On 8/10/20 5:37 AM, peterz@xxxxxxxxxxxxx wrote:
> On Sat, Aug 08, 2020 at 12:34:38PM -0600, Jens Axboe wrote:
>> Some callers may need to make signaling decisions based on the state
>> of the targeted task, and that can only safely be done post adding
>> the task_work to the task. Split task_work_add() into:
>>
>> __task_work_add() - adds the work item
>> __task_work_notify() - sends the notification
>>
>> No functional changes in this patch.
>
> Might be nice to mention __task_work_add() is now inline.
OK, will mention that.
>> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
>> Cc: stable@xxxxxxxxxxxxxxx # v5.7+
>> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
>> ---
>> include/linux/task_work.h | 19 ++++++++++++++++
>> kernel/task_work.c | 48 +++++++++++++++++++++------------------
>> 2 files changed, 45 insertions(+), 22 deletions(-)
>
>
>
>> +struct callback_head work_exited = {
>> + .next = NULL /* all we need is ->next == NULL */
>> +};
>
> Would it make sense to make this const ? Esp. with the thing exposed,
> sticking it in R/O memory might avoid a mistake somewhere.
That was my original intent, but that makes 'head' in task_work_run()
const as well, and cmpxchg() doesn't like that:
kernel/task_work.c: In function ‘task_work_run’:
./arch/x86/include/asm/cmpxchg.h:89:29: warning: initialization discards ‘const’ qualifier from pointer target type [-Wdiscarded-qualifiers]
89 | __typeof__(*(ptr)) __new = (new); \
| ^
./arch/x86/include/asm/cmpxchg.h:134:2: note: in expansion of macro ‘__raw_cmpxchg’
134 | __raw_cmpxchg((ptr), (old), (new), (size), LOCK_PREFIX)
| ^~~~~~~~~~~~~
./arch/x86/include/asm/cmpxchg.h:149:2: note: in expansion of macro ‘__cmpxchg’
149 | __cmpxchg(ptr, old, new, sizeof(*(ptr)))
| ^~~~~~~~~
./include/asm-generic/atomic-instrumented.h:1685:2: note: in expansion of macro ‘arch_cmpxchg’
1685 | arch_cmpxchg(__ai_ptr, __VA_ARGS__); \
| ^~~~~~~~~~~~
kernel/task_work.c:126:12: note: in expansion of macro ‘cmpxchg’
126 | } while (cmpxchg(&task->task_works, work, head) != work);
| ^~~~~~~
which is somewhat annoying. Because there's really no good reason why it
can't be const, it'll just require the changes to dig a bit deeper.
--
Jens Axboe
