[test] pagetable OOM handling (again, but 64-bit this time)

[test] pagetable OOM handling (again, but 64-bit this time)

[William Lee Irwin III]

[-- Attachment #1: Type: text/plain, Size: 8604 bytes --]

/*
 * This appears to panic an Altix. An earlier version of this without
 * the VSZ limiting appeared to panic a 4x logical x86-64 box, and yet
 * another version panicked an E3K. panic string:
 * "Kernel panic - not syncing: Out of memory and no killable processes..."
 * The basic idea is to keep mm->total_vm low enough to evade VSZ rlimits
 * and avoid the OOM killer long enough to get into a state where the
 * SIGKILL will be ignored. mm->total_vm is kept low with the 1 pte per
 * pmd trick. Getting into 'D' state is done by faulting in pte pages.
 * It also relies on a race that clobbers PF_MEMDIE instead of getting
 * stuck in a non-allocating codepath where it waits in 'D' state in
 * order to ignore signals. "Normally" the task->flags setting works;
 * however, the restoration of ->flags after the try_to_free_pages()
 * call in __alloc_pages() causes tasks to honor OOM kills while they
 * don't honor signals per se; the race means that if it's caught
 * between the the load and the store of the restoration (x86 addressing
 * modes make this either far less likely or impossible) PF_MEMDIE is
 * lost. To do this more reliably on x86-64 finding another way to get
 * into 'D' state outside the allocator for a sacrificial victim or
 * something on that order is needed. Having little enough RAM to be
 * able to use the VSZ of pid 1 as a limit would also work. Of course,
 * I can't be 100% certain how it does what it does without more
 * instrumentation, but these things are at least what motivated how I
 * wrote it. Originally zeromap_page_range() on a large (ca. 512GB)
 * virtual area and fork() of only a few processes was what triggered
 * the panic() (and much faster than this does on ia64), but that
 * didn't seem to do anything on x86-64. Altix' console log included.
 * This test is useless on 32-bit architectures.
 */
#include <unistd.h>
#include <fcntl.h>
#include <stdlib.h>
#include <stdio.h>
#include <signal.h>
#include <sys/mman.h>
#include <sys/stat.h>

extern char _end[];

static int nr_signals;

void ignore(int sig, siginfo_t *info, void *ucontext)
{
	(void)sig;
	(void)info;
	(void)ucontext;
	++nr_signals;
}

int main(void)
{
	long page_size, pmd_size, physpages, pages, nr_children, max_fds,
		init_vsz, total_vsz = 0;
	char *start, *p, *end, *pos;
	const int flags = MAP_ANONYMOUS|MAP_PRIVATE|MAP_FIXED,
				prot = PROT_READ|PROT_WRITE;
	int fd, *child_fds, child;
	struct sigaction action = {
		.sa_sigaction = ignore,
		.sa_flags = SA_SIGINFO,
	};
	FILE *file;

	file = fopen("/proc/1/status", "r");
	if (!file) {
		fprintf(stderr, "couldn't open /proc/1/status\n");
		exit(EXIT_FAILURE);
	} else {
		char *buf = NULL;
		size_t bufsz;

		while (getline(&buf, &bufsz, file) > 0) {
			if (sscanf(buf, "VmSize: %ld", &init_vsz) == 1) {
				init_vsz <<= 17;
				fclose(file);
				goto self_vsz;
			}
		}
		fprintf(stderr, "failed to parse /proc/1/status\n");
		exit(EXIT_FAILURE);
self_vsz:
		if (!(file = fopen("/proc/self/status", "r"))) {
			fprintf(stderr, "couldn't open /proc/self/status\n");
			exit(EXIT_FAILURE);
		}
		while (getline(&buf, &bufsz, file) > 0) {
			if (sscanf(buf, "VmSize: %ld", &total_vsz) == 1) {
				total_vsz <<= 10;
				fclose(file);
				free(buf);
				goto setup;
			}
		}
		fprintf(stderr, "failed to parse /proc/self/status\n");
		exit(EXIT_FAILURE);
	}
setup:
	nr_signals = 0;
	sigemptyset(&action.sa_mask);
	sigaction(SIGSEGV, &action, NULL);
	sigaction(SIGBUS, &action, NULL);
	setvbuf(stdin, NULL, _IONBF, 0);
	setvbuf(stdout, NULL, _IONBF, 0);
	setvbuf(stderr, NULL, _IONBF, 0);
	fd = open("/dev/zero", O_RDONLY|O_NOCTTY);
	if ((page_size = sysconf(_SC_PAGE_SIZE)) < 0)
		exit(EXIT_FAILURE);
	pmd_size = (page_size*page_size)/sizeof(long);
	start = (char *)(((long)_end + pmd_size - 1) & ~(pmd_size - 1));
	if ((unsigned long)start < (unsigned long)_end + page_size)
		start = start + pmd_size - page_size;
	else
		start -= page_size;
	printf("pmd_size = 0x%lx\n", pmd_size);
	printf("pgdir_size = 0x%lx\n", pmd_size*page_size/sizeof(long));
	close(STDIN_FILENO);
	close(STDOUT_FILENO);
	/* close(STDERR_FILENO); */
	if (fork())
		exit(EXIT_SUCCESS);
	setsid();
	pos = start;
	while ((p = mmap(pos, 2*page_size, prot, flags, 0, 0)) != MAP_FAILED &&
			total_vsz < init_vsz) {
		pos += 2*pmd_size;
		total_vsz += 2*page_size;
		if ((long)pos < 0) {
			fprintf(stderr, "starting address wrapped\n");
			break;
		} else if (pos >= (char *)((long)&p & ~(pmd_size - 1))) {
			fprintf(stderr, "starting address hit the stack\n");
			break;
		}
	}
	end = pos;
	if ((physpages = sysconf(_SC_PHYS_PAGES)) < 0) {
		FILE *file;
		char *buf = NULL;
		size_t bufsz;

