 0x0, >>>> ki_complete = 0x0, >>>> private = 0x0, >>>> ki_flags = 0x0 >>>> } >>>> >>>> And the backtrace shows: >>>> ocfs2_file_write_iter+0xcaa/0xd00 [ocfs2] >>>> ? ocfs2_check_range_for_refcount+0x150/0x150 [ocfs2] >>>> aio_run_iocb+0x229/0x2f0 >>>> ? try_to_wake_up+0x380/0x380 >>>> do_io_submit+0x291/0x540 >>>> ? syscall_trace_leave+0xad/0x130 >>>> SyS_io_submit+0x10/0x20 >>>> system_call_fastpath+0x16/0x75 >>>> >>>> Signed-off-by: Changwei Ge <ge.changwei@h3c.com> >>>> --- >>>> fs/ocfs2/file.c | 4 ++-- >>>> 1 file changed, 2 insertions(+), 2 deletions(-) >>>> >>>> diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c >>>> index 5d1784a..1393ff2 100644 >>>> --- a/fs/ocfs2/file.c >>>> +++ b/fs/ocfs2/file.c >>>> @@ -2343,7 +2343,7 @@ static ssize_t ocfs2_file_write_iter(struct kiocb >> *iocb, >>>> >>>> written = __generic_file_write_iter(iocb, from); >>>> /* buffered aio wouldn't have proper lock coverage today */ >>>> - BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>>> + BUG_ON(written == -EIOCBQUEUED && !direct_io); >>>> >>>> /* >>>> * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io >>>> @@ -2463,7 +2463,7 @@ static ssize_t ocfs2_file_read_iter(struct kiocb *iocb, >>>> trace_generic_file_aio_read_ret(ret); >>>> >>>> /* buffered aio wouldn't have proper lock coverage today */ >>>> - BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>>> + BUG_ON(ret == -EIOCBQUEUED && !direct_io); >>>> >>>> /* see ocfs2_file_write_iter */ >>>> if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) { >>>> >> _______________________________________________ >> Ocfs2-devel mailing list >> Ocfs2-devel@oss.oracle.com >> https://oss.oracle.com/mailman/listinfo/ocfs2-devel > > _______________________________________________ > Ocfs2-devel mailing list > Ocfs2-devel@oss.oracle.com > https://oss.oracle.com/mailman/listinfo/ocfs2-devel
Hi Changwei, I agree with Gang that currently we still haven't figured out why iocb was freed. Though you fix won't bring any side effect, it looks like a workaround. That means, the freed iocb may still have risk in other place. Thanks, Joseph On 18/5/8 23:23, Changwei Ge wrote: > Hi Gang, > > I don't think this patch is a workaround trick. > > We do face the risk using freed iocb although it is actually indeed hard > to encounter, it still exists. > > So I propose to fix it making ocfs2 more reliable. > > > Moreover, this patch has been kept in -mm tree for one month. Can anyone > help review it with ack or nack? So I can do some improvement for it. :-) > > > Thanks, > > Changwei > > > On 04/11/2018 10:51 AM, Gang He wrote: >> Hi Changwei, >> >> The code change just works around the problem, but theoretically the IOCB object should not be freed before which is handled. >> Anyway, if we can find the root cause behind via some way (e.g. inject delay in some place), the result is more perfect. >> >> >> Thanks >> Gang >> >> >>> Hi Jun, >>> >>> On 2018/4/11 8:52, piaojun wrote: >>>> Hi Changwei, >>>> >>>> It looks like a code bug, and 'iocb' should not be freed at this place. >>>> Could this BUG reproduced easily? >>> Actually, it's not easy to be reproduced since IO is much slower than CPU >>> executing instructions. But the logic here is broken, we'd better fix this. >>> >>> Thanks, >>> Changwei >>> >>>> thanks, >>>> Jun >>>> >>>> On 2018/4/10 20:00, Changwei Ge wrote: >>>>> When -EIOCBQUEUED returns, it means that aio_complete() will be called >>>>> from dio_complete(), which is an asynchronous progress against write_iter. >>>>> Generally, IO is a very slow progress than executing instruction, but we >>>>> still can't take the risk to access a freed iocb. >>>>> >>>>> And we do face a BUG crash issue. >>>>> >From crash