/*
* stress_vm_rw
* stress vm_read_v/vm_write_v
*/
int stress_vm_rw(
uint64_t *const counter,
const uint32_t instance,
const uint64_t max_ops,
const char *name)
{
context_t ctxt;
uint8_t stack[64*1024];
const ssize_t stack_offset =
stress_get_stack_direction(&ctxt) * (STACK_SIZE - 64);
uint8_t *stack_top = stack + stack_offset;
(void)instance;
if (!set_vm_rw_bytes) {
if (opt_flags & OPT_FLAGS_MAXIMIZE)
opt_vm_rw_bytes = MAX_VM_RW_BYTES;
if (opt_flags & OPT_FLAGS_MINIMIZE)
opt_vm_rw_bytes = MIN_VM_RW_BYTES;
}
ctxt.name = name;
ctxt.page_size = stress_get_pagesize();
ctxt.sz = opt_vm_rw_bytes & ~(ctxt.page_size - 1);
ctxt.counter = counter;
ctxt.max_ops = max_ops;
if (pipe(ctxt.pipe_wr) < 0) {
pr_fail_dbg(name, "pipe");
return EXIT_FAILURE;
}
if (pipe(ctxt.pipe_rd) < 0) {
(void)close(ctxt.pipe_wr[0]);
(void)close(ctxt.pipe_wr[1]);
pr_fail_dbg(name, "pipe");
return EXIT_FAILURE;
}
ctxt.pid = clone(stress_vm_child, align_stack(stack_top),
SIGCHLD | CLONE_VM, &ctxt);
if (ctxt.pid < 0) {
(void)close(ctxt.pipe_wr[0]);
(void)close(ctxt.pipe_wr[1]);
(void)close(ctxt.pipe_rd[0]);
(void)close(ctxt.pipe_rd[1]);
pr_fail_dbg(name, "clone");
return EXIT_FAILURE;
}
return stress_vm_parent(&ctxt);
}
/*
* stress_bind_mount()
* stress bind mounting
*/
int stress_bind_mount(
uint64_t *const counter,
const uint32_t instance,
const uint64_t max_ops,
const char *name)
{
int pid = 0, status;
context_t context;
const ssize_t stack_offset =
stress_get_stack_direction(&pid) *
(CLONE_STACK_SIZE - 64);
char stack[CLONE_STACK_SIZE];
char *stack_top = stack + stack_offset;
(void)instance;
context.name = name;
context.max_ops = max_ops;
context.counter = counter;
pid = clone(stress_bind_mount_child,
stack_top, CLONE_NEWUSER | CLONE_NEWNS | CLONE_NEWPID | CLONE_VM, &context, 0);
if (pid < 0) {
int rc = exit_status(errno);
pr_fail(stderr, "%s: clone failed: errno=%d (%s)\n",
name, errno, strerror(errno));
return rc;
}
do {
/* Twiddle thumbs */
sleep(1);
} while (opt_do_run && (!max_ops || *counter < max_ops));
(void)kill(pid, SIGKILL);
(void)waitpid(pid, &status, 0);
return EXIT_SUCCESS;
}
/*
* stress_userfaultfd_oomable()
* stress userfaultfd system call, this
* is an OOM-able child process that the
* parent can restart
*/
static int stress_userfaultfd_oomable(
const args_t *args,
const size_t userfaultfd_bytes)
{
const size_t page_size = args->page_size;
size_t sz;
uint8_t *data;
void *zero_page = NULL;
int fd = -1, fdinfo = -1, status, rc = EXIT_SUCCESS, count = 0;
const unsigned int uffdio_copy = 1 << _UFFDIO_COPY;
const unsigned int uffdio_zeropage = 1 << _UFFDIO_ZEROPAGE;
pid_t pid;
struct uffdio_api api;
struct uffdio_register reg;
context_t c;
bool do_poll = true;
char filename[PATH_MAX];
/* Child clone stack */
static uint8_t stack[STACK_SIZE];
const ssize_t stack_offset =
stress_get_stack_direction() * (STACK_SIZE - 64);
uint8_t *stack_top = stack + stack_offset;
sz = userfaultfd_bytes & ~(page_size - 1);
if (posix_memalign(&zero_page, page_size, page_size)) {
pr_err("%s: zero page allocation failed\n", args->name);
return EXIT_NO_RESOURCE;
}
data = mmap(NULL, sz, PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
if (data == MAP_FAILED) {
rc = EXIT_NO_RESOURCE;
pr_err("%s: mmap failed\n", args->name);
goto free_zeropage;
}
/* Get userfault fd */
if ((fd = shim_userfaultfd(0)) < 0) {
if (errno == ENOSYS) {
pr_inf("%s: stressor will be skipped, "
"userfaultfd not supported\n",
args->name);
rc = EXIT_NOT_IMPLEMENTED;
goto unmap_data;
}
rc = exit_status(errno);
pr_err("%s: userfaultfd failed, errno = %d (%s)\n",
args->name, errno, strerror(errno));
goto unmap_data;
}
(void)snprintf(filename, sizeof(filename), "/proc/%d/fdinfo/%d",
getpid(), fd);
fdinfo = open(filename, O_RDONLY);
if (stress_set_nonblock(fd) < 0)
do_poll = false;
/* API sanity check */
(void)memset(&api, 0, sizeof(api));
api.api = UFFD_API;
api.features = 0;
if (ioctl(fd, UFFDIO_API, &api) < 0) {
pr_err("%s: ioctl UFFDIO_API failed, errno = %d (%s)\n",
args->name, errno, strerror(errno));
rc = EXIT_FAILURE;
goto unmap_data;
}
if (api.api != UFFD_API) {
pr_err("%s: ioctl UFFDIO_API API check failed\n",
args->name);
rc = EXIT_FAILURE;
goto unmap_data;
}
/* Register fault handling mode */
(void)memset(®, 0, sizeof(reg));
reg.range.start = (unsigned long)data;
reg.range.len = sz;
reg.mode = UFFDIO_REGISTER_MODE_MISSING;
if (ioctl(fd, UFFDIO_REGISTER, ®) < 0) {
pr_err("%s: ioctl UFFDIO_REGISTER failed, errno = %d (%s)\n",
args->name, errno, strerror(errno));
rc = EXIT_FAILURE;
goto unmap_data;
}
/* OK, so do we have copy supported? */
if ((reg.ioctls & uffdio_copy) != uffdio_copy) {
pr_err("%s: ioctl UFFDIO_REGISTER did not support _UFFDIO_COPY\n",
args->name);
rc = EXIT_FAILURE;
goto unmap_data;
}
/* OK, so do we have zeropage supported? */
if ((reg.ioctls & uffdio_zeropage) != uffdio_zeropage) {
pr_err("%s: ioctl UFFDIO_REGISTER did not support _UFFDIO_ZEROPAGE\n",
args->name);
rc = EXIT_FAILURE;
goto unmap_data;
}
/* Set up context for child */
c.args = args;
c.data = data;
c.sz = sz;
c.page_size = page_size;
c.parent = getpid();
/*
* We need to clone the child and share the same VM address space
* as parent so we can perform the page fault handling
*/
pid = clone(stress_userfaultfd_child, align_stack(stack_top),
SIGCHLD | CLONE_FILES | CLONE_FS | CLONE_SIGHAND | CLONE_VM, &c);
if (pid < 0) {
pr_err("%s: fork failed, errno = %d (%s)\n",
args->name, errno, strerror(errno));
goto unreg;
}
/* Parent */
do {
struct uffd_msg msg;
ssize_t ret;
/* check we should break out before we block on the read */
if (!g_keep_stressing_flag)
break;
/*
* polled wait exercises userfaultfd_poll
* in the kernel, but only works if fd is NONBLOCKing
*/
if (do_poll) {
struct pollfd fds[1];
(void)memset(fds, 0, sizeof fds);
fds[0].fd = fd;
fds[0].events = POLLIN;
/* wait for 1 second max */
ret = poll(fds, 1, 1000);
if (ret == 0)
continue; /* timed out, redo the poll */
if (ret < 0) {
if (errno == EINTR)
continue;
if (errno != ENOMEM) {
pr_fail_err("poll userfaultfd");
if (!g_keep_stressing_flag)
break;
}
/*
* poll ran out of free space for internal
* fd tables, so give up and block on the
* read anyway
*/
goto do_read;
}
/* No data, re-poll */
if (!(fds[0].revents & POLLIN))
continue;
if (LIKELY(fdinfo > -1) &&
UNLIKELY(count++ >= COUNT_MAX)) {
ret = lseek(fdinfo, 0, SEEK_SET);
if (ret == 0) {
char buffer[4096];
ret = read(fdinfo, buffer, sizeof(buffer));
(void)ret;
}
count = 0;
}
}
do_read:
if ((ret = read(fd, &msg, sizeof(msg))) < 0) {
if (errno == EINTR)
continue;
pr_fail_err("read userfaultfd");
if (!g_keep_stressing_flag)
break;
continue;
}
/* We only expect a page fault event */
if (msg.event != UFFD_EVENT_PAGEFAULT) {
pr_fail_err("userfaultfd msg not pagefault event");
continue;
}
/* We only expect a write fault */
if (!(msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)) {
pr_fail_err("userfaultfd msg not write page fault event");
continue;
}
/* Go handle the page fault */
if (handle_page_fault(args, fd, (uint8_t *)(ptrdiff_t)msg.arg.pagefault.address,
zero_page, data, data + sz, page_size) < 0)
break;
inc_counter(args);
} while (keep_stressing());
/* Run it over, zap child */
(void)kill(pid, SIGKILL);
if (shim_waitpid(pid, &status, 0) < 0) {
pr_dbg("%s: waitpid failed, errno = %d (%s)\n",
args->name, errno, strerror(errno));
}
unreg:
if (ioctl(fd, UFFDIO_UNREGISTER, ®) < 0) {
pr_err("%s: ioctl UFFDIO_UNREGISTER failed, errno = %d (%s)\n",
args->name, errno, strerror(errno));
rc = EXIT_FAILURE;
goto unmap_data;
}
unmap_data:
(void)munmap(data, sz);
free_zeropage:
free(zero_page);
if (fdinfo > -1)
(void)close(fdinfo);
if (fd > -1)
(void)close(fd);
return rc;
}