lkml.org 
[lkml]   [2019]   [Jul]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] staging: kpc2000: Convert put_page() to put_user_page*()
On Mon, Jul 15, 2019 at 03:01:43PM -0700, John Hubbard wrote:
> On 7/15/19 2:47 PM, Matt Sickler wrote:
> > It looks like Outlook is going to absolutely trash this email. Hopefully it comes through okay.
> >
> ...
> >>
> >> Because this is a common pattern, and because the code here doesn't likely
> >> need to set page dirty before the dma_unmap_sg call, I think the following
> >> would be better (it's untested), instead of the above diff hunk:
> >>
> >> diff --git a/drivers/staging/kpc2000/kpc_dma/fileops.c
> >> b/drivers/staging/kpc2000/kpc_dma/fileops.c
> >> index 48ca88bc6b0b..d486f9866449 100644
> >> --- a/drivers/staging/kpc2000/kpc_dma/fileops.c
> >> +++ b/drivers/staging/kpc2000/kpc_dma/fileops.c
> >> @@ -211,16 +211,13 @@ void transfer_complete_cb(struct aio_cb_data
> >> *acd, size_t xfr_count, u32 flags)
> >> BUG_ON(acd->ldev == NULL);
> >> BUG_ON(acd->ldev->pldev == NULL);
> >>
> >> - for (i = 0 ; i < acd->page_count ; i++) {
> >> - if (!PageReserved(acd->user_pages[i])) {
> >> - set_page_dirty(acd->user_pages[i]);
> >> - }
> >> - }
> >> -
> >> dma_unmap_sg(&acd->ldev->pldev->dev, acd->sgt.sgl, acd->sgt.nents, acd->ldev->dir);
> >>
> >> for (i = 0 ; i < acd->page_count ; i++) {
> >> - put_page(acd->user_pages[i]);
> >> + if (!PageReserved(acd->user_pages[i])) {
> >> + put_user_pages_dirty(&acd->user_pages[i], 1);
> >> + else
> >> + put_user_page(acd->user_pages[i]);
> >> }
> >>
> >> sg_free_table(&acd->sgt);
> >
> > I don't think I ever really knew the right way to do this.
> >
> > The changes Bharath suggested look okay to me. I'm not sure about the check for PageReserved(), though. At first glance it appears to be equivalent to what was there before, but maybe I should learn what that Reserved page flag really means.
> > From [1], the only comment that seems applicable is
> > * - MMIO/DMA pages. Some architectures don't allow to ioremap pages that are
> > * not marked PG_reserved (as they might be in use by somebody else who does
> > * not respect the caching strategy).
> >
> > These pages should be coming from anonymous (RAM, not file backed) memory in userspace. Sometimes it comes from hugepage backed memory, though I don't think that makes a difference. I should note that transfer_complete_cb handles both RAM to device and device to RAM DMAs, if that matters.
Yes. file_operations->read passes a userspace buffer which AFAIK is
anonymous memory.
> > [1] https://elixir.bootlin.com/linux/v5.2/source/include/linux/page-flags.h#L17
> >
>
> I agree: the PageReserved check looks unnecessary here, from my outside-the-kpc_2000-team
> perspective, anyway. Assuming that your analysis above is correct, you could collapse that
> whole think into just:
Since the file_operations->read passes a userspace buffer, I doubt that
the pages of the userspace buffer will be reserved.
> @@ -211,17 +209,8 @@ void transfer_complete_cb(struct aio_cb_data *acd, size_t xfr_count, u32 flags)
> BUG_ON(acd->ldev == NULL);
> BUG_ON(acd->ldev->pldev == NULL);
>
> - for (i = 0 ; i < acd->page_count ; i++) {
> - if (!PageReserved(acd->user_pages[i])) {
> - set_page_dirty(acd->user_pages[i]);
> - }
> - }
> -
> dma_unmap_sg(&acd->ldev->pldev->dev, acd->sgt.sgl, acd->sgt.nents, acd->ldev->dir);
> -
> - for (i = 0 ; i < acd->page_count ; i++) {
> - put_page(acd->user_pages[i]);
> - }
> + put_user_pages_dirty(&acd->user_pages[i], acd->page_count);
>
> sg_free_table(&acd->sgt);
>
> (Also, Matt, I failed to Cc: you on a semi-related cleanup that I just sent out for this
> driver, as long as I have your attention:
>
> https://lore.kernel.org/r/20190715212123.432-1-jhubbard@nvidia.com
> )
Matt will you be willing to pick this up for testing or do you want a
different patch?
> thanks,
> --
> John Hubbard
> NVIDIA
Thank you
Bharath

\
 
 \ /
  Last update: 2019-07-16 12:29    [W:0.044 / U:45.052 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site