Quelle  bugs.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/*
 * This is for all the tests related to logic bugs (e.g. bad dereferences,
 * bad alignment, bad loops, bad locking, bad scheduling, deep stacks, and
 * lockups) along with other things that don't fit well into existing LKDTM
 * test source files.
 */
#include "lkdtm.h"
#include <linux/cpu.h>
#include <linux/list.h>
#include <linux/sched.h>
#include <linux/sched/signal.h>
#include <linux/sched/task_stack.h>
#include <linux/slab.h>
#include <linux/stop_machine.h>
#include <linux/uaccess.h>

#if IS_ENABLED(CONFIG_X86_32) && !IS_ENABLED(CONFIG_UML)
#include <asm/desc.h>
#endif

struct lkdtm_list {
 struct list_head node;
};

/*
 * Make sure our attempts to over run the kernel stack doesn't trigger
 * a compiler warning when CONFIG_FRAME_WARN is set. Then make sure we
 * recurse past the end of THREAD_SIZE by default.
 */
#if defined(CONFIG_FRAME_WARN) && (CONFIG_FRAME_WARN > 0)
#define REC_STACK_SIZE (_AC(CONFIG_FRAME_WARN, UL) / 2)
#else
#define REC_STACK_SIZE (THREAD_SIZE / 8UL)
#endif
#define REC_NUM_DEFAULT ((THREAD_SIZE / REC_STACK_SIZE) * 2)

static int recur_count = REC_NUM_DEFAULT;

static DEFINE_SPINLOCK(lock_me_up);

/*
 * Make sure compiler does not optimize this function or stack frame away:
 * - function marked noinline
 * - stack variables are marked volatile
 * - stack variables are written (memset()) and read (buf[..] passed as arg)
 * - function may have external effects (memzero_explicit())
 * - no tail recursion possible
 */
static int noinline recursive_loop(int remaining)
{
 volatile char buf[REC_STACK_SIZE];
 volatile int ret;

 memset((void *)buf, remaining & 0xFF, sizeof(buf));
 if (!remaining)
  ret = 0;
 else
  ret = recursive_loop((int)buf[remaining % sizeof(buf)] - 1);
 memzero_explicit((void *)buf, sizeof(buf));
 return ret;
}

/* If the depth is negative, use the default, otherwise keep parameter. */
void __init lkdtm_bugs_init(int *recur_param)
{
 if (*recur_param < 0)
  *recur_param = recur_count;
 else
  recur_count = *recur_param;
}

static void lkdtm_PANIC(void)
{
 panic("dumptest");
}

static int panic_stop_irqoff_fn(void *arg)
{
 atomic_t *v = arg;

 /*
 * As stop_machine() disables interrupts, all CPUs within this function
 * have interrupts disabled and cannot take a regular IPI.
 *
 * The last CPU which enters here will trigger a panic, and as all CPUs
 * cannot take a regular IPI, we'll only be able to stop secondaries if
 * smp_send_stop() or crash_smp_send_stop() uses an NMI.
 */
 if (atomic_inc_return(v) == num_online_cpus())
  panic("panic stop irqoff test");

 for (;;)
  cpu_relax();
}

static void lkdtm_PANIC_STOP_IRQOFF(void)
{
 atomic_t v = ATOMIC_INIT(0);
 stop_machine(panic_stop_irqoff_fn, &v, cpu_online_mask);
}

static void lkdtm_BUG(void)
{
 BUG();
}

static int warn_counter;

static void lkdtm_WARNING(void)
{
 WARN_ON(++warn_counter);
}

static void lkdtm_WARNING_MESSAGE(void)
{
 WARN(1, "Warning message trigger count: %d\n", ++warn_counter);
}

static void lkdtm_EXCEPTION(void)
{
 *((volatile int *) 0) = 0;
}

static void lkdtm_LOOP(void)
{
 for (;;)
  ;
}

static void lkdtm_EXHAUST_STACK(void)
{
 pr_info("Calling function with %lu frame size to depth %d ...\n",
  REC_STACK_SIZE, recur_count);
 recursive_loop(recur_count);
 pr_info("FAIL: survived without exhausting stack?!\n");
}

static noinline void __lkdtm_CORRUPT_STACK(void *stack)
{
 memset(stack, '\xff', 64);
}

