On Mon, Apr 29, 2024 at 04:06:39PM +0300, Jarkko Sakkinen wrote:
> On Mon Apr 29, 2024 at 1:43 PM EEST, Dmitrii Kuvaiskii wrote:
> > SGX runtimes such as Gramine may implement EDMM-based lazy allocation of
> > enclave pages and may support MADV_DONTNEED semantics [1]. The former
> > implies #PF-based page allocation, and the latter implies the usage of
> > SGX_IOC_ENCLAVE_REMOVE_PAGES ioctl.
> >
> > A trivial program like below (run under Gramine and with EDMM enabled)
> > stresses these two flows in the SGX driver and hangs:
> >
> > /* repeatedly touch different enclave pages at random and mix with
> >  * `madvise(MADV_DONTNEED)` to stress EAUG/EREMOVE flows */
> > static void* thread_func(void* arg) {
> >     size_t num_pages = 0xA000 / page_size;
> >     for (int i = 0; i < 5000; i++) {
> >         size_t page = get_random_ulong() % num_pages;
> >         char data = READ_ONCE(((char*)arg)[page * page_size]);
> >
> >         page = get_random_ulong() % num_pages;
> >         madvise(arg + page * page_size, page_size, MADV_DONTNEED);
> >     }
> > }
> >
> > addr = mmap(NULL, 0xA000, PROT_READ | PROT_WRITE, MAP_ANONYMOUS, -1, 0);
> > pthread_t threads[16];
> > for (int i = 0; i < 16; i++)
> >     pthread_create(&threads[i], NULL, thread_func, addr);
> 
> I'm not convinced that kernel is the problem here but it could be also
> how Gramine is implemented.
> 
> So maybe you could make a better case of that. The example looks a bit
> artificial to me.

I believe that these are the bugs in the kernel (in the SGX driver). I
provided more detailed descriptions of the races and ensuing bugs in the
other two replies, please check them.

The example is a stress test written to debug very infrequent hangs of
real-world applications that are run with Gramine, EDMM, and two
optimizations (lazy allocation and MADV_DONTNEED semantics). We observed
hangs of Node.js, PyTorch, R, iperf, Blender, Nginx. To root cause these
hangs, we wrote this artificial stress test. This test succeeds on vanilla
Linux, so ideally it should also pass on Gramine.

Please also note that the optimizations of lazy allocation and
MADV_DONTNEED provide significant performance improvement for some
workloads that run on Gramine. For example, a Java workload with a 16GB
enclave size has approx. 57x improvement in total runtime. Thus, we
consider it important to permit these optimizations in Gramine, which IIUC
requires bug fixes in the SGX driver.

You can find more info at
https://github.com/gramineproject/gramine/pull/1513.

Which parts do you consider artificial, and how could I modify the stress
test?

--
Dmitrii Kuvaiskii

Reply via email to