On 4/16/20 12:02 PM, Dave Hansen wrote:
> On 4/16/20 9:34 AM, Claudio Imbrenda wrote:
>>> Ahh, so this is *just* intended to precede I/O done on the page, when
>>> a non-host entity is touching the memory?
>> yep
> OK, so we've got to do an action that precedes *all* I/O to a page.
> That's not too bad.
>
> I still don't understand how this could work generally, though There
> are lots of places where I/O is done to a page without either going
> through __test_set_page_writeback() or gup() with FOLL_PIN set.
>
> sendfile() is probably the best example of this:
>
> fd = open("/normal/ext4/file", O_RDONLY);
> sendfile(socket_fd, fd, &off, count);
>
> There's no gup in sight since the file doesn't have an address and it's
> not being written to so there's no writeback.
>
> How does sendfile work?
Did you manage to see if sendfile works (or any other operation that
DMAs file-backed data without being preceded by a gup)?
I suspect it's actually not that hard to fix. As long as you have a
dma_ops for the devices in question either via dev->dma_ops or you add
an s390 get_arch_vm_ops(), you can fix *all* the DMA sites, sendfile()
included.
BTW, device drivers do need to know how to use the DMA mapping API. If
s390 has drivers that need to be updated, I think that's vastly
preferable to incomplete hooks in core mm code.
