percpu: don't assume existence of cpu0
authorTejun Heo <tj@kernel.org>
Tue, 1 Sep 2009 12:12:28 +0000 (21:12 +0900)
committerTejun Heo <tj@kernel.org>
Tue, 1 Sep 2009 12:23:18 +0000 (21:23 +0900)
percpu incorrectly assumed that cpu0 was always there which led to the
following warning and eventual oops on sparc machines w/o cpu0.

  WARNING: at mm/percpu.c:651 pcpu_map+0xdc/0x100()
  Modules linked in:
  Call Trace:
    [000000000045eb70] warn_slowpath_common+0x50/0xa0
    [000000000045ebdc] warn_slowpath_null+0x1c/0x40
    [00000000004d493c] pcpu_map+0xdc/0x100
    [00000000004d59a4] pcpu_alloc+0x3e4/0x4e0
    [00000000004d5af8] __alloc_percpu+0x18/0x40
    [00000000005b112c] __percpu_counter_init+0x4c/0xc0
  ...
  Unable to handle kernel NULL pointer dereference
  ...
   I7: <sysfs_new_dirent+0x30/0x120>
   Disabling lock debugging due to kernel taint
   Caller[000000000053c1b0]: sysfs_new_dirent+0x30/0x120
   Caller[000000000053c7a4]: create_dir+0x24/0xc0
   Caller[000000000053c870]: sysfs_create_dir+0x30/0x80
   Caller[00000000005990e8]: kobject_add_internal+0xc8/0x200
  ...
   Kernel panic - not syncing: Attempted to kill the idle task!

This patch fixes the problem by backporting parts from devel branch to
make percpu core not depend on the existence of cpu0.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Meelis Roos <mroos@linux.ee>
Cc: David Miller <davem@davemloft.net>
mm/percpu.c

index 5fe37842e0ea3689bcdf24ec1d77e681ba27a55c..3311c8919f375fa062faad78b2439176f436301a 100644 (file)
@@ -197,7 +197,12 @@ static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
 static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk,
                                     int page_idx)
 {
-       return *pcpu_chunk_pagep(chunk, 0, page_idx) != NULL;
+       /*
+        * Any possible cpu id can be used here, so there's no need to
+        * worry about preemption or cpu hotplug.
+        */
+       return *pcpu_chunk_pagep(chunk, raw_smp_processor_id(),
+                                page_idx) != NULL;
 }
 
 /* set the pointer to a chunk in a page struct */
@@ -297,6 +302,14 @@ static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
                return pcpu_first_chunk;
        }
 
+       /*
+        * The address is relative to unit0 which might be unused and
+        * thus unmapped.  Offset the address to the unit space of the
+        * current processor before looking it up in the vmalloc
+        * space.  Note that any possible cpu id can be used here, so
+        * there's no need to worry about preemption or cpu hotplug.
+        */
+       addr += raw_smp_processor_id() * pcpu_unit_size;
        return pcpu_get_page_chunk(vmalloc_to_page(addr));
 }