[SCM] strace branch, master, updated. v4.10-25-g8497b62

[Denys Vlasenko]

On 03/24/2015 01:49 AM, Dmitry V. Levin wrote:
> On Mon, Mar 23, 2015 at 05:50:39AM +0300, Dmitry V. Levin wrote:
>> On Sat, Mar 21, 2015 at 05:06:32PM +0000, Denys Vlasenko wrote:
>> [...]
>>> commit 8497b6222ed8ef606996d0ceb2bae260d82f95e2
>>> Author: Denys Vlasenko <dvlasenk at redhat.com>
>>> Date:   Sat Mar 21 17:51:52 2015 +0100
>>>
>>>     Show the syscall name in "resuming interrupted call" message
>>
>> It's nice when it works, but you've broken ARM and probably other
>> architectures:
>>
>> $ ./strace true
>> pid 1234 stray syscall exit
>> Segmentation fault
> 
> SPARC has also suffered from this change:
> 
> sparc$ ./strace -e, true
> syscall: unknown syscall trap 1a80000d f7aceb48
> +++ exited with 0 +++
> 
> Not as fatal as on ARM, but enough to trigger bexecve.test failure.

sparc seems to be one of a few architectures needing to examine
syscall insn to determine personality.

We seem to be carrying a lot of non-Linux cruft in this particular
corner.

Linux kernel does not support Solaris syscalls:
0x91d02008 is a "ta 0x8" instruction, and Linux kernel
has no syscall trap gate set up for it. It won't work
if a (presumably Solaris) userspace binary would try executing "ta 0x8".

Same for 0x91d02027.

So this is bogus too:

if (trap == 0x91d02027) {
       scno = 156;
}

The gist is, only 0x91d02010 and 0x91d0206d are valid -
those are 32-bit and 64-bit Linux syscall entry points.
We need to check trap insn only on SPARC64,
so for 32-bit strace we can save one round-trip to kernel.


And this is bogus too:

scno = sparc_regs.u_regs[U_REG_G1];
if (scno == 0) {
       scno = sparc_regs.u_regs[U_REG_O0];
       ...


because Linux kernel _always_ takes syscall# from %g1,
zero value is no exception.


I think entire personality 1 for SPARC is dead (doesn't work
for many years) - all solaris_FOO()'s are printargs,
except for lonely solaris_open().


> BTW, this "resuming interrupted system call" feature is broken on SPARC:
> 
> $ sleep 2 & sleep 1; ./strace -p $!
> [1] 24450
> Process 24450 attached
> syscall: unknown syscall trap 1a800004 f7ebea10
> syscall_516(0, 0x14980, 0x117e0, 0x1, 0x2, 0xffaef328) = 0
> close(1)                                = 0
> close(2)                                = 0
> exit_group(0)                           = ?
> +++ exited with 0 +++
> [1]+  Done                    sleep 2
> 
> ERESTART_RESTARTBLOCK is 516; what a coincidence! :)

That's the result of bogus "if (scno == 0)" check.
restart_syscall is syscall#0. We mishandle it.

I'm going to push a run-tested fix.