Re: [usb-storage] Re: cma: deadlock using usb-storage and fs

[Laura Abbott <labbott@redhat.com>]

On 12/17/18 10:29 AM, Gaël PORTAY wrote:
> Alan,
> 
> On Mon, Dec 17, 2018 at 10:45:17AM -0500, Alan Stern wrote:
>> On Sun, 16 Dec 2018, Gaël PORTAY wrote:
>> ...
>>
>>> The second task wants to writeback/flush the pages through USB, which, I
>>> assume, is due to the page migration. The usb-storage triggers a CMA allocation
>>> but get locked in cma_alloc since the first task hold the mutex (It is a FAT
>>> formatted partition, if it helps).
>>>
>>> 	usb-storage     D    0   349      2 0x00000000
>>> 	Backtrace:
>> ...
>>> 	[<bf1c7550>] (usb_sg_wait [usbcore]) from [<bf2bd618>] (usb_stor_bulk_transfer_sglist.part.2+0x80/0xdc [usb_storage]) r9:0001e000 r8:eca594ac r7:0001e000 r6:c0008200 r5:eca59514 r4:eca59488
>>
>> It looks like there is a logical problem in the CMA allocator.  The
>> call in usb_sg_wait() specifies GFP_NOIO, which is supposed to prevent
>> allocations from blocking on any I/O operations.  Therefore we
>> shouldn't be waiting for the CMA mutex.
>>
> 
> Right.
> 
>> Perhaps the CMA allocator needs to drop the mutex while doing
>> writebacks/flushes, or perhaps it needs to be reorganized some other
>> way.  I don't know anything about it.
>>
>> Does the CMA code have any maintainers who might need to know about
>> this, or is it all handled by the MM maintainers?
> 
> I did not find maintainers neither for CMA nor MM.
> 
> That is why I have sent this mail to mm mailing list but to no one in
> particular.
> 

Last time I looked at this, we needed the cma_mutex for serialization
so unless we want to rework that, I think we need to not use CMA in the
writeback case (i.e. GFP_IO).

The ARM dma layer uses gfpflags_allow_blocking to decide if it should
use CMA vs. the atomic pool:

static inline bool gfpflags_allow_blocking(const gfp_t gfp_flags)
{
         return !!(gfp_flags & __GFP_DIRECT_RECLAIM);
}

That's not sufficient to cover the writeback case. This is
used in multiple DMA allocations (arm64 and intel-iommu at
first pass) so I think we need a new gfpflags_allow_writeback
for deciding if CMA should be used.

Thanks,
Laura