		fprintf(stderr, "sysconf(_SC_PHYS_PAGES) failed, ret = %ld\n",
				physpages);
		fprintf(stderr, "trying to parse /proc/meminfo\n");
		file = fopen("/proc/meminfo", "r");
		if (!file) {
			fprintf(stderr, "failed to open /proc/meminfo\n");
			exit(EXIT_FAILURE);
		}
		if (getline(&buf, &bufsz, file) <= 0) {
			fprintf(stderr, "failed to read first line of /proc/meminfo\n");
			exit(EXIT_FAILURE);
		}
		if (sscanf(buf, "MemTotal: %ld", &physpages) != 1) {
			fprintf(stderr, "failed to parse first line of /proc/meminfo\n");
			exit(EXIT_FAILURE);
		}
		if (physpages <= 0) {
			fprintf(stderr, "parsed /proc/meminfo, but result was crap\n");
			exit(EXIT_FAILURE);
		}
		fclose(file);
		free(buf);
	}
	if ((max_fds = sysconf(_SC_OPEN_MAX)) < 0) {
		fprintf(stderr, "sysconf(_SC_OPEN_MAX) failed\n");
		exit(EXIT_FAILURE);
	}
	pages = (end - start)/pmd_size;
	nr_children = (physpages + pages - 1)/pages;
	fprintf(stderr, "blocking\n");
	fprintf(stderr, "start = %p, end = %p\n", start, end);
	fprintf(stderr, "%ld physical pages\n", physpages);
	fprintf(stderr, "%ld pagetable pages per process\n", pages);
	fprintf(stderr, "%ld children needed to exhaust RAM\n", nr_children);
	fprintf(stderr, "%ld open files maximum\n", max_fds);
	fprintf(stderr, "%ld KB VSZ per process\n", total_vsz >> 10);
	nr_children *= 8;
	if (nr_children*sizeof(int [2]) > (size_t)(2*page_size)) {
		fprintf(stderr, "Either your kernel port or your arch sucks\n"
				"You need more than two pages worth of "
				"storage to hold the fd's to communicate "
				"with the children\n"
				"This thing wants %ld fd's, %zd bytes of "
				"RAM to store them, and thinks the "
				"page size on your box is %ld\n",
				2*nr_children,
				nr_children*sizeof(int [2]),
				page_size);
		exit(EXIT_FAILURE);
	}
	if (2*nr_children + 1 > max_fds) {
		fprintf(stderr, "Either your rlimits are utterly retarded "
				"or your arch sucks. You have an open fd "
				"limit of %ld fd's per process and this "
				"thing wants to spawn %ld processes, and "
				"open 2 fd's for each for %ld total fd's.\n",
				max_fds,
				nr_children,
				2*nr_children);
	}
	if ((child = fork()) < 0)
		exit(EXIT_FAILURE);
	else if (child)
		exit(EXIT_SUCCESS);
	else
		setsid();
	child_fds = (int *)start;
	for (child = 0; child < nr_children; ++child) {
		if (pipe(&child_fds[2*child])) {
			fprintf(stderr, "pipe(2) failed, bailing out\n");
			exit(EXIT_FAILURE);
		}
	}
	for (child = 0; child < nr_children; ++child) {
		pid_t pid = fork();

		if (pid < 0)
			fprintf(stderr, "fork() appears to have failed "
					"for child %d\n",
					child);
		else if (!pid) {
			char c;

			read(child_fds[2*child], &c, 1);
			action.sa_sigaction = ignore,
			action.sa_flags = SA_SIGINFO,
			sigemptyset(&action.sa_mask);
			sigaction(SIGSEGV, &action, NULL);
			sigaction(SIGBUS, &action, NULL);
			sigfillset(&action.sa_mask);
			sigprocmask(SIG_BLOCK, &action.sa_mask, NULL);
			for (pos = start; pos < end; pos += 2*pmd_size) {
				if (*pos && pos - start >= pmd_size)
					fprintf(stderr, "bad kernel!\n"
						"zeromapped memory appears to "
						"be corrupted! (check 1)\n");
				if (pos[page_size] && pos - start >= pmd_size)
					fprintf(stderr, "bad kernel!\n"
						"zeromapped memory appears to "
						"be corrupted! (check 1)\n");
			}
			pause();
			return EXIT_SUCCESS;
		}
	}
	for (pos = start + pmd_size; pos < end; pos += pmd_size)
		munmap(pos, 2*page_size);
	for (child = 0; child < nr_children; ++child) {
		char c = '\0';

		fprintf(stderr, "pretending to wake child %d\n", child);
		write(child_fds[2*child+1], &c, 1);
	}
	/*
	 * The parent now offers itself as the sacrificial victim to be
	 * OOM killed in the place of its children.
	 */
	start += pmd_size;
	p = mmap(start, (((size_t)&child) & ~(pmd_size - 1)) - (size_t)start,
			PROT_READ, MAP_PRIVATE|MAP_FIXED, fd, 0);
	if (p == MAP_FAILED)
		fprintf(stderr, "mmap in parent failed!\n");
	close(fd);
	pause();
	return EXIT_SUCCESS;
}

[-- Attachment #2: altix.log.51.bz2 --]
[-- Type: application/octet-stream, Size: 19946 bytes --]