[PATCH 04/15] mm: remove the pgmap field from struct hmm_vma_walk
rcampbell at nvidia.com
Fri Aug 16 21:10:47 UTC 2019
On 8/16/19 10:28 AM, Jason Gunthorpe wrote:
> On Fri, Aug 16, 2019 at 10:21:41AM -0700, Dan Williams wrote:
>>> We can do a get_dev_pagemap inside the page_walk and touch the pgmap,
>>> or we can do the 'device mutex && retry' pattern and touch the pgmap
>>> in the driver, under that lock.
>>> However in all cases the current get_dev_pagemap()'s in the page walk
>>> are not necessary, and we can delete them.
>> Yes, as long as 'struct page' instances resulting from that lookup are
>> not passed outside of that lock.
> Also, I was reflecting over lunch that the hmm_range_fault should only
> return DEVICE_PRIVATE pages for the caller's device (see other thread
> with HCH), and in this case, the caller should also be responsible to
> ensure that the driver is not calling hmm_range_fault at the same time
> it is deleting it's own DEVICE_PRIVATE mapping - ie by fencing its
> page fault handler.
Yes, that would make it a one step process to access another
device's migrated memory pages.
Right now, it has to be a two step process where the caller calls
hmm_range_fault, check the struct page to see if it is device
private and not owned, then call hmm_range_fault again with
range->pfns[i] |= range->flags[HMM_PFN_DEVICE_PRIVATE] to cause
the other device to migrate the page back to system memory.
More information about the amd-gfx