proc/vmcore: fix clearing user buffer by properly using clear_user()
authorDavid Hildenbrand <david@redhat.com>
Sat, 20 Nov 2021 00:43:58 +0000 (16:43 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 20 Nov 2021 18:35:55 +0000 (10:35 -0800)
To clear a user buffer we cannot simply use memset, we have to use
clear_user().  With a virtio-mem device that registers a vmcore_cb and
has some logically unplugged memory inside an added Linux memory block,
I can easily trigger a BUG by copying the vmcore via "cp":

  systemd[1]: Starting Kdump Vmcore Save Service...
  kdump[420]: Kdump is using the default log level(3).
  kdump[453]: saving to /sysroot/var/crash/127.0.0.1-2021-11-11-14:59:22/
  kdump[458]: saving vmcore-dmesg.txt to /sysroot/var/crash/127.0.0.1-2021-11-11-14:59:22/
  kdump[465]: saving vmcore-dmesg.txt complete
  kdump[467]: saving vmcore
  BUG: unable to handle page fault for address: 00007f2374e01000
  #PF: supervisor write access in kernel mode
  #PF: error_code(0x0003) - permissions violation
  PGD 7a523067 P4D 7a523067 PUD 7a528067 PMD 7a525067 PTE 800000007048f867
  Oops: 0003 [#1] PREEMPT SMP NOPTI
  CPU: 0 PID: 468 Comm: cp Not tainted 5.15.0+ #6
  Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.14.0-27-g64f37cc530f1-prebuilt.qemu.org 04/01/2014
  RIP: 0010:read_from_oldmem.part.0.cold+0x1d/0x86
  Code: ff ff ff e8 05 ff fe ff e9 b9 e9 7f ff 48 89 de 48 c7 c7 38 3b 60 82 e8 f1 fe fe ff 83 fd 08 72 3c 49 8d 7d 08 4c 89 e9 89 e8 <49> c7 45 00 00 00 00 00 49 c7 44 05 f8 00 00 00 00 48 83 e7 f81
  RSP: 0018:ffffc9000073be08 EFLAGS: 00010212
  RAX: 0000000000001000 RBX: 00000000002fd000 RCX: 00007f2374e01000
  RDX: 0000000000000001 RSI: 00000000ffffdfff RDI: 00007f2374e01008
  RBP: 0000000000001000 R08: 0000000000000000 R09: ffffc9000073bc50
  R10: ffffc9000073bc48 R11: ffffffff829461a8 R12: 000000000000f000
  R13: 00007f2374e01000 R14: 0000000000000000 R15: ffff88807bd421e8
  FS:  00007f2374e12140(0000) GS:ffff88807f000000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00007f2374e01000 CR3: 000000007a4aa000 CR4: 0000000000350eb0
  Call Trace:
   read_vmcore+0x236/0x2c0
   proc_reg_read+0x55/0xa0
   vfs_read+0x95/0x190
   ksys_read+0x4f/0xc0
   do_syscall_64+0x3b/0x90
   entry_SYSCALL_64_after_hwframe+0x44/0xae

Some x86-64 CPUs have a CPU feature called "Supervisor Mode Access
Prevention (SMAP)", which is used to detect wrong access from the kernel
to user buffers like this: SMAP triggers a permissions violation on
wrong access.  In the x86-64 variant of clear_user(), SMAP is properly
handled via clac()+stac().

To fix, properly use clear_user() when we're dealing with a user buffer.

Link: https://lkml.kernel.org/r/20211112092750.6921-1-david@redhat.com
Fixes: 997c136f518c ("fs/proc/vmcore.c: add hook to read_from_oldmem() to check for non-ram pages")
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Baoquan He <bhe@redhat.com>
Cc: Dave Young <dyoung@redhat.com>
Cc: Baoquan He <bhe@redhat.com>
Cc: Vivek Goyal <vgoyal@redhat.com>
Cc: Philipp Rudo <prudo@redhat.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/proc/vmcore.c

index 30a3b66..509f851 100644 (file)
@@ -154,9 +154,13 @@ ssize_t read_from_oldmem(char *buf, size_t count,
                        nr_bytes = count;
 
                /* If pfn is not ram, return zeros for sparse dump files */
-               if (!pfn_is_ram(pfn))
-                       memset(buf, 0, nr_bytes);
-               else {
+               if (!pfn_is_ram(pfn)) {
+                       tmp = 0;
+                       if (!userbuf)
+                               memset(buf, 0, nr_bytes);
+                       else if (clear_user(buf, nr_bytes))
+                               tmp = -EFAULT;
+               } else {
                        if (encrypted)
                                tmp = copy_oldmem_page_encrypted(pfn, buf,
                                                                 nr_bytes,
@@ -165,12 +169,12 @@ ssize_t read_from_oldmem(char *buf, size_t count,
                        else
                                tmp = copy_oldmem_page(pfn, buf, nr_bytes,
                                                       offset, userbuf);
-
-                       if (tmp < 0) {
-                               up_read(&vmcore_cb_rwsem);
-                               return tmp;
-                       }
                }
+               if (tmp < 0) {
+                       up_read(&vmcore_cb_rwsem);
+                       return tmp;
+               }
+
                *ppos += nr_bytes;
                count -= nr_bytes;
                buf += nr_bytes;