]> git.ipfire.org Git - thirdparty/qemu.git/commitdiff
os: don't corrupt pre-existing memory-backend data with prealloc
authorDaniel P. Berrange <berrange@redhat.com>
Fri, 3 Mar 2017 11:32:55 +0000 (11:32 +0000)
committerStefan Hajnoczi <stefanha@redhat.com>
Wed, 15 Mar 2017 03:55:41 +0000 (11:55 +0800)
When using a memory-backend object with prealloc turned on, QEMU
will memset() the first byte in every memory page to zero. While
this might have been acceptable for memory backends associated
with RAM, this corrupts application data for NVDIMMs.

Instead of setting every page to zero, read the current byte
value and then just write that same value back, so we are not
corrupting the original data. Directly write the value instead
of memset()ing it, since there's no benefit to memset for a
single byte write.

Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Message-id: 20170303113255.28262-1-berrange@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
util/oslib-posix.c

index 956f66ab4a5caf6a7a59586283d341e48a964147..94d81b9ec0991fb319bcc62910bf05ccdab98bba 100644 (file)
@@ -361,7 +361,19 @@ static void *do_touch_pages(void *arg)
         memset_thread_failed = true;
     } else {
         for (i = 0; i < numpages; i++) {
-            memset(addr, 0, 1);
+            /*
+             * Read & write back the same value, so we don't
+             * corrupt existing user/app data that might be
+             * stored.
+             *
+             * 'volatile' to stop compiler optimizing this away
+             * to a no-op
+             *
+             * TODO: get a better solution from kernel so we
+             * don't need to write at all so we don't cause
+             * wear on the storage backing the region...
+             */
+            *(volatile char *)addr = *addr;
             addr += hpagesize;
         }
     }