crypto/af_alg02: thread termination fixes
PATCH v2 1/1 crypto/af_alg02: thread termination fixes
On musl, pthread_kill() doesn't return ESRCH if thread id is not found (POSIX only recommends to return ESRCH). Use tst_atomic_store/load() instead, when waiting for the thread.
Also, the thread's resources wasn't properly freed after the run(), so adding pthread_join() should fix that.
---
v2: convert pthread_tryjoin_np() into atomic_load/store() + pthread_join()
testcases/kernel/crypto/af_alg02.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-)
testcases/kernel/crypto/af_alg02.c +7 -2
@@ -18,11 +18,13 @@#include "tst_test.h" #include "tst_af_alg.h" #include "tst_safe_pthread.h" +#include "tst_atomic.h" #include <pthread.h> #include <errno.h> #define SALSA20_IV_SIZE 8#define SALSA20_MIN_KEY_SIZE 16 +static int completed;static void *verify_encrypt (void *arg) {
@@ -48,19 +50,21 @@ static void *verify_encrypt(void *arg)tst_res (TPASS ,"Successfully \"encrypted\" an empty message" ); else tst_res(TFAIL, "read() didn't return 0"); + +tst_atomic_store (1 , &completed);return arg; }static void run (void ) {pthread_t thr; - +tst_atomic_store (0 , &completed);pthread_setcancelstate (PTHREAD_CANCEL_ENABLE ,NULL );SAFE_PTHREAD_CREATE (&thr,NULL , verify_encrypt,NULL );TST_CHECKPOINT_WAIT (0 ); -while (pthread_kill (thr,0 ) !=ESRCH ) { +while (!tst_atomic_load (&completed)) {if (tst_timeout_remaining () <=10 ) {pthread_cancel (thr); tst_brk(TBROK,
@@ -68,6 +72,7 @@ static void run(void) }usleep (1000 ); } +pthread_join (thr,NULL ); }static struct tst_test test = {
Signature · 1 line
2.25.1
PATCH v2 1/1 crypto/af_alg02: thread termination fixes
Hi!
Quoted 6 lines
On musl,
pthread_kill()doesn't return ESRCH if thread id is not found (POSIX only recommends to return ESRCH). Use tst_atomic_store/load() instead, when waiting for the thread.Also, the thread's resources wasn't properly freed after the
run(), so addingpthread_join()should fix that.
I do not think that we even need atomic operations here as we do not have competing threads setting the value, it should work fine with regular assignments as long as the completed variable is marked as volatile (which will prevent compiler mis-optimizations).
Signature · 2 lines
Cyril Hrubis [email protected]
PATCH v2 1/1 crypto/af_alg02: thread termination fixes
Hi
Quoted 11 lines · Cyril Hrubis
Hi!
Quoted 6 lines
On musl,
pthread_kill()doesn't return ESRCH if thread id is not found (POSIX only recommends to return ESRCH). Use tst_atomic_store/load() instead, when waiting for the thread.Also, the thread's resources wasn't properly freed after the
run(), so addingpthread_join()should fix that.I do not think that we even need atomic operations here as we do not have competing threads setting the value, it should work fine with regular assignments as long as the completed variable is marked as volatile (which will prevent compiler mis-optimizations).
+1 Using a volatile variable should be enough here.
If only pthread_timedjoin_np was not _np... This is exactly the function that could be used here without any boilerplate.
But is the custom timeout handling in this test even required? Does the default timeout using SIGALRM not work?
The thread was introduced, because SIG_ALRM was apparently not able to interrupt the read call on linux < 4.14.
But even there it should be possible to interrupt pthread_join. So just replacing the whole loop with pthread_join should be enough.
Joerg
PATCH v2 1/1 crypto/af_alg02: thread termination fixes
Quoted 8 lines · Joerg Vehlow
Hi
Quoted 5 lines · Cyril Hrubis
Hi!
Quoted 4 lines
On musl,
pthread_kill()doesn't return ESRCH if thread id is not found (POSIX only recommends to return ESRCH). Use tst_atomic_store/load() instead, when waiting for the thread.
Quoted 9 lines
Quoted 6 lines
Quoted 2 lines
Also, the thread's resources wasn't properly freed after the
run(), so addingpthread_join()should fix that.I do not think that we even need atomic operations here as we do not have competing threads setting the value, it should work fine with regular assignments as long as the completed variable is marked as volatile (which will prevent compiler mis-optimizations).
+1 Using a volatile variable should be enough here.
I have no preference for atomic or volatile methods. Both should be fine.
Quoted 5 lines
If only
pthread_timedjoin_npwas not_np... This is exactly the function that could be used here without any boilerplate.
But is the custom timeout handling in this test even required? Does the default timeout using SIGALRM not work?
The default timeout obviously works.
But with introducing the thread_B (custom timeout) can get a fine-grained report when the read() was stuck. That's the advantage I can think of.
Quoted 5 lines
The thread was introduced, because
SIG_ALRMwas apparently not able to interrupt the read call on linux < 4.14.
But even there it should be possible to interruptpthread_join. So just replacing the whole loop withpthread_joinshould be enough.