tool, iocb is obviously freed already. >>>>> crash> struct -x kiocb ffff881a350f5900 >>>>> struct kiocb { >>>>> ki_filp = 0xffff881a350f5a80, >>>>> ki_pos = 0x0, >>>>> ki_complete = 0x0, >>>>> private = 0x0, >>>>> ki_flags = 0x0 >>>>> } >>>>> >>>>> And the backtrace shows: >>>>> ocfs2_file_write_iter+0xcaa/0xd00 [ocfs2] >>>>> ? ocfs2_check_range_for_refcount+0x150/0x150 [ocfs2] >>>>> aio_run_iocb+0x229/0x2f0 >>>>> ? try_to_wake_up+0x380/0x380 >>>>> do_io_submit+0x291/0x540 >>>>> ? syscall_trace_leave+0xad/0x130 >>>>> SyS_io_submit+0x10/0x20 >>>>> system_call_fastpath+0x16/0x75 >>>>> >>>>> Signed-off-by: Changwei Ge <ge.changwei@h3c.com> >>>>> --- >>>>> fs/ocfs2/file.c | 4 ++-- >>>>> 1 file changed, 2 insertions(+), 2 deletions(-) >>>>> >>>>> diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c >>>>> index 5d1784a..1393ff2 100644 >>>>> --- a/fs/ocfs2/file.c >>>>> +++ b/fs/ocfs2/file.c >>>>> @@ -2343,7 +2343,7 @@ static ssize_t ocfs2_file_write_iter(struct kiocb >>> *iocb, >>>>> >>>>> written = __generic_file_write_iter(iocb, from); >>>>> /* buffered aio wouldn't have proper lock coverage today */ >>>>> - BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>>>> + BUG_ON(written == -EIOCBQUEUED && !direct_io); >>>>> >>>>> /* >>>>> * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io >>>>> @@ -2463,7 +2463,7 @@ static ssize_t ocfs2_file_read_iter(struct kiocb *iocb, >>>>> trace_generic_file_aio_read_ret(ret); >>>>> >>>>> /* buffered aio wouldn't have proper lock coverage today */ >>>>> - BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>>>> + BUG_ON(ret == -EIOCBQUEUED && !direct_io); >>>>> >>>>> /* see ocfs2_file_write_iter */ >>>>> if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) { >>>>> >>> _______________________________________________ >>> Ocfs2-devel mailing list >>> Ocfs2-devel@oss.oracle.com >>> https://oss.oracle.com/mailman/listinfo/ocfs2-devel >> >> _______________________________________________ >> Ocfs2-devel mailing list >> Ocfs2-devel@oss.oracle.com >> https://oss.oracle.com/mailman/listinfo/ocfs2-devel >
Hi Joseph, Thanks a lot for your comments. Actually, I think I have found the root cause why the iocb is freed. As __generic_file_write_iter() returns -EIOCBQUEUED, aio_complete() will be invoked asynchronously against write_iter()->ocfs2_file_write_iter(). If -EIOCBQUEUED is not returned, aio_complete() should be invoked after return from ocfs2_file_write_iter(), which is a synchronous progress. If I still don't make my point clear, please let me know. Thanks, Changwei On 2018/5/9 9:41, Joseph Qi wrote: > Hi Changwei, > > I agree with Gang that currently we still haven't figured out why iocb > was freed. Though you fix won't bring any side effect, it looks like a > workaround. > That means, the freed iocb may still have risk in other place. > > Thanks, > Joseph > > On 18/5/8 23:23, Changwei Ge wrote: >> Hi Gang, >> >> I don't think this patch is a workaround trick. >> >> We do face the risk using freed iocb although it is actually indeed hard >> to encounter, it still exists. >> >> So I propose to fix it making ocfs2 more reliable. >> >> >> Moreover, this patch has been kept in -mm tree for one month. Can anyone >> help review it with ack or nack? So I can do some improvement for it. :-) >> >> >> Thanks, >> >> Changwei >> >> >> On 04/11/2018 10:51 AM, Gang He wrote: >>> Hi Changwei, >>> >>> The code change just works around the problem, but theoretically the IOCB object should not be freed before which is handled. >>> Anyway, if we can find the root cause behind via some way (e.g. inject delay in some place), the result is more perfect. >>> >>> >>> Thanks >>> Gang >>> >>> >>>> Hi Jun, >>>> >>>> On 2018/4/11 8:52, piaojun