On Feb 2, 2015, at 12:48 PM, David Rientjes wrote:
> On Sun, 1 Feb 2015, green@xxxxxxxxxxxxxx wrote:
>
>> From: Bruno Faccini <bruno.faccini@xxxxxxxxx>
>>
>> When possible, try to use of __vmalloc_node() instead of
>> vzalloc/vzalloc_node which allows for protection flag specification,
>> and particularly to not set __GFP_FS, which can cause some deadlock
>> situations in our code due to recursive calls.
> You're saying that all usage of OBD_ALLOC_LARGE() and
> OBD_CPT_ALLOC_LARGE() are in contexts where we need GFP_NOFS? It would be
Most of them fore sure (hm, there's only one OBD_CPT_ALLOC_LARGE in the client
and I imagine it better be GFP_NOFS even though the condition for that is
very unlikely, but that's what happens when you have tens of thousands nodes
all doing the same code all the time - all sorts of unlikely things trigger a lot).
> much better to keep using vzalloc{,_node)() in contexts that permit
> __GFP_FS for a higher likelihood of being able to allocate the memory.
While it's certainly possible to go audit all the OBD_ALLOC_LARGE and
isolate the ones where __GFP_FS is not detrimential, I just found yesterday that
vmalloc possibly does GFP_KERNEL allocations in its guts no matter what.
I saw all the rants and stuff about that too (but somewhat old).
Yet I cannot help but ask too if perhaps something could be done about it now?
>
>> Additionally fixed a typo in the macro name: VEROBSE->VERBOSE
>>
>> Signed-off-by: Bruno Faccini <bruno.faccini@xxxxxxxxx>
>> Signed-off-by: Oleg Drokin <oleg.drokin@xxxxxxxxx>
>> Reviewed-on: http://review.whamcloud.com/11190
>> Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-5349
>> ---
>> drivers/staging/lustre/lustre/include/obd_support.h | 18 ++++++++++++------
>> 1 file changed, 12 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/staging/lustre/lustre/include/obd_support.h b/drivers/staging/lustre/lustre/include/obd_support.h
>> index 2991d2e..c90a88e 100644
>> --- a/drivers/staging/lustre/lustre/include/obd_support.h
>> +++ b/drivers/staging/lustre/lustre/include/obd_support.h
>> @@ -655,11 +655,17 @@ do { \
>> #define OBD_CPT_ALLOC_PTR(ptr, cptab, cpt) \
>> OBD_CPT_ALLOC(ptr, cptab, cpt, sizeof(*(ptr)))
>>
>> -# define __OBD_VMALLOC_VEROBSE(ptr, cptab, cpt, size) \
>> +/* Direct use of __vmalloc_node() allows for protection flag specification
>> + * (and particularly to not set __GFP_FS, which is likely to cause some
>> + * deadlock situations in our code).
>> + */
>> +# define __OBD_VMALLOC_VERBOSE(ptr, cptab, cpt, size) \
>> do { \
>> - (ptr) = cptab == NULL ? \
>> - vzalloc(size) : \
>> - vzalloc_node(size, cfs_cpt_spread_node(cptab, cpt)); \
>> + (ptr) = __vmalloc_node(size, 1, GFP_NOFS | __GFP_HIGHMEM | __GFP_ZERO,\
>> + PAGE_KERNEL, \
>> + cptab == NULL ? NUMA_NO_NODE : \
>> + cfs_cpt_spread_node(cptab, cpt),\
>> + __builtin_return_address(0)); \
>> if (unlikely((ptr) == NULL)) { \
>> CERROR("vmalloc of '" #ptr "' (%d bytes) failed\n", \
>> (int)(size)); \
>> @@ -671,9 +677,9 @@ do { \
>> } while (0)
>>
>> # define OBD_VMALLOC(ptr, size) \
>> - __OBD_VMALLOC_VEROBSE(ptr, NULL, 0, size)
>> + __OBD_VMALLOC_VERBOSE(ptr, NULL, 0, size)
>> # define OBD_CPT_VMALLOC(ptr, cptab, cpt, size) \
>> - __OBD_VMALLOC_VEROBSE(ptr, cptab, cpt, size)
>> + __OBD_VMALLOC_VERBOSE(ptr, cptab, cpt, size)
>>
>>
>> /* Allocations above this size are considered too big and could not be done
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href