nscd time_t size mismatch problem

[Adhemerval Zanella Netto]

On 25/10/22 22:04, Zev Weiss wrote:
> On Tue, Oct 25, 2022 at 02:13:23PM PDT, Adhemerval Zanella Netto wrote:
>>
>>
>> On 25/10/22 16:54, Zev Weiss via Libc-alpha wrote:
>>> Hello glibc devs,
>>>
>>> We've recently been seeing some misbehavior from nscd in OpenBMC.  It
>>> manifests as lots of log messages like:
>>>
>>>     disabled inotify-based monitoring for file /passwd': No such file or directory
>>>     stat failed for file /passwd'; will try again later: No such file or directory
>>>     disabled inotify-based monitoring for file /group': No such file or directory
>>>     stat failed for file /group'; will try again later: No such file or directory
>>>     disabled inotify-based monitoring for file /hosts': No such file or directory
>>>     stat failed for file /hosts'; will try again later: No such file or directory
>>>     disabled inotify-based monitoring for file /resolv.conf': No such file or directory
>>>     stat failed for file /resolv.conf'; will try again later: No such file or directory
>>>
>>> and so forth.  I initially assumed it was a configure-time --sysconfdir mixup, but after digging into it I found that it actually stems from a time_t size mismatch (this is a 32-bit ARM gnueabi target):
>>>
>>>     $ gdb -batch -ex 'pt time_t' -ex 'p sizeof(time_t)' time/time.o
>>>     type = long
>>>     $1 = 4
>>>     $ gdb -batch -ex 'pt time_t' -ex 'p sizeof(time_t)' nscd/nscd.o
>>>     type = long long
>>>     $1 = 8
>>>
>>> The confusing log messages are thus just the result of the coincidence that sizeof(long long) - sizeof(long) == strlen("/etc"), which causes the disagreement in the layout of struct traced_file to make it look like the 'fname' member just had its directory prefix chopped off.
>>>
>>> In the discussion of the bug in the OpenBMC issue tracker [0], Wayne Tung (CCed) came up with the patch below, which does seem to solve the immediate problem, but if I'm understanding things right does so by just reverting nscd to a 32-bit time_t, and so I'd expect probably wouldn't be considered the "right" fix -- however I don't presently know enough about the 32/64-bit time_t transition and ensuing compatibility concerns to know what the right fix really is.  Should nscd perhaps be using __time64_t or something instead of time_t?
>>
>> Reverting to 32 bits time_t only means that we are postponing some potential
>> failure to y2038, we really move everything to 64 bit time_t.  Could you check
>> if the following patch fix it?
>>
>> The issue is we do build nss modules with 64 time_t, however some features
>> are built on libc.so itself and in such cases we need to explicit use the
>> internal __time64_t type.
>>
>> diff --git a/nscd/nscd.h b/nscd/nscd.h
>> index 368091aef8..f15321585b 100644
>> --- a/nscd/nscd.h
>> +++ b/nscd/nscd.h
>> @@ -65,7 +65,7 @@ typedef enum
>> struct traced_file
>> {
>>   /* Tracks the last modified time of the traced file.  */
>> -  time_t mtime;
>> +  __time64_t mtime;
>>   /* Support multiple registered files per database.  */
>>   struct traced_file *next;
>>   int call_res_init;
>>
> 
> Ah, great -- after testing that out I can confirm that it appears to fix the problem.  Thanks!
> 
> Also, after sending that email I discovered that there's an existing bugzilla issue for this same problem (https://sourceware.org/bugzilla/show_bug.cgi?id=29402), so that can presumably be closed once a fix is committed.

I will send a proper path and I will check if there are any other instance of time_t
that is used internally as this.