[CRIU] zdtm/static/fd failure on aarch64

[Dmitry Safonov]

2018-06-12 17:23 GMT+01:00 Adrian Reber <adrian at lisas.de>:
> On Tue, Jun 12, 2018 at 04:13:05PM +0100, Dmitry Safonov wrote:
>> 2018-06-12 7:48 GMT+01:00 Adrian Reber <adrian at lisas.de>:
>> > Output is always like this:
>> >
>> > 02:20:50.173: 11691: ERR: ../lib/lock.h:149: futex *0xffff86100000 = 0, c = 2249195520 (errno = 11 (Resource temporarily unavailable))
>> >
>> > 11691 mmap(NULL, 65536, PROT_READ|PROT_WRITE, MAP_SHARED|MAP_ANONYMOUS, -1, 0) = 0xffff86100000
>> > 11691 futex(0xffff86100000, FUTEX_WAIT, 2249195521, NULL) = -1 EAGAIN (Resource temporarily unavailable)
>> > 11691 brk(NULL)                         = 0x149f0000
>> > 11691 brk(0x14a20000)                   = 0x14a20000
>> > 11691 brk(NULL)                         = 0x14a20000
>> > 11691 openat(AT_FDCWD, "/etc/localtime", O_RDONLY|O_CLOEXEC) = 4
>> > 11691 fstat(4, {st_mode=S_IFREG|0644, st_size=3519, ...}) = 0
>> > 11691 fstat(4, {st_mode=S_IFREG|0644, st_size=3519, ...}) = 0
>> > 11691 mmap(NULL, 65536, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0xffff860f0000
>> > 11691 read(4, "TZif2\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\4\0\0\0\4\0\0\0\0"..., 8192) = 3519
>> > 11691 lseek(4, -2252, SEEK_CUR)         = 1267
>> > 11691 read(4, "TZif2\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\5\0\0\0\5\0\0\0\0"..., 8192) = 2252
>> > 11691 close(4)                          = 0
>> > 11691 munmap(0xffff860f0000, 65536)     = 0
>> > 11691 write(2, "02:20:50.173: 11691: ERR: ../lib"..., 135) = 135
>> > 11691 futex(0xffff86100000, FUTEX_WAIT, 2249195521, NULL) = -1 EAGAIN (Resource temporarily unavailable)
>> > 11691 newfstatat(AT_FDCWD, "/etc/localtime", {st_mode=S_IFREG|0644, st_size=3519, ...}, 0) = 0
>> > 11691 write(2, "02:20:50.175: 11691: ERR: ../lib"..., 135) = 135
>> >
>> > How I understand it is that atomic_inc() should return the new value but the returned value
>> > is completely different from what atomic_get() returns, right?
>>
>> Well, atomic_inc() should return the old value, AFAICS from code.
>> atomic_inc_return() should return the new value.
>>
>> I think it's not due cast we've some garbage value, it looks as if
>> we're doing 16-bit operations on 32-bit values.. But we aren't as far
>> as I can catch from arm64 asm.
>>
>> Could you try with this to see, if atomic ops work?
>
> 11:47:58.974: 15442: ERR: ../lib/lock.h:152: futex *0xffffb1d50000 = 0, c = 2983526400, +0 = 2983526400, -0 = 2983526400 (errno = 11 (Resource temporarily unavailable))
>
> Looks wrong. But looking at the lower bits of 0xffffb1d50000 I see that
> 0xb1d50000 is the same as 2983526400. So instead of the value we get the
> lower bits of the address.
>
>> Probably also worth to try to zero-init `tmp'/`result' inside ops to check
>> if the result is garbage from the stack in ops function.
>
> I already tried that, but it did not change anything.

Probably, could you try with this?
(haven't compile-tested on arm64)

-- 
             Dmitry
-------------- next part --------------
A non-text attachment was scrubbed...
Name: arm64-atomics-zdtm.diff
Type: text/x-patch
Size: 780 bytes
Desc: not available
URL: <http://lists.openvz.org/pipermail/criu/attachments/20180612/8233a38c/attachment-0001.bin>