That should work but no precise log to indicate where goes wrong, so I vote to go the loop way:).
Signature · 5 lines
Regards, Li Wang -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.linux.it/pipermail/ltp/attachments/20210730/ca8828e3/attachment.htm>
PATCH v2 1/1 crypto/af_alg02: thread termination fixes
Hi Li,
Quoted 49 lines · Li Wang
On Fri, Jul 30, 2021 at 5:32 PM Joerg Vehlow <[email protected] <mailto:[email protected]>> wrote:
Hi
On 7/30/2021 11:12 AM, Cyril Hrubis wrote:
> Hi!
>> On musl,pthread_kill()doesn't return ESRCH if thread id is not found
>> (POSIX only recommends to return ESRCH). Use tst_atomic_store/load()
>> instead, when waiting for the thread.
>>
>> Also, the thread's resources wasn't properly freed after therun(), >> so addingpthread_join()should fix that.
> I do not think that we even need atomic operations here as we do not > have competing threads setting the value, it should work fine with > regular assignments as long as the completed variable is marked as > volatile (which will prevent compiler mis-optimizations).+1 Using a volatile variable should be enough here.
I have no preference for atomic or volatile methods. Both should be fine.
If only
pthread_timedjoin_npwas not_np... This is exactly the function
that could be used here without any boilerplate.
But is the custom timeout handling in this test even required? Does the
default timeout using SIGALRM not work?The default timeout obviously works.
But with introducing the
thread_B(custom timeout) can get a fine-grained report when theread()was stuck. That's the advantage I can think of.The thread was introduced, because
SIG_ALRMwas apparently not able to
interrupt the read call on linux < 4.14.
But even there it should be possible to interruptpthread_join. So just
replacing the whole loop withpthread_joinshould be enough.
Quoted 2 lines
That should work but no precise log to indicate where goes wrong, so I vote to go the loop way:).
Does it really matter? The being stuck in the read does not seem to be the failure case tested with this test. Otherwise it would TFAIL. Additionally the loop and "custom" timeout was only introduced at a later stage of the test. Initially it relied on the default timeout handling.
If my assumption, that being stuck in the read is not the failure case of this test, then your argument is invalid. Your argument would work for each and every line of code, that might be executed, when the timeout hits.
Joerg
PATCH v2 1/1 crypto/af_alg02: thread termination fixes
Hi Joerg,
Quoted 8 lines
Quoted 2 lines
That should work but no precise log to indicate where goes wrong, so I vote to go the loop way:).
Does it really matter? The being stuck in the read does not seem to be the failure case tested with this test. Otherwise it would TFAIL. Additionally the loop and "custom" timeout was only introduced at a later stage of the test. Initially it relied on the default timeout handling.
It's doesn't matter actually.
Quoted 4 lines
If my assumption, that being stuck in the read is not the failure case of this test, then your argument is invalid. Your argument would work for each and every line of code, that might be executed, when the timeout hits.
Yes right, but I was not arguing for that.
I just provide more opinions for Alexey to make a final decision:).
Signature · 5 lines
Regards, Li Wang -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.linux.it/pipermail/ltp/attachments/20210730/0ac04989/attachment.htm>
PATCH v2 1/1 crypto/af_alg02: thread termination fixes
In reply to Joerg VehlowHi!
Quoted 10 lines
Quoted 2 lines
That should work but no precise log to indicate where goes wrong, so I vote to go the loop way:).
Does it really matter? The being stuck in the read does not seem to be the failure case tested with this test. Otherwise it would TFAIL. Additionally the loop and "custom" timeout was only introduced at a later stage of the test. Initially it relied on the default timeout handling.
If my assumption, that being stuck in the read is not the failure case of this test, then your argument is invalid. Your argument would work for each and every line of code, that might be executed, when the timeout hits.
The top level in the test actually says that for some kernels the read() does not return, which is a kernel bug so the test should report this case as a failure as well and also the commit that fixes this should be added to the tags structure.
Signature · 2 lines
Cyril Hrubis [email protected]
PATCH v2 1/1 crypto/af_alg02: thread termination fixes
Hi,
Quoted 11 lines · Cyril Hrubis
Hi!
Quoted 10 lines
Quoted 2 lines
That should work but no precise log to indicate where goes wrong, so I vote to go the loop way:).
Does it really matter? The being stuck in the read does not seem to be the failure case tested with this test. Otherwise it would TFAIL. Additionally the loop and "custom" timeout was only introduced at a later stage of the test. Initially it relied on the default timeout handling.
If my assumption, that being stuck in the read is not the failure case of this test, then your argument is invalid. Your argument would work for each and every line of code, that might be executed, when the timeout hits.
Quoted 4 lines
The top level in the test actually says that for some kernels the
read()does not return, which is a kernel bug so the test should report this case as a failure as well and also the commit that fixes this should be added to the tags structure.
Ups, yes sorry, should have read the whole test and not just assumed, that the not returning read was something else.
In that case +1 on keeping the custom handling, but maybe use FAIL and not BROK in that case?
I guess there is no other good reason, why the test can trigger the timeout
Joerg