Re: [PATCH] tests/tcg/s390x: Add cdsg.c

[Ilya Leoshkevich <iii@linux.ibm.com>]

On Tue, Nov 29, 2022 at 09:54:13AM +0100, David Hildenbrand wrote:
> On 29.11.22 00:48, Ilya Leoshkevich wrote:
> > Add a simple test to prevent regressions.
> > 
> > Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
> > ---
> >   tests/tcg/s390x/Makefile.target |  4 ++
> >   tests/tcg/s390x/cdsg.c          | 73 +++++++++++++++++++++++++++++++++
> >   2 files changed, 77 insertions(+)
> >   create mode 100644 tests/tcg/s390x/cdsg.c
> > 
> > diff --git a/tests/tcg/s390x/Makefile.target b/tests/tcg/s390x/Makefile.target
> > index 1d454270c0e..523214dac33 100644
> > --- a/tests/tcg/s390x/Makefile.target
> > +++ b/tests/tcg/s390x/Makefile.target
> > @@ -27,6 +27,7 @@ TESTS+=noexec
> >   TESTS+=div
> >   TESTS+=clst
> >   TESTS+=long-double
> > +TESTS+=cdsg
> >   Z13_TESTS=vistr
> >   $(Z13_TESTS): CFLAGS+=-march=z13 -O2
> > @@ -66,3 +67,6 @@ sha512-mvx: sha512.c
> >   	$(CC) $(CFLAGS) $(EXTRA_CFLAGS) $< -o $@ $(LDFLAGS)
> >   TESTS+=sha512-mvx
> > +
> > +cdsg: CFLAGS+=-pthread
> > +cdsg: LDFLAGS+=-pthread
> > diff --git a/tests/tcg/s390x/cdsg.c b/tests/tcg/s390x/cdsg.c
> > new file mode 100644
> > index 00000000000..83313699f7d
> > --- /dev/null
> > +++ b/tests/tcg/s390x/cdsg.c
> > @@ -0,0 +1,73 @@
> > +#include <assert.h>
> > +#include <pthread.h>
> > +#include <stdbool.h>
> > +#include <stdlib.h>
> > +
> > +static volatile bool start;
> > +static unsigned long val[2] __attribute__((__aligned__(16)));
> > +
> > +void *cdsg_loop(void *arg)
> > +{
> > +    unsigned long orig0, orig1, new0, new1;
> > +    register unsigned long r0 asm("r0");
> > +    register unsigned long r1 asm("r1");
> > +    register unsigned long r2 asm("r2");
> > +    register unsigned long r3 asm("r3");
> > +    int cc;
> > +    int i;
> > +
> > +    while (!start) {
> > +    }
> > +
> > +    orig0 = val[0];
> > +    orig1 = val[1];
> > +    for (i = 0; i < 1000;) {
> 
> Are 1000 iterations sufficient to catch the race window reliably?

Good point, I had to raise it to 10k.
If I break the code like this:

--- a/tcg/tcg-op.c
+++ b/tcg/tcg-op.c
@@ -3509,7 +3509,7 @@ void tcg_gen_atomic_cmpxchg_i128(TCGv_i128 retv, TCGv addr, TCGv_i128 cmpv,
 {
     gen_atomic_cx_i128 gen;
 
-    if (!(tcg_ctx->tb_cflags & CF_PARALLEL)) {
+    if (true) {
         tcg_gen_nonatomic_cmpxchg_i128(retv, addr, cmpv, newv, idx, memop);
         return;
     }
 
the test with 10k iterations fails consistently.
And it's still fast:

$ time -p ./qemu-s390x ./tests/tcg/s390x-linux-user/cdsg
real 0.01

> > +        new0 = orig0 + 1;
> > +        new1 = orig1 + 2;
> > +
> > +        r0 = orig0;
> > +        r1 = orig1;
> > +        r2 = new0;
> > +        r3 = new1;
> > +        asm("cdsg %[r0],%[r2],%[db2]\n"
> > +            "ipm %[cc]"
> > +            : [r0] "+r" (r0)
> > +            , [r1] "+r" (r1)
> > +            , [db2] "=m" (val)
> > +            , [cc] "=r" (cc)
> > +            : [r2] "r" (r2)
> > +            , [r3] "r" (r3)
> > +            : "cc");
> 
> Nit: I'd suggest a simple cdsg helper function that makes this code easier
> to digest.

Ok.

> 
> > +        orig0 = r0;
> > +        orig1 = r1;
> > +        cc = (cc >> 28) & 3;
> > +
> > +        if (cc == 0) {
> > +            orig0 = new0;
> > +            orig1 = new1;
> > +            i++;
> > +        } else {
> > +            assert(cc == 1);
> > +        }
> > +    }
> > +
> > +    return NULL;
> > +}
> > +
> > +int main(void)
> > +{
> > +    pthread_t thread;
> > +    int ret;
> > +
> > +    ret = pthread_create(&thread, NULL, cdsg_loop, NULL);
> > +    assert(ret == 0);
> > +    start = true;
> > +    cdsg_loop(NULL);
> > +    ret = pthread_join(thread, NULL);
> > +    assert(ret == 0);
> > +
> > +    assert(val[0] == 2000);
> > +    assert(val[1] == 4000);
> > +
> > +    return EXIT_SUCCESS;
> > +}
> 
> -- 
> Thanks,
> 
> David / dhildenb
> 
>