Re: [PATCH 3/4] scsi: ipr: use kmalloc() to allocate IPR dump buffer memory

[Brian King]

On 7/2/26 1:55 AM, Mike Rapoport wrote:
> On Wed, Jul 01, 2026 at 04:03:48PM -0500, Brian King wrote:
>> On 7/1/26 4:52 AM, Mike Rapoport wrote:
>>> On Wed, Jul 01, 2026 at 09:03:06AM +0200, Hannes Reinecke wrote:
>>>> On 6/30/26 12:54 PM, Mike Rapoport (Microsoft) wrote:
>>>>> IPR dump machinery allocates memory to save adapter's crash dump using
>>>>> __get_free_page().
>>>>>
>>>>> This memory can be allocated with kmalloc() as there's nothing special
>>>>> about it to go directly to the page allocator.
>>>>>
>>>>> kmalloc() provides a better API that does not require ugly casts and
>>>>> kfree() does not need to know the size of the freed object.
>>>>>
>>>>> Replace use of __get_free_page() with kmalloc().
>>>>>
>>>>> Link: https://lore.kernel.org/all/635405e4-9423-4a25-a6e7-e03c8ea0bcbe@xxxxxxxxxx
>>>>> Signed-off-by: Mike Rapoport (Microsoft) <rppt@xxxxxxxxxx>
>>>>> ---
>>>>>   drivers/scsi/ipr.c | 4 ++--
>>>>>   1 file changed, 2 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/drivers/scsi/ipr.c b/drivers/scsi/ipr.c
>>>>> index d207e5e81afe..5a212bfdeec2 100644
>>>>> --- a/drivers/scsi/ipr.c
>>>>> +++ b/drivers/scsi/ipr.c
>>>>> @@ -2893,7 +2893,7 @@ static int ipr_sdt_copy(struct ipr_ioa_cfg *ioa_cfg,
>>>>>   	       (ioa_dump->hdr.len + bytes_copied) < max_dump_size) {
>>>>>   		if (ioa_dump->page_offset >= PAGE_SIZE ||
>>>>>   		    ioa_dump->page_offset == 0) {
>>>>> -			page = (__be32 *)__get_free_page(GFP_ATOMIC);
>>>>> +			page = kmalloc(PAGE_SIZE, GFP_ATOMIC);
>>>>>   			if (!page) {
>>>>>   				ipr_trace;
>>>>> @@ -3226,7 +3226,7 @@ static void ipr_release_dump(struct kref *kref)
>>>>>   	spin_unlock_irqrestore(ioa_cfg->host->host_lock, lock_flags);
>>>>>   	for (i = 0; i < dump->ioa_dump.next_page_index; i++)
>>>>> -		free_page((unsigned long) dump->ioa_dump.ioa_data[i]);
>>>>> +		kfree(dump->ioa_dump.ioa_data[i]);
>>>>>   	vfree(dump->ioa_dump.ioa_data);
>>>>>   	kfree(dump);
>>>>>
>>>>
>>>> I _think_ we can replace this with kvmalloc, and allocate the entire
>>>> dump buffer in one go. Once switched to kmalloc() it's kinda pointless
>>>> to allocate separate page-sized buffers here.
>>>
>>> kmalloc() performance is on par with __get_free_page(), but kvmalloc()
>>> would be slower if it falls back to vmalloc(). 
>>>
>>> I'm not familiar with the driver to say if this could be an issue here.
>>
>> This code only runs when the adapter has hit a fatal error, so should be
>> extremely rare. The memory is getting allocated while the storage adapter
>> is in a failed state, so anything running on the system at the time could
>> be stalled until recovery is completed. This memory is allocated and should
>> be freed soon after the adapter recovers. In order for this code to
>> run, the iprdump daemon must be running, which will then read out the dump
>> after the adapter is recovered, and write it to disk, after which time, the
>> ipr driver will free the kernel memory.
> 
> Thanks for the explanation!
> 
> So the allocation is not on the hot path and it seems to me that
> performance differences between kmalloc() and slower vmalloc() won't move a
> needle relatively to the entire dump procedure.
> 
> And it seems that GFP_ATOMIC here is a historical artifact rather than
> actual necessity. ipr_get_ioa_dump() runs in a workqueue context without
> locks held so nothing warrants GFP_ATOMIC.

Agree that GFP_ATOMIC is not needed here, but we probably do need
GFP_NOIO since we don't want these allocations starting I/O as the
storage adapter is not in a state where it can handle I/O until
the card is reset.

Thanks,

Brian


-- 
Brian King
Power Linux I/O
IBM Linux Technology Center