Re: [PATCH 2/2] selftests/mm: compaction_test: Fix trivial test success and reduce probability of OOM-killer invocation

[Dev Jain]

On 5/20/24 05:33, Andrew Morton wrote:
> On Wed, 15 May 2024 15:06:33 +0530 Dev Jain <dev.jain@xxxxxxx> wrote:
>
> > Reset nr_hugepages to zero before the start of the test.
> >
> > If a non-zero number of hugepages is already set before the start of the
> > test, the following problems arise:
> >
> >   - The probability of the test getting OOM-killed increases.
> > Proof: The test wants to run on 80% of available memory to prevent
> > OOM-killing (see original code comments). Let the value of mem_free at the
> > start of the test, when nr_hugepages = 0, be x. In the other case, when
> > nr_hugepages > 0, let the memory consumed by hugepages be y. In the former
> > case, the test operates on 0.8 * x of memory. In the latter, the test
> > operates on 0.8 * (x - y) of memory, with y already filled, hence, memory
> > consumed is y + 0.8 * (x - y) = 0.8 * x + 0.2 * y > 0.8 * x. Q.E.D
> >
> >   - The probability of a bogus test success increases.
> > Proof: Let the memory consumed by hugepages be greater than 25% of x,
> > with x and y defined as above. The definition of compaction_index is
> > c_index = (x - y)/z where z is the memory consumed by hugepages after
> > trying to increase them again. In check_compaction(), we set the number
> > of hugepages to zero, and then increase them back; the probability that
> > they will be set back to consume at least y amount of memory again is
> > very high (since there is not much delay between the two attempts of
> > changing nr_hugepages). Hence, z >= y > (x/4) (by the 25% assumption).
> > Therefore,
> > c_index = (x - y)/z <= (x - y)/y = x/y - 1 < 4 - 1 = 3
> > hence, c_index can always be forced to be less than 3, thereby the test
> > succeeding always. Q.E.D
> >
> > NOTE: This patch depends on the previous one.
> >
> > -int check_compaction(unsigned long mem_free, unsigned int hugepage_size)
> > +int check_compaction(unsigned long mem_free, unsigned int hugepage_size,
> > +		     int initial_nr_hugepages)
> >   {
> >   	int fd, ret = -1;
> >   	int compaction_index = 0;
> > -	char initial_nr_hugepages[10] = {0};
> >   	char nr_hugepages[10] = {0};
> > +	char init_nr_hugepages[10] = {0};
> > +
> > +	sprintf(init_nr_hugepages, "%d", initial_nr_hugepages);
> Well, [10] isn't really large enough.  "-1111111111" requires 12 chars,
> with the trailing \0.  And I'd suggest an unsigned type and a %u -
> negative initial_nr_hugepages doesn't make a lot of sense.
>
> >   
> > +int set_zero_hugepages(int *initial_nr_hugepages)
> > +{
> > +	int fd, ret = -1;
> > +	char nr_hugepages[10] = {0};
> Ditto?


Sure, makes sense. I'll just change that to 20 and make it unsigned.

> > +	fd = open("/proc/sys/vm/nr_hugepages", O_RDWR | O_NONBLOCK);
> > +	if (fd < 0) {
> > +		ksft_print_msg("Failed to open /proc/sys/vm/nr_hugepages: %s\n",
> > +			       strerror(errno));
> > +		goto out;
> > +	}
> > +
> > +	if (read(fd, nr_hugepages, sizeof(nr_hugepages)) <= 0) {
> > +		ksft_print_msg("Failed to read from /proc/sys/vm/nr_hugepages: %s\n",
> > +			       strerror(errno));
> > +		goto close_fd;
> > +	}
> > +
> > +	lseek(fd, 0, SEEK_SET);
> > +
> > +	/* Start with the initial condition of 0 huge pages */
> > +	if (write(fd, "0", sizeof(char)) != sizeof(char)) {
> > +		ksft_print_msg("Failed to write 0 to /proc/sys/vm/nr_hugepages: %s\n",
> > +			       strerror(errno));
> > +		goto close_fd;
> > +	}
> > +
> > +	*initial_nr_hugepages = atoi(nr_hugepages);
> > +	ret = 0;
> > +
> > + close_fd:
> > +	close(fd);
> > +
> > + out:
> > +	return ret;
> > +}