/* This should trip the stack canary, not corrupt the return address. */
static noinline void lkdtm_CORRUPT_STACK(void)
{
 /* Use default char array length that triggers stack protection. */
 char data[8] __aligned(sizeof(void *));

 pr_info("Corrupting stack containing char array ...\n");
 __lkdtm_CORRUPT_STACK((void *)&data);
}

/* Same as above but will only get a canary with -fstack-protector-strong */
static noinline void lkdtm_CORRUPT_STACK_STRONG(void)
{
 union {
  unsigned short shorts[4];
  unsigned long *ptr;
 } data __aligned(sizeof(void *));

 pr_info("Corrupting stack containing union ...\n");
 __lkdtm_CORRUPT_STACK((void *)&data);
}

static pid_t stack_pid;
static unsigned long stack_addr;

static void lkdtm_REPORT_STACK(void)
{
 volatile uintptr_t magic;
 pid_t pid = task_pid_nr(current);

 if (pid != stack_pid) {
  pr_info("Starting stack offset tracking for pid %d\n", pid);
  stack_pid = pid;
  stack_addr = (uintptr_t)&magic;
 }

 pr_info("Stack offset: %d\n", (int)(stack_addr - (uintptr_t)&magic));
}

static pid_t stack_canary_pid;
static unsigned long stack_canary;
static unsigned long stack_canary_offset;

static noinline void __lkdtm_REPORT_STACK_CANARY(void *stack)
{
 int i = 0;
 pid_t pid = task_pid_nr(current);
 unsigned long *canary = (unsigned long *)stack;
 unsigned long current_offset = 0, init_offset = 0;

 /* Do our best to find the canary in a 16 word window ... */
 for (i = 1; i < 16; i++) {
  canary = (unsigned long *)stack + i;
#ifdef CONFIG_STACKPROTECTOR
  if (*canary == current->stack_canary)
   current_offset = i;
  if (*canary == init_task.stack_canary)
   init_offset = i;
#endif
 }

 if (current_offset == 0) {
  /*
 * If the canary doesn't match what's in the task_struct,
 * we're either using a global canary or the stack frame
 * layout changed.
 */
  if (init_offset != 0) {
   pr_err("FAIL: global stack canary found at offset %ld (canary for pid %d matches init_task's)!\n",
          init_offset, pid);
  } else {
   pr_warn("FAIL: did not correctly locate stack canary :(\n");
   pr_expected_config(CONFIG_STACKPROTECTOR);
  }

  return;
 } else if (init_offset != 0) {
  pr_warn("WARNING: found both current and init_task canaries nearby?!\n");
 }

 canary = (unsigned long *)stack + current_offset;
 if (stack_canary_pid == 0) {
  stack_canary = *canary;
  stack_canary_pid = pid;
  stack_canary_offset = current_offset;
  pr_info("Recorded stack canary for pid %d at offset %ld\n",
   stack_canary_pid, stack_canary_offset);
 } else if (pid == stack_canary_pid) {
  pr_warn("ERROR: saw pid %d again -- please use a new pid\n", pid);
 } else {
  if (current_offset != stack_canary_offset) {
   pr_warn("ERROR: canary offset changed from %ld to %ld!?\n",
    stack_canary_offset, current_offset);
   return;
  }

  if (*canary == stack_canary) {
   pr_warn("FAIL: canary identical for pid %d and pid %d at offset %ld!\n",
    stack_canary_pid, pid, current_offset);
  } else {
   pr_info("ok: stack canaries differ between pid %d and pid %d at offset %ld.\n",
    stack_canary_pid, pid, current_offset);
   /* Reset the test. */
   stack_canary_pid = 0;
  }
 }
}

static void lkdtm_REPORT_STACK_CANARY(void)
{
 /* Use default char array length that triggers stack protection. */
 char data[8] __aligned(sizeof(void *)) = { };

