Re: [PATCH] Support single byte reads from integers published in procfs by kernel/sysctl.c

[ebiederm@xmission.com (Eric W. Biederman)]

Earl Chew <echew@ixiacom.com> writes:

> On 23/01/2012 6:47 AM, Eric W. Biederman wrote:
>> The difference between strings and integers is that with strings
>> single byte reads/writes actually make sense.
>> 
>> Single decimal digits reads don't make much sense at all for integers.
>> 
>> Is there any compelling reason to support this complete idiocy from user
>> space programs?
>
> Eric,
>
> [ I also note that the patch allows multiple short reads (not necessarily
>   just a single byte at a time) from these entries in procfs. ]
>
> I think there is a reasonable userspace expectation that entities
> that present themselves as text files to produce results that are
> consistent with the userspace model of a text file.
>
> For a more concrete example, consider the following BusyBox ash(1) code:
>
> read X < /proc/sys/kernel/threads-max
> Y=$(cat /proc/sys/kernel/threads-max)
>
> read A < /proc/sys/kernel/core_pattern
> B=$(cat /proc/sys/kernel/core_pattern)


> The fact that these yield different results is surprising given that
> the user sees these as text files.
>
> I think it is worthwhile asking the opposite question,
>
> " Is there any compelling reason for the kernel to require userspace
>   programs to access /proc/sys/kernel/threads-max any differently
>   from, say, /etc/fstab ? "
>
> If the answer is in the affirmative, then it requires some layer in
> the userspace program (presumably the IO layer) to make some
> special accommodation for some entries in /proc.
>
> As can be seen in the case of BusyBox, this approach might be difficult
> to swallow in something as generic as ash(1), so it is left to the user to
> puzzle out why read fails, but cat works -- but only sometimes.

read X < file seems like a reasonable case to support.
Bash doesn't have that problem so presumably BusyBox is simply
inefficient.

If you are interested in fixing this properly with a tiny buffer
reachable from struct file I think this can be worth fixing.  I think
this is doable by using seq_file in proc_sys_read.

Rereading different bytes of the integer multiple times when the integer
may be changing does not seem like a reasonable implementation.

Eric