mbox series

[RFC,0/2] Ability to allocate DMAable pages using unpopulated-alloc

Message ID 1652810658-27810-1-git-send-email-olekstysh@gmail.com (mailing list archive)
Headers show
Series Ability to allocate DMAable pages using unpopulated-alloc | expand


Oleksandr Tyshchenko May 17, 2022, 6:04 p.m. UTC
From: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com>

Hello all.

The purpose of this RFC patch series is to get feedback about supporting the allocation
of DMAable pages by Linux's unpopulated-alloc.

The unpopulated-alloc feature has been enabled on Arm since the extended-regions support
reached upstream. With that (if, of course, we run new Xen version and Xen was able to
allocate extended regions), we don't allocate the real RAM pages from host memory and balloon
them out (in order to obtain physical memory space to map the guest pages into) anymore, we use
the unpopulated pages instead. And it seems that all users I have played with on Arm (I mean,
Xen PV and virtio backends) are happy with the pages provided by xen_alloc_unpopulated_pages().
It is worth mentioning that these pages are not contiguous, but this wasn't an issue so far.

There is one place, where we still steal RAM pages if user-space Xen PV backend tries
to establish grant mapping with a need to be backed by a DMA buffer for the sake of zero-copy
(see dma_alloc*() usage in gnttab_dma_alloc_pages()).

And, if I am not mistaken (there might be pitfalls which I am not aware of), we could avoid
wasting real RAM pages in that particular case also by adding an ability to allocate
unpopulated DMAable pages (which are guaranteed to be contiguous in IPA).
The benefit is quite clear here:
1. Avoid wasting real RAM pages (reducing the amount of CMA memory usable) for allocating
   physical memory space to map the granted buffer into (which can be big enough if
   we deal with Xen PV Display driver using multiple Full HD buffers) 
2. Avoid superpage shattering in Xen P2M when establishing stage-2 mapping for that
   granted buffer
3. Avoid extra operations needed for the granted buffer to be properly mapped and
   unmapped such as ballooning in/out real RAM pages

Please note, there are several TODOs (described in corresponding commit subjects),
which I will try to eliminate in next versions if we find a common ground regarding
the approach.

Any feedback/help would be highly appreciated.

Oleksandr Tyshchenko (2):
  xen/unpopulated-alloc: Introduce helpers for DMA allocations
  xen/grant-table: Use unpopulated DMAable pages instead of real RAM

 drivers/xen/grant-table.c       |  27 +++++++
 drivers/xen/unpopulated-alloc.c | 167 ++++++++++++++++++++++++++++++++++++++++
 include/xen/xen.h               |  15 ++++
 3 files changed, 209 insertions(+)