 __lkdtm_REPORT_STACK_CANARY((void *)&data);
}

static void lkdtm_UNALIGNED_LOAD_STORE_WRITE(void)
{
 static u8 data[5] __attribute__((aligned(4))) = {1, 2, 3, 4, 5};
 u32 *p;
 u32 val = 0x12345678;

 p = (u32 *)(data + 1);
 if (*p == 0)
  val = 0x87654321;
 *p = val;

 if (IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
  pr_err("XFAIL: arch has CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS\n");
}

static void lkdtm_SOFTLOCKUP(void)
{
 preempt_disable();
 for (;;)
  cpu_relax();
}

static void lkdtm_HARDLOCKUP(void)
{
 local_irq_disable();
 for (;;)
  cpu_relax();
}

static void __lkdtm_SMP_CALL_LOCKUP(void *unused)
{
 for (;;)
  cpu_relax();
}

static void lkdtm_SMP_CALL_LOCKUP(void)
{
 unsigned int cpu, target;

 cpus_read_lock();

 cpu = get_cpu();
 target = cpumask_any_but(cpu_online_mask, cpu);

 if (target >= nr_cpu_ids) {
  pr_err("FAIL: no other online CPUs\n");
  goto out_put_cpus;
 }

 smp_call_function_single(target, __lkdtm_SMP_CALL_LOCKUP, NULL, 1);

 pr_err("FAIL: did not hang\n");

out_put_cpus:
 put_cpu();
 cpus_read_unlock();
}

static void lkdtm_SPINLOCKUP(void)
{
 /* Must be called twice to trigger. */
 spin_lock(&lock_me_up);
 /* Let sparse know we intended to exit holding the lock. */
 __release(&lock_me_up);
}

static void __noreturn lkdtm_HUNG_TASK(void)
{
 set_current_state(TASK_UNINTERRUPTIBLE);
 schedule();
 BUG();
}

static volatile unsigned int huge = INT_MAX - 2;
static volatile unsigned int ignored;

static void lkdtm_OVERFLOW_SIGNED(void)
{
 int value;

 value = huge;
 pr_info("Normal signed addition ...\n");
 value += 1;
 ignored = value;

 pr_info("Overflowing signed addition ...\n");
 value += 4;
 ignored = value;
}


static void lkdtm_OVERFLOW_UNSIGNED(void)
{
 unsigned int value;

 value = huge;
 pr_info("Normal unsigned addition ...\n");
 value += 1;
 ignored = value;

 pr_info("Overflowing unsigned addition ...\n");
 value += 4;
 ignored = value;
}

/* Intentionally using unannotated flex array definition. */
struct array_bounds_flex_array {
 int one;
 int two;
 char data[];
};

struct array_bounds {
 int one;
 int two;
 char data[8];
 int three;
};

static void lkdtm_ARRAY_BOUNDS(void)
{
 struct array_bounds_flex_array *not_checked;
 struct array_bounds *checked;
 volatile int i;

 not_checked = kmalloc(sizeof(*not_checked) * 2, GFP_KERNEL);
 checked = kmalloc(sizeof(*checked) * 2, GFP_KERNEL);
 if (!not_checked || !checked) {
  kfree(not_checked);
  kfree(checked);
  return;
 }

 pr_info("Array access within bounds ...\n");
 /* For both, touch all bytes in the actual member size. */
 for (i = 0; i < sizeof(checked->data); i++)
  checked->data[i] = 'A';
 /*
 * For the uninstrumented flex array member, also touch 1 byte
 * beyond to verify it is correctly uninstrumented.
 */
 for (i = 0; i < 2; i++)
  not_checked->data[i] = 'A';

 pr_info("Array access beyond bounds ...\n");
 for (i = 0; i < sizeof(checked->data) + 1; i++)
  checked->data[i] = 'B';

 kfree(not_checked);
 kfree(checked);
 pr_err("FAIL: survived array bounds overflow!\n");
 if (IS_ENABLED(CONFIG_UBSAN_BOUNDS))
  pr_expected_config(CONFIG_UBSAN_TRAP);
 else
  pr_expected_config(CONFIG_UBSAN_BOUNDS);
}

struct lkdtm_annotated {
 unsigned long flags;
 int count;
 int array[] __counted_by(count);
};

static volatile int fam_count = 4;

static void lkdtm_FAM_BOUNDS(void)
{
 struct lkdtm_annotated *inst;

 inst = kzalloc(struct_size(inst, array, fam_count + 1), GFP_KERNEL);
 if (!inst) {
  pr_err("FAIL: could not allocate test struct!\n");
  return;
 }

 inst->count = fam_count;
 pr_info("Array access within bounds ...\n");
 inst->array[1] = fam_count;
 ignored = inst->array[1];

 pr_info("Array access beyond bounds ...\n");
 inst->array[fam_count] = fam_count;
 ignored = inst->array[fam_count];

 kfree(inst);

 pr_err("FAIL: survived access of invalid flexible array member index!\n");

 if (!IS_ENABLED(CONFIG_CC_HAS_COUNTED_BY))
  pr_warn("This is expected since this %s was built with a compiler that does not support __counted_by\n",
   lkdtm_kernel_info);
 else if (IS_ENABLED(CONFIG_UBSAN_BOUNDS))
  pr_expected_config(CONFIG_UBSAN_TRAP);
 else
  pr_expected_config(CONFIG_UBSAN_BOUNDS);
}

static void lkdtm_CORRUPT_LIST_ADD(void)
{
 /*
 * Initially, an empty list via LIST_HEAD:
 * test_head.next = &test_head
 * test_head.prev = &test_head
 */
 LIST_HEAD(test_head);
 struct lkdtm_list good, bad;
 void *target[2] = { };
 void *redirection = ⌖

 pr_info("attempting good list addition\n");

 /*
 * Adding to the list performs these actions:
 * test_head.next->prev = &good.node
 * good.node.next = test_head.next
 * good.node.prev = test_head
 * test_head.next = good.node
 */
 list_add(&good.node, &test_head);

 pr_info("attempting corrupted list addition\n");
 /*
 * In simulating this "write what where" primitive, the "what" is
 * the address of &bad.node, and the "where" is the address held
 * by "redirection".
 */
 test_head.next = redirection;
 list_add(&bad.node, &test_head);

 if (target[0] == NULL && target[1] == NULL)
  pr_err("Overwrite did not happen, but no BUG?!\n");
 else {
  pr_err("list_add() corruption not detected!\n");
  pr_expected_config(CONFIG_LIST_HARDENED);
 }
}

static void lkdtm_CORRUPT_LIST_DEL(void)
{
 LIST_HEAD(test_head);
 struct lkdtm_list item;
 void *target[2] = { };
 void *redirection = ⌖

 list_add(&item.node, &test_head);

 pr_info("attempting good list removal\n");
 list_del(&item.node);

 pr_info("attempting corrupted list removal\n");
 list_add(&item.node, &test_head);

 /* As with the list_add() test above, this corrupts "next". */
 item.node.next = redirection;
 list_del(&item.node);

 if (target[0] == NULL && target[1] == NULL)
  pr_err("Overwrite did not happen, but no BUG?!\n");
 else {
  pr_err("list_del() corruption not detected!\n");
  pr_expected_config(CONFIG_LIST_HARDENED);
 }
}

/* Test that VMAP_STACK is actually allocating with a leading guard page */
static void lkdtm_STACK_GUARD_PAGE_LEADING(void)
{
 const unsigned char *stack = task_stack_page(current);
 const unsigned char *ptr = stack - 1;
 volatile unsigned char byte;

 pr_info("attempting bad read from page below current stack\n");

 byte = *ptr;

 pr_err("FAIL: accessed page before stack! (byte: %x)\n", byte);
}

/* Test that VMAP_STACK is actually allocating with a trailing guard page */
static void lkdtm_STACK_GUARD_PAGE_TRAILING(void)
{
 const unsigned char *stack = task_stack_page(current);
 const unsigned char *ptr = stack + THREAD_SIZE;
 volatile unsigned char byte;

 pr_info("attempting bad read from page above current stack\n");

 byte = *ptr;

 pr_err("FAIL: accessed page after stack! (byte: %x)\n", byte);
}

static void lkdtm_UNSET_SMEP(void)
{
#if IS_ENABLED(CONFIG_X86_64) && !IS_ENABLED(CONFIG_UML)
#define MOV_CR4_DEPTH 64
 void (*direct_write_cr4)(unsigned long val);
 unsigned char *insn;
 unsigned long cr4;
 int i;

 cr4 = native_read_cr4();

 if ((cr4 & X86_CR4_SMEP) != X86_CR4_SMEP) {
  pr_err("FAIL: SMEP not in use\n");
  return;
 }
 cr4 &= ~(X86_CR4_SMEP);

 pr_info("trying to clear SMEP normally\n");
 native_write_cr4(cr4);
 if (cr4 == native_read_cr4()) {
  pr_err("FAIL: pinning SMEP failed!\n");
  cr4 |= X86_CR4_SMEP;
  pr_info("restoring SMEP\n");
  native_write_cr4(cr4);
  return;
 }
 pr_info("ok: SMEP did not get cleared\n");

 /*
 * To test the post-write pinning verification we need to call
 * directly into the middle of native_write_cr4() where the
 * cr4 write happens, skipping any pinning. This searches for
 * the cr4 writing instruction.
 */
 insn = (unsigned char *)native_write_cr4;
 OPTIMIZER_HIDE_VAR(insn);
 for (i = 0; i < MOV_CR4_DEPTH; i++) {
  /* mov %rdi, %cr4 */
  if (insn[i] == 0x0f && insn[i+1] == 0x22 && insn[i+2] == 0xe7)
   break;
  /* mov %rdi,%rax; mov %rax, %cr4 */
  if (insn[i]   == 0x48 && insn[i+1] == 0x89 &&
      insn[i+2] == 0xf8 && insn[i+3] == 0x0f &&
      insn[i+4] == 0x22 && insn[i+5] == 0xe0)
   break;
 }
 if (i >= MOV_CR4_DEPTH) {
  pr_info("ok: cannot locate cr4 writing call gadget\n");
  return;
 }
 direct_write_cr4 = (void *)(insn + i);

 pr_info("trying to clear SMEP with call gadget\n");
 direct_write_cr4(cr4);
 if (native_read_cr4() & X86_CR4_SMEP) {
  pr_info("ok: SMEP removal was reverted\n");
 } else {
  pr_err("FAIL: cleared SMEP not detected!\n");
  cr4 |= X86_CR4_SMEP;
  pr_info("restoring SMEP\n");
  native_write_cr4(cr4);
 }
#else
 pr_err("XFAIL: this test is x86_64-only\n");
#endif
}

static void lkdtm_DOUBLE_FAULT(void)
{
#if IS_ENABLED(CONFIG_X86_32) && !IS_ENABLED(CONFIG_UML)
 /*
 * Trigger #DF by setting the stack limit to zero.  This clobbers
 * a GDT TLS slot, which is okay because the current task will die
 * anyway due to the double fault.
 */
 struct desc_struct d = {
  .type = 3, /* expand-up, writable, accessed data */
  .p = 1,  /* present */
  .d = 1,  /* 32-bit */
  .g = 0,  /* limit in bytes */
  .s = 1,  /* not system */
 };

 local_irq_disable();
 write_gdt_entry(get_cpu_gdt_rw(smp_processor_id()),
   GDT_ENTRY_TLS_MIN, &d, DESCTYPE_S);

