perf_counter: Make sure we dont leak kernel memory to userspace
Anton Blanchard [Thu, 16 Jul 2009 13:15:52 +0000 (15:15 +0200)]
There are a few places we are leaking tiny amounts of kernel
memory to userspace. This happens when writing out strings
because we always align the end to 64 bits.

To avoid this we should always use an appropriately sized
temporary buffer and ensure it is zeroed.

Since d_path assembles the string from the end of the buffer
backwards, we need to add 64 bits after the buffer to allow for
alignment.

We also need to copy arch_vma_name to the temporary buffer,
because if we use it directly we may end up copying to
userspace a number of bytes after the end of the string
constant.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090716104817.273972048@samba.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

kernel/perf_counter.c

index c6c38fb..f7a8ab9 100644 (file)
@@ -2968,8 +2968,10 @@ static void perf_counter_comm_event(struct perf_comm_event *comm_event)
        struct perf_cpu_context *cpuctx;
        struct perf_counter_context *ctx;
        unsigned int size;
-       char *comm = comm_event->task->comm;
+       char comm[TASK_COMM_LEN];
 
+       memset(comm, 0, sizeof(comm));
+       strncpy(comm, comm_event->task->comm, sizeof(comm));
        size = ALIGN(strlen(comm)+1, sizeof(u64));
 
        comm_event->comm = comm;
@@ -3088,8 +3090,15 @@ static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
        char *buf = NULL;
        const char *name;
 
+       memset(tmp, 0, sizeof(tmp));
+
        if (file) {
-               buf = kzalloc(PATH_MAX, GFP_KERNEL);
+               /*
+                * d_path works from the end of the buffer backwards, so we
+                * need to add enough zero bytes after the string to handle
+                * the 64bit alignment we do later.
+                */
+               buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
                if (!buf) {
                        name = strncpy(tmp, "//enomem", sizeof(tmp));
                        goto got_name;
@@ -3100,9 +3109,11 @@ static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
                        goto got_name;
                }
        } else {
-               name = arch_vma_name(mmap_event->vma);
-               if (name)
+               if (arch_vma_name(mmap_event->vma)) {
+                       name = strncpy(tmp, arch_vma_name(mmap_event->vma),
+                                      sizeof(tmp));
                        goto got_name;
+               }
 
                if (!vma->vm_mm) {
                        name = strncpy(tmp, "[vdso]", sizeof(tmp));