From 13cccafe0edcd03bf1c841de8ab8a1c8e34f77d9 Mon Sep 17 00:00:00 2001 From: Brian Foster Date: Tue, 16 Aug 2022 11:54:07 -0400 Subject: s390: fix double free of GS and RI CBs on fork() failure The pointers for guarded storage and runtime instrumentation control blocks are stored in the thread_struct of the associated task. These pointers are initially copied on fork() via arch_dup_task_struct() and then cleared via copy_thread() before fork() returns. If fork() happens to fail after the initial task dup and before copy_thread(), the newly allocated task and associated thread_struct memory are freed via free_task() -> arch_release_task_struct(). This results in a double free of the guarded storage and runtime info structs because the fields in the failed task still refer to memory associated with the source task. This problem can manifest as a BUG_ON() in set_freepointer() (with CONFIG_SLAB_FREELIST_HARDENED enabled) or KASAN splat (if enabled) when running trinity syscall fuzz tests on s390x. To avoid this problem, clear the associated pointer fields in arch_dup_task_struct() immediately after the new task is copied. Note that the RI flag is still cleared in copy_thread() because it resides in thread stack memory and that is where stack info is copied. Signed-off-by: Brian Foster Fixes: 8d9047f8b967c ("s390/runtime instrumentation: simplify task exit handling") Fixes: 7b83c6297d2fc ("s390/guarded storage: simplify task exit handling") Cc: # 4.15 Reviewed-by: Gerald Schaefer Reviewed-by: Heiko Carstens Link: https://lore.kernel.org/r/20220816155407.537372-1-bfoster@redhat.com Signed-off-by: Vasily Gorbik --- arch/s390/kernel/process.c | 22 ++++++++++++++++------ 1 file changed, 16 insertions(+), 6 deletions(-) diff --git a/arch/s390/kernel/process.c b/arch/s390/kernel/process.c index 89949b9f3cf8..d5119e039d85 100644 --- a/arch/s390/kernel/process.c +++ b/arch/s390/kernel/process.c @@ -91,6 +91,18 @@ int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src) memcpy(dst, src, arch_task_struct_size); dst->thread.fpu.regs = dst->thread.fpu.fprs; + + /* + * Don't transfer over the runtime instrumentation or the guarded + * storage control block pointers. These fields are cleared here instead + * of in copy_thread() to avoid premature freeing of associated memory + * on fork() failure. Wait to clear the RI flag because ->stack still + * refers to the source thread. + */ + dst->thread.ri_cb = NULL; + dst->thread.gs_cb = NULL; + dst->thread.gs_bc_cb = NULL; + return 0; } @@ -150,13 +162,11 @@ int copy_thread(struct task_struct *p, const struct kernel_clone_args *args) frame->childregs.flags = 0; if (new_stackp) frame->childregs.gprs[15] = new_stackp; - - /* Don't copy runtime instrumentation info */ - p->thread.ri_cb = NULL; + /* + * Clear the runtime instrumentation flag after the above childregs + * copy. The CB pointer was already cleared in arch_dup_task_struct(). + */ frame->childregs.psw.mask &= ~PSW_MASK_RI; - /* Don't copy guarded storage control block */ - p->thread.gs_cb = NULL; - p->thread.gs_bc_cb = NULL; /* Set a new TLS ? */ if (clone_flags & CLONE_SETTLS) { -- cgit v1.2.3 From 41ac42f137080bc230b5882e3c88c392ab7f2d32 Mon Sep 17 00:00:00 2001 From: Gerald Schaefer Date: Wed, 17 Aug 2022 15:26:03 +0200 Subject: s390/mm: do not trigger write fault when vma does not allow VM_WRITE For non-protection pXd_none() page faults in do_dat_exception(), we call do_exception() with access == (VM_READ | VM_WRITE | VM_EXEC). In do_exception(), vma->vm_flags is checked against that before calling handle_mm_fault(). Since commit 92f842eac7ee3 ("[S390] store indication fault optimization"), we call handle_mm_fault() with FAULT_FLAG_WRITE, when recognizing that it was a write access. However, the vma flags check is still only checking against (VM_READ | VM_WRITE | VM_EXEC), and therefore also calling handle_mm_fault() with FAULT_FLAG_WRITE in cases where the vma does not allow VM_WRITE. Fix this by changing access check in do_exception() to VM_WRITE only, when recognizing write access. Link: https://lkml.kernel.org/r/20220811103435.188481-3-david@redhat.com Fixes: 92f842eac7ee3 ("[S390] store indication fault optimization") Cc: Reported-by: David Hildenbrand Reviewed-by: Heiko Carstens Signed-off-by: Gerald Schaefer Signed-off-by: Vasily Gorbik --- arch/s390/mm/fault.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/arch/s390/mm/fault.c b/arch/s390/mm/fault.c index 13449941516c..09b6e756d521 100644 --- a/arch/s390/mm/fault.c +++ b/arch/s390/mm/fault.c @@ -379,7 +379,9 @@ static inline vm_fault_t do_exception(struct pt_regs *regs, int access) flags = FAULT_FLAG_DEFAULT; if (user_mode(regs)) flags |= FAULT_FLAG_USER; - if (access == VM_WRITE || is_write) + if (is_write) + access = VM_WRITE; + if (access == VM_WRITE) flags |= FAULT_FLAG_WRITE; mmap_read_lock(mm); -- cgit v1.2.3