[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Thu, 20 Jan 2022 02:19:51 +0000
From: "zhaohang (F)" <zhaohang14@...wei.com>
To: "musl@...ts.openwall.com" <musl@...ts.openwall.com>
CC: "zhangwentao (M)" <zhangwentao234@...wei.com>
Subject: 答复: Re:Re:Re: [pthread] pthread_barrier_wait invalid case
Maybe the following patch can solve this lacking issue
diff --git a/src/time/timer_create.c b/src/time/timer_create.c
index 0a29f05c2..dcd24fdcc 100644
--- a/src/time/timer_create.c
+++ b/src/time/timer_create.c
@@ -103,6 +103,10 @@ static void *start(void *arg)
union sigval val = args->sev->sigev_value;
__child_sync(&args->b);
+
+ if (self->timer_id < 0)
+ return 0;
+
for (;;) {
siginfo_t si;
while (sigwaitinfo(SIGTIMER_SET, &si) < 0);
发件人: zuotina [mailto:zuotingyang@....com]
发送时间: 2022年1月19日 22:56
收件人: musl@...ts.openwall.com
主题: [musl] Re:[musl] Re:Re: [musl] [pthread] pthread_barrier_wait invalid case
Hi Team,
Simple feedback on this issue
First, replace pthread_barrier_wait in timer_create with a custom sync function (implemented by __wait, __wake),
then the problem of panic is solved
But I still think the best way is fixing pthread_barrier_wait.
In addition, it is also the problem of the timer_create function. Continue to ask for advice.
```c
timer_create:
case SIGEV_THREAD:
r = pthread_create(&td, &attr, start, &args);
...
if (syscall(SYS_timer_create, clk, &ksev, &timerid) < 0)
timerid = -1;
```
If this syscall fails, the 'start' thread will reside permanently,
so the above only sets timerid = -1, which should not be perfect ?
```c
start:
for (;;) {
while (sigwaitinfo(SIGTIMER_SET, &si) < 0);
}
```
At 2021-12-17 22:28:14, "zuotina" <zuotingyang@....com<mailto:zuotingyang@....com>> wrote:
At 2021-12-17 02:16:07, "Rich Felker" <dalias@...c.org<mailto:dalias@...c.org>> wrote:
>On Thu, Dec 16, 2021 at 11:25:35PM +0800, zuotina wrote:
>> Hi everrone
>>
>>
>> I encountered a panic problem when using timer_create recently.
>> Although the probability is small, it still happened.
>> Finaly I found there is a problem in the code of phtread_barrier_wait,
>> and review code found that there may be problems in the following place,
>> 81 a_store(&b->_b_lock, 0);
>> 82 if (b->_b_waiters) __wake(&b->_b_lock, 1, 1);
>> If scheduling occurs between lines 81 and 82, it will be not good.
>> So I did an experiment and modified the source code of pthread_barrier_wait to verify my guess
>> ```c
>> 81 a_store(&b->_b_lock, 0);
>> /* If it is scheduled out here, when another thread executes pthread_barrier_wait again,
>> it can go through the entire function happily, that is, it will not be blocked */
>> syscall(yiled); // new add for test
>> // When the dispatch comes back, this b has been released
>> 82 if (b->_b_waiters) __wake(&b->_b_lock, 1, 1);
>> ```
>
>The intent here is that it's not possible that b has been released,
>because all waiters have to synchronize on b->_b_inst. It's possible
>there's a bug here. I'll look. What arch are you running on?
running on aarch64.
Looking forward to fix, thank you
>> Here is an example of timer_create (src/time/timer_create.c)
>> There are two threads A and B call pthread_barrier_wait.
>> The call is as follows
>> A thread: (timer_create // parent thread)
>> {
>> .....
>> // new add for test---begin
>> while(b->_b_inst == NULL) {
>> syscall(yield);
>> }
>> // new add for test---end
>> pthread_barrier_wait();
>> }
>> B thread: (start // child thread)
>> {
>> .....
>> // Ensure that this function is advanced to the if (!inst) {} branch of barrier_wait
>> pthread_barrier_wait();
>> }
>>
>>
>> In short, the reason for panic is that pthread_barrier_wait is not blocked as expected;
>> I hope you help to confirm whether there is a problem with the implementation
>> of pthread_barrier_wait or am I wrong?
>>
>>
>> Looking forward to your reply. Thank you.
>
>Thanks for the report.
>
>Rich
Content of type "text/html" skipped
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
