Quelle  demand_paging_test.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/*
 * KVM demand paging test
 * Adapted from dirty_log_test.c
 *
 * Copyright (C) 2018, Red Hat, Inc.
 * Copyright (C) 2019, Google, Inc.
 */
#include <inttypes.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <pthread.h>
#include <linux/userfaultfd.h>
#include <sys/syscall.h>

#include "kvm_util.h"
#include "test_util.h"
#include "memstress.h"
#include "guest_modes.h"
#include "ucall_common.h"
#include "userfaultfd_util.h"

#ifdef __NR_userfaultfd

static int nr_vcpus = 1;
static uint64_t guest_percpu_mem_size = DEFAULT_PER_VCPU_MEM_SIZE;

static size_t demand_paging_size;
static char *guest_data_prototype;

static void vcpu_worker(struct memstress_vcpu_args *vcpu_args)
{
 struct kvm_vcpu *vcpu = vcpu_args->vcpu;
 int vcpu_idx = vcpu_args->vcpu_idx;
 struct kvm_run *run = vcpu->run;
 struct timespec start;
 struct timespec ts_diff;
 int ret;

 clock_gettime(CLOCK_MONOTONIC, &start);

 /* Let the guest access its memory */
 ret = _vcpu_run(vcpu);
 TEST_ASSERT(ret == 0, "vcpu_run failed: %d", ret);
 if (get_ucall(vcpu, NULL) != UCALL_SYNC) {
  TEST_ASSERT(false,
       "Invalid guest sync status: exit_reason=%s",
       exit_reason_str(run->exit_reason));
 }

 ts_diff = timespec_elapsed(start);
 PER_VCPU_DEBUG("vCPU %d execution time: %ld.%.9lds\n", vcpu_idx,
         ts_diff.tv_sec, ts_diff.tv_nsec);
}

static int handle_uffd_page_request(int uffd_mode, int uffd,
  struct uffd_msg *msg)
{
 pid_t tid = syscall(__NR_gettid);
 uint64_t addr = msg->arg.pagefault.address;
 struct timespec start;
 struct timespec ts_diff;
 int r;

 clock_gettime(CLOCK_MONOTONIC, &start);

 if (uffd_mode == UFFDIO_REGISTER_MODE_MISSING) {
  struct uffdio_copy copy;

  copy.src = (uint64_t)guest_data_prototype;
  copy.dst = addr;
  copy.len = demand_paging_size;
  copy.mode = 0;

  r = ioctl(uffd, UFFDIO_COPY, ©);
  /*
 * With multiple vCPU threads fault on a single page and there are
 * multiple readers for the UFFD, at least one of the UFFDIO_COPYs
 * will fail with EEXIST: handle that case without signaling an
 * error.
 *
 * Note that this also suppress any EEXISTs occurring from,
 * e.g., the first UFFDIO_COPY/CONTINUEs on a page. That never
 * happens here, but a realistic VMM might potentially maintain
 * some external state to correctly surface EEXISTs to userspace
 * (or prevent duplicate COPY/CONTINUEs in the first place).
 */
  if (r == -1 && errno != EEXIST) {
   pr_info("Failed UFFDIO_COPY in 0x%lx from thread %d, errno = %d\n",
    addr, tid, errno);
   return r;
  }
 } else if (uffd_mode == UFFDIO_REGISTER_MODE_MINOR) {
  struct uffdio_continue cont = {0};

  cont.range.start = addr;
  cont.range.len = demand_paging_size;

  r = ioctl(uffd, UFFDIO_CONTINUE, &cont);
  /*
 * With multiple vCPU threads fault on a single page and there are
 * multiple readers for the UFFD, at least one of the UFFDIO_COPYs
 * will fail with EEXIST: handle that case without signaling an
 * error.
 *
 * Note that this also suppress any EEXISTs occurring from,
 * e.g., the first UFFDIO_COPY/CONTINUEs on a page. That never
 * happens here, but a realistic VMM might potentially maintain
 * some external state to correctly surface EEXISTs to userspace
 * (or prevent duplicate COPY/CONTINUEs in the first place).
 */
  if (r == -1 && errno != EEXIST) {
   pr_info("Failed UFFDIO_CONTINUE in 0x%lx, thread %d, errno = %d\n",
    addr, tid, errno);
   return r;
  }
 } else {
  TEST_FAIL("Invalid uffd mode %d", uffd_mode);
 }

 ts_diff = timespec_elapsed(start);