wrote: >>>>> Hi Changwei, >>>>> >>>>> It looks like a code bug, and 'iocb' should not be freed at this place. >>>>> Could this BUG reproduced easily? >>>> Actually, it's not easy to be reproduced since IO is much slower than CPU >>>> executing instructions. But the logic here is broken, we'd better fix this. >>>> >>>> Thanks, >>>> Changwei >>>> >>>>> thanks, >>>>> Jun >>>>> >>>>> On 2018/4/10 20:00, Changwei Ge wrote: >>>>>> When -EIOCBQUEUED returns, it means that aio_complete() will be called >>>>>> from dio_complete(), which is an asynchronous progress against write_iter. >>>>>> Generally, IO is a very slow progress than executing instruction, but we >>>>>> still can't take the risk to access a freed iocb. >>>>>> >>>>>> And we do face a BUG crash issue. >>>>>> >From crash tool, iocb is obviously freed already. >>>>>> crash> struct -x kiocb ffff881a350f5900 >>>>>> struct kiocb { >>>>>> ki_filp = 0xffff881a350f5a80, >>>>>> ki_pos = 0x0, >>>>>> ki_complete = 0x0, >>>>>> private = 0x0, >>>>>> ki_flags = 0x0 >>>>>> } >>>>>> >>>>>> And the backtrace shows: >>>>>> ocfs2_file_write_iter+0xcaa/0xd00 [ocfs2] >>>>>> ? ocfs2_check_range_for_refcount+0x150/0x150 [ocfs2] >>>>>> aio_run_iocb+0x229/0x2f0 >>>>>> ? try_to_wake_up+0x380/0x380 >>>>>> do_io_submit+0x291/0x540 >>>>>> ? syscall_trace_leave+0xad/0x130 >>>>>> SyS_io_submit+0x10/0x20 >>>>>> system_call_fastpath+0x16/0x75 >>>>>> >>>>>> Signed-off-by: Changwei Ge <ge.changwei@h3c.com> >>>>>> --- >>>>>> fs/ocfs2/file.c | 4 ++-- >>>>>> 1 file changed, 2 insertions(+), 2 deletions(-) >>>>>> >>>>>> diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c >>>>>> index 5d1784a..1393ff2 100644 >>>>>> --- a/fs/ocfs2/file.c >>>>>> +++ b/fs/ocfs2/file.c >>>>>> @@ -2343,7 +2343,7 @@ static ssize_t ocfs2_file_write_iter(struct kiocb >>>> *iocb, >>>>>> >>>>>> written = __generic_file_write_iter(iocb, from); >>>>>> /* buffered aio wouldn't have proper lock coverage today */ >>>>>> - BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>>>>> + BUG_ON(written == -EIOCBQUEUED && !direct_io); >>>>>> >>>>>> /* >>>>>> * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io >>>>>> @@ -2463,7 +2463,7 @@ static ssize_t ocfs2_file_read_iter(struct kiocb *iocb, >>>>>> trace_generic_file_aio_read_ret(ret); >>>>>> >>>>>> /* buffered aio wouldn't have proper lock coverage today */ >>>>>> - BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); >>>>>> + BUG_ON(ret == -EIOCBQUEUED && !direct_io); >>>>>> >>>>>> /* see ocfs2_file_write_iter */ >>>>>> if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) { >>>>>> >>>> _______________________________________________ >>>> Ocfs2-devel mailing list >>>> Ocfs2-devel@oss.oracle.com >>>> https://oss.oracle.com/mailman/listinfo/ocfs2-devel >>> _______________________________________________ >>> Ocfs2-devel mailing list >>> Ocfs2-devel@oss.oracle.com >>> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c index 5d1784a..1393ff2 100644 --- a/fs/ocfs2/file.c +++ b/fs/ocfs2/file.c @@ -2343,7 +2343,7 @@ static ssize_t ocfs2_file_write_iter(struct kiocb *iocb, written = __generic_file_write_iter(iocb, from); /* buffered aio wouldn't have proper lock coverage today */ - BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); + BUG_ON(written == -EIOCBQUEUED && !direct_io); /* * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io @@ -2463,7 +2463,7 @@ static ssize_t ocfs2_file_read_iter(struct kiocb *iocb, trace_generic_file_aio_read_ret(ret); /* buffered aio wouldn't have proper lock coverage today */ - BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT)); + BUG_ON(ret == -EIOCBQUEUED && !direct_io); /* see ocfs2_file_write_iter */ if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {