--- /dev/null
+// SPDX-License-Identifier: GPL-2.0+
+//
+// Copyright 2022, Michael Ellerman, IBM Corp.
+//
+// Test that the 4PB address space SLB handling doesn't corrupt userspace registers
+// (r9-r13) due to a SLB fault while saving the PPR.
+//
+// The bug was introduced in f384796c4 ("powerpc/mm: Add support for handling > 512TB
+// address in SLB miss") and fixed in 4c2de74cc869 ("powerpc/64: Interrupts save PPR on
+// stack rather than thread_struct").
+//
+// To hit the bug requires the task struct and kernel stack to be in different segments.
+// Usually that requires more than 1TB of RAM, or if that's not practical, boot the kernel
+// with "disable_1tb_segments".
+//
+// The test works by creating mappings above 512TB, to trigger the large address space
+// support. It creates 64 mappings, double the size of the SLB, to cause SLB faults on
+// each access (assuming naive replacement). It then loops over those mappings touching
+// each, and checks that r9-r13 aren't corrupted.
+//
+// It then forks another child and tries again, because a new child process will get a new
+// kernel stack and thread struct allocated, which may be more optimally placed to trigger
+// the bug. It would probably be better to leave the previous child processes hanging
+// around, so that kernel stack & thread struct allocations are not reused, but that would
+// amount to a 30 second fork bomb. The current design reliably triggers the bug on
+// unpatched kernels.
+
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <sys/mman.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+#include "utils.h"
+
+#ifndef MAP_FIXED_NOREPLACE
+#define MAP_FIXED_NOREPLACE MAP_FIXED // "Should be safe" above 512TB
+#endif
+
+#define BASE_ADDRESS (1ul << 50) // 1PB
+#define STRIDE (2ul << 40) // 2TB
+#define SLB_SIZE 32
+#define NR_MAPPINGS (SLB_SIZE * 2)
+
+static volatile sig_atomic_t signaled;
+
+static void signal_handler(int sig)
+{
+ signaled = 1;
+}
+
+#define CHECK_REG(_reg) \
+ if (_reg != _reg##_orig) { \
+ printf(str(_reg) " corrupted! Expected 0x%lx != 0x%lx\n", _reg##_orig, \
+ _reg); \
+ _exit(1); \
+ }
+
+static int touch_mappings(void)
+{
+ unsigned long r9_orig, r10_orig, r11_orig, r12_orig, r13_orig;
+ unsigned long r9, r10, r11, r12, r13;
+ unsigned long addr, *p;
+ int i;
+
+ for (i = 0; i < NR_MAPPINGS; i++) {
+ addr = BASE_ADDRESS + (i * STRIDE);
+ p = (unsigned long *)addr;
+
+ asm volatile("mr %0, %%r9 ;" // Read original GPR values
+ "mr %1, %%r10 ;"
+ "mr %2, %%r11 ;"
+ "mr %3, %%r12 ;"
+ "mr %4, %%r13 ;"
+ "std %10, 0(%11) ;" // Trigger SLB fault
+ "mr %5, %%r9 ;" // Save possibly corrupted values
+ "mr %6, %%r10 ;"
+ "mr %7, %%r11 ;"
+ "mr %8, %%r12 ;"
+ "mr %9, %%r13 ;"
+ "mr %%r9, %0 ;" // Restore original values
+ "mr %%r10, %1 ;"
+ "mr %%r11, %2 ;"
+ "mr %%r12, %3 ;"
+ "mr %%r13, %4 ;"
+ : "=&b"(r9_orig), "=&b"(r10_orig), "=&b"(r11_orig),
+ "=&b"(r12_orig), "=&b"(r13_orig), "=&b"(r9), "=&b"(r10),
+ "=&b"(r11), "=&b"(r12), "=&b"(r13)
+ : "b"(i), "b"(p)
+ : "r9", "r10", "r11", "r12", "r13");
+
+ CHECK_REG(r9);
+ CHECK_REG(r10);
+ CHECK_REG(r11);
+ CHECK_REG(r12);
+ CHECK_REG(r13);
+ }
+
+ return 0;
+}
+
+static int test(void)
+{
+ unsigned long page_size, addr, *p;
+ struct sigaction action;
+ bool hash_mmu;
+ int i, status;
+ pid_t pid;
+
+ // This tests a hash MMU specific bug.
+ FAIL_IF(using_hash_mmu(&hash_mmu));
+ SKIP_IF(!hash_mmu);
+
+ page_size = sysconf(_SC_PAGESIZE);
+
+ for (i = 0; i < NR_MAPPINGS; i++) {
+ addr = BASE_ADDRESS + (i * STRIDE);
+
+ p = mmap((void *)addr, page_size, PROT_READ | PROT_WRITE,
+ MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED_NOREPLACE, -1, 0);
+ if (p == MAP_FAILED) {
+ perror("mmap");
+ printf("Error: couldn't mmap(), confirm kernel has 4PB support?\n");
+ return 1;
+ }
+ }
+
+ action.sa_handler = signal_handler;
+ action.sa_flags = SA_RESTART;
+ FAIL_IF(sigaction(SIGALRM, &action, NULL) < 0);
+
+ // Seen to always crash in under ~10s on affected kernels.
+ alarm(30);
+
+ while (!signaled) {
+ // Fork new processes, to increase the chance that we hit the case where
+ // the kernel stack and task struct are in different segments.
+ pid = fork();
+ if (pid == 0)
+ exit(touch_mappings());
+
+ FAIL_IF(waitpid(-1, &status, 0) == -1);
+ FAIL_IF(WIFSIGNALED(status));
+ FAIL_IF(!WIFEXITED(status));
+ FAIL_IF(WEXITSTATUS(status));
+ }
+
+ return 0;
+}
+
+int main(void)
+{
+ return test_harness(test, "large_vm_gpr_corruption");
+}
projects / linux-2.6-microblaze.git / commitdiff