 PER_PAGE_DEBUG("UFFD page-in %d \t%ld ns\n", tid,
         timespec_to_ns(ts_diff));
 PER_PAGE_DEBUG("Paged in %ld bytes at 0x%lx from thread %d\n",
         demand_paging_size, addr, tid);

 return 0;
}

struct test_params {
 int uffd_mode;
 bool single_uffd;
 useconds_t uffd_delay;
 int readers_per_uffd;
 enum vm_mem_backing_src_type src_type;
 bool partition_vcpu_memory_access;
};

static void prefault_mem(void *alias, uint64_t len)
{
 size_t p;

 TEST_ASSERT(alias != NULL, "Alias required for minor faults");
 for (p = 0; p < (len / demand_paging_size); ++p) {
  memcpy(alias + (p * demand_paging_size),
         guest_data_prototype, demand_paging_size);
 }
}

static void run_test(enum vm_guest_mode mode, void *arg)
{
 struct memstress_vcpu_args *vcpu_args;
 struct test_params *p = arg;
 struct uffd_desc **uffd_descs = NULL;
 uint64_t uffd_region_size;
 struct timespec start;
 struct timespec ts_diff;
 double vcpu_paging_rate;
 struct kvm_vm *vm;
 int i, num_uffds = 0;

 vm = memstress_create_vm(mode, nr_vcpus, guest_percpu_mem_size, 1,
     p->src_type, p->partition_vcpu_memory_access);

 demand_paging_size = get_backing_src_pagesz(p->src_type);

 guest_data_prototype = malloc(demand_paging_size);
 TEST_ASSERT(guest_data_prototype,
      "Failed to allocate buffer for guest data pattern");
 memset(guest_data_prototype, 0xAB, demand_paging_size);

 if (p->uffd_mode == UFFDIO_REGISTER_MODE_MINOR) {
  num_uffds = p->single_uffd ? 1 : nr_vcpus;
  for (i = 0; i < num_uffds; i++) {
   vcpu_args = &memstress_args.vcpu_args[i];
   prefault_mem(addr_gpa2alias(vm, vcpu_args->gpa),
         vcpu_args->pages * memstress_args.guest_page_size);
  }
 }

 if (p->uffd_mode) {
  num_uffds = p->single_uffd ? 1 : nr_vcpus;
  uffd_region_size = nr_vcpus * guest_percpu_mem_size / num_uffds;

  uffd_descs = malloc(num_uffds * sizeof(struct uffd_desc *));
  TEST_ASSERT(uffd_descs, "Memory allocation failed");
  for (i = 0; i < num_uffds; i++) {
   struct memstress_vcpu_args *vcpu_args;
   void *vcpu_hva;

   vcpu_args = &memstress_args.vcpu_args[i];

   /* Cache the host addresses of the region */
   vcpu_hva = addr_gpa2hva(vm, vcpu_args->gpa);
   /*
 * Set up user fault fd to handle demand paging
 * requests.
 */
   uffd_descs[i] = uffd_setup_demand_paging(
    p->uffd_mode, p->uffd_delay, vcpu_hva,
    uffd_region_size,
    p->readers_per_uffd,
    &handle_uffd_page_request);
  }
 }

 pr_info("Finished creating vCPUs and starting uffd threads\n");

 clock_gettime(CLOCK_MONOTONIC, &start);
 memstress_start_vcpu_threads(nr_vcpus, vcpu_worker);
 pr_info("Started all vCPUs\n");

 memstress_join_vcpu_threads(nr_vcpus);
 ts_diff = timespec_elapsed(start);
 pr_info("All vCPU threads joined\n");

 if (p->uffd_mode) {
  /* Tell the user fault fd handler threads to quit */
  for (i = 0; i < num_uffds; i++)
   uffd_stop_demand_paging(uffd_descs[i]);
 }

 pr_info("Total guest execution time:\t%ld.%.9lds\n",
  ts_diff.tv_sec, ts_diff.tv_nsec);

 vcpu_paging_rate = memstress_args.vcpu_args[0].pages /
      ((double)ts_diff.tv_sec + (double)ts_diff.tv_nsec / NSEC_PER_SEC);
 pr_info("Per-vcpu demand paging rate:\t%f pgs/sec/vcpu\n",
  vcpu_paging_rate);
 pr_info("Overall demand paging rate:\t%f pgs/sec\n",
  vcpu_paging_rate * nr_vcpus);

 memstress_destroy_vm(vm);