 /*
 * Put our zero-limit segment in SS and then trigger a fault.  The
 * 4-byte access to (%esp) will fault with #SS, and the attempt to
 * deliver the fault will recursively cause #SS and result in #DF.
 * This whole process happens while NMIs and MCEs are blocked by the
 * MOV SS window.  This is nice because an NMI with an invalid SS
 * would also double-fault, resulting in the NMI or MCE being lost.
 */
 asm volatile ("movw %0, %%ss; addl $0, (%%esp)" ::
        "r" ((unsigned short)(GDT_ENTRY_TLS_MIN << 3)));

 pr_err("FAIL: tried to double fault but didn't die\n");
#else
 pr_err("XFAIL: this test is ia32-only\n");
#endif
}

#ifdef CONFIG_ARM64
static noinline void change_pac_parameters(void)
{
 if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL)) {
  /* Reset the keys of current task */
  ptrauth_thread_init_kernel(current);
  ptrauth_thread_switch_kernel(current);
 }
}
#endif

static noinline void lkdtm_CORRUPT_PAC(void)
{
#ifdef CONFIG_ARM64
#define CORRUPT_PAC_ITERATE 10
 int i;

 if (!IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL))
  pr_err("FAIL: kernel not built with CONFIG_ARM64_PTR_AUTH_KERNEL\n");

 if (!system_supports_address_auth()) {
  pr_err("FAIL: CPU lacks pointer authentication feature\n");
  return;
 }

 pr_info("changing PAC parameters to force function return failure...\n");
 /*
 * PAC is a hash value computed from input keys, return address and
 * stack pointer. As pac has fewer bits so there is a chance of
 * collision, so iterate few times to reduce the collision probability.
 */
 for (i = 0; i < CORRUPT_PAC_ITERATE; i++)
  change_pac_parameters();

 pr_err("FAIL: survived PAC changes! Kernel may be unstable from here\n");
#else
 pr_err("XFAIL: this test is arm64-only\n");
#endif
}

static struct crashtype crashtypes[] = {
 CRASHTYPE(PANIC),
 CRASHTYPE(PANIC_STOP_IRQOFF),
 CRASHTYPE(BUG),
 CRASHTYPE(WARNING),
 CRASHTYPE(WARNING_MESSAGE),
 CRASHTYPE(EXCEPTION),
 CRASHTYPE(LOOP),
 CRASHTYPE(EXHAUST_STACK),
 CRASHTYPE(CORRUPT_STACK),
 CRASHTYPE(CORRUPT_STACK_STRONG),
 CRASHTYPE(REPORT_STACK),
 CRASHTYPE(REPORT_STACK_CANARY),
 CRASHTYPE(UNALIGNED_LOAD_STORE_WRITE),
 CRASHTYPE(SOFTLOCKUP),
 CRASHTYPE(HARDLOCKUP),
 CRASHTYPE(SMP_CALL_LOCKUP),
 CRASHTYPE(SPINLOCKUP),
 CRASHTYPE(HUNG_TASK),
 CRASHTYPE(OVERFLOW_SIGNED),
 CRASHTYPE(OVERFLOW_UNSIGNED),
 CRASHTYPE(ARRAY_BOUNDS),
 CRASHTYPE(FAM_BOUNDS),
 CRASHTYPE(CORRUPT_LIST_ADD),
 CRASHTYPE(CORRUPT_LIST_DEL),
 CRASHTYPE(STACK_GUARD_PAGE_LEADING),
 CRASHTYPE(STACK_GUARD_PAGE_TRAILING),
 CRASHTYPE(UNSET_SMEP),
 CRASHTYPE(DOUBLE_FAULT),
 CRASHTYPE(CORRUPT_PAC),
};

struct crashtype_category bugs_crashtypes = {
 .crashtypes = crashtypes,
 .len     = ARRAY_SIZE(crashtypes),
};