 free(guest_data_prototype);
 if (p->uffd_mode)
  free(uffd_descs);
}

static void help(char *name)
{
 puts("");
 printf("usage: %s [-h] [-m vm_mode] [-u uffd_mode] [-a]\n"
     " [-d uffd_delay_usec] [-r readers_per_uffd] [-b memory]\n"
     " [-s type] [-v vcpus] [-c cpu_list] [-o]\n", name);
 guest_modes_help();
 printf(" -u: use userfaultfd to handle vCPU page faults. Mode is a\n"
        " UFFD registration mode: 'MISSING' or 'MINOR'.\n");
 kvm_print_vcpu_pinning_help();
 printf(" -a: Use a single userfaultfd for all of guest memory, instead of\n"
        " creating one for each region paged by a unique vCPU\n"
        " Set implicitly with -o, and no effect without -u.\n");
 printf(" -d: add a delay in usec to the User Fault\n"
        " FD handler to simulate demand paging\n"
        " overheads. Ignored without -u.\n");
 printf(" -r: Set the number of reader threads per uffd.\n");
 printf(" -b: specify the size of the memory region which should be\n"
        " demand paged by each vCPU. e.g. 10M or 3G.\n"
        " Default: 1G\n");
 backing_src_help("-s");
 printf(" -v: specify the number of vCPUs to run.\n");
 printf(" -o: Overlap guest memory accesses instead of partitioning\n"
        " them into a separate region of memory for each vCPU.\n");
 puts("");
 exit(0);
}

int main(int argc, char *argv[])
{
 int max_vcpus = kvm_check_cap(KVM_CAP_MAX_VCPUS);
 const char *cpulist = NULL;
 struct test_params p = {
  .src_type = DEFAULT_VM_MEM_SRC,
  .partition_vcpu_memory_access = true,
  .readers_per_uffd = 1,
  .single_uffd = false,
 };
 int opt;

 guest_modes_append_default();

 while ((opt = getopt(argc, argv, "ahom:u:d:b:s:v:c:r:")) != -1) {
  switch (opt) {
  case 'm':
   guest_modes_cmdline(optarg);
   break;
  case 'u':
   if (!strcmp("MISSING", optarg))
    p.uffd_mode = UFFDIO_REGISTER_MODE_MISSING;
   else if (!strcmp("MINOR", optarg))
    p.uffd_mode = UFFDIO_REGISTER_MODE_MINOR;
   TEST_ASSERT(p.uffd_mode, "UFFD mode must be 'MISSING' or 'MINOR'.");
   break;
  case 'a':
   p.single_uffd = true;
   break;
  case 'd':
   p.uffd_delay = strtoul(optarg, NULL, 0);
   TEST_ASSERT(p.uffd_delay >= 0, "A negative UFFD delay is not supported.");
   break;
  case 'b':
   guest_percpu_mem_size = parse_size(optarg);
   break;
  case 's':
   p.src_type = parse_backing_src_type(optarg);
   break;
  case 'v':
   nr_vcpus = atoi_positive("Number of vCPUs", optarg);
   TEST_ASSERT(nr_vcpus <= max_vcpus,
        "Invalid number of vcpus, must be between 1 and %d", max_vcpus);
   break;
  case 'c':
   cpulist = optarg;
   break;
  case 'o':
   p.partition_vcpu_memory_access = false;
   p.single_uffd = true;
   break;
  case 'r':
   p.readers_per_uffd = atoi(optarg);
   TEST_ASSERT(p.readers_per_uffd >= 1,
        "Invalid number of readers per uffd %d: must be >=1",
        p.readers_per_uffd);
   break;
  case 'h':
  default:
   help(argv[0]);
   break;
  }
 }

 if (p.uffd_mode == UFFDIO_REGISTER_MODE_MINOR &&
     !backing_src_is_shared(p.src_type)) {
  TEST_FAIL("userfaultfd MINOR mode requires shared memory; pick a different -s");
 }

 if (cpulist) {
  kvm_parse_vcpu_pinning(cpulist, memstress_args.vcpu_to_pcpu,
           nr_vcpus);
  memstress_args.pin_vcpus = true;
 }

 for_each_guest_mode(run_test, &p);

 return 0;
}

#else /* __NR_userfaultfd */

#warning "missing __NR_userfaultfd definition"

int main(void)
{
 print_skip("__NR_userfaultfd must be present for userfaultfd test");
 return KSFT_SKIP;
}

#endif /* __NR_userfaultfd */