[v6,0/9] Huge page-table entries for TTM

Message

Thomas Hellström (Intel) March 4, 2020, 10:28 a.m. UTC

In order to reduce CPU usage [1] and in theory TLB misses this patchset enables
huge- and giant page-table entries for TTM and TTM-enabled graphics drivers.

Patch 1 and 2 introduce a vma_is_special_huge() function to make the mm code
take the same path as DAX when splitting huge- and giant page table entries,
(which currently means zapping the page-table entry and rely on re-faulting).

Patch 3 makes the mm code split existing huge page-table entries
on huge_fault fallbacks. Typically on COW or on buffer-objects that want
write-notify. COW and write-notification is always done on the lowest
page-table level. See the patch log message for additional considerations.

Patch 4 introduces functions to allow the graphics drivers to manipulate
the caching- and encryption flags of huge page-table entries without ugly
hacks.

Patch 5 implements the huge_fault handler in TTM.
This enables huge page-table entries, provided that the kernel is configured
to support transhuge pages, either by default or using madvise().
However, they are unlikely to be inserted unless the kernel buffer object
pfns and user-space addresses align perfectly. There are various options
here, but since buffer objects that reside in system pages typically start
at huge page boundaries if they are backed by huge pages, we try to enforce
buffer object starting pfns and user-space addresses to be huge page-size
aligned if their size exceeds a huge page-size. If pud-size transhuge
("giant") pages are enabled by the arch, the same holds for those.

Patch 6 implements a specialized huge_fault handler for vmwgfx.
The vmwgfx driver may perform dirty-tracking and needs some special code
to handle that correctly.

Patch 7 implements a drm helper to align user-space addresses according
to the above scheme, if possible.

Patch 8 implements a TTM range manager for vmwgfx that does the same for
graphics IO memory. This may later be reused by other graphics drivers
if necessary.

Patch 9 finally hooks up the helpers of patch 7 and 8 to the vmwgfx driver.
A similar change is needed for graphics drivers that want a reasonable
likelyhood of actually using huge page-table entries.

If a buffer object size is not huge-page or giant-page aligned,
its size will NOT be inflated by this patchset. This means that the buffer
object tail will use smaller size page-table entries and thus no memory
overhead occurs. Drivers that want to pay the memory overhead price need to
implement their own scheme to inflate buffer-object sizes.

PMD size huge page-table-entries have been tested with vmwgfx and found to
work well both with system memory backed and IO memory backed buffer objects.

PUD size giant page-table-entries have seen limited (fault and COW) testing
using a modified kernel (to support 1GB page allocations) and a fake vmwgfx
TTM memory type. The vmwgfx driver does otherwise not support 1GB-size IO
memory resources.

Comments and suggestions welcome.
Thomas

Changes since RFC:
* Check for buffer objects present in contigous IO Memory (Christian König)
* Rebased on the vmwgfx emulated coherent memory functionality. That rebase
  adds patch 5.
Changes since v1:
* Make the new TTM range manager vmwgfx-specific. (Christian König)
* Minor fixes for configs that don't support or only partially support
  transhuge pages.
Changes since v2:
* Minor coding style and doc fixes in patch 5/9 (Christian König)
* Patch 5/9 doesn't touch mm. Remove from the patch title.
Changes since v3:
* Added reviews and acks
* Implemented ugly but generic ttm_pgprot_is_wrprotecting() instead of arch
  specific code.
Changes since v4:
* Added timings (Andrew Morton)
* Updated function documentation (Andrew Morton)
Changes since v6:
* Fix drm build error with !CONFIG_MMU

[1]
The below test program generates the following gnu time output when run on a
vmwgfx-enabled kernel without the patch series:

4.78user 6.02system 0:10.91elapsed 99%CPU (0avgtext+0avgdata 1624maxresident)k
0inputs+0outputs (0major+640077minor)pagefaults 0swaps

and with the patch series:

1.71user 3.60system 0:05.40elapsed 98%CPU (0avgtext+0avgdata 1656maxresident)k
0inputs+0outputs (0major+20079minor)pagefaults 0swaps

A consistent number of reduced graphics page-faults can be seen with normal
graphics applications, but due to the aggressive buffer object caching in
vmwgfx user-space drivers the CPU time reduction is within the error marginal.

#include <unistd.h>
#include <string.h>
#include <sys/mman.h>
#include <xf86drm.h>

static void checkerr(int ret, const char *name)
{
  if (ret < 0) {
    perror(name);
    exit(-1);
  }
}

int main(int agc, const char *argv[])
{
    struct drm_mode_create_dumb c_arg = {0};
    struct drm_mode_map_dumb m_arg = {0};
    struct drm_mode_destroy_dumb d_arg = {0};
    int ret, i, fd;
    void *map;

    fd = open("/dev/dri/card0", O_RDWR);
    checkerr(fd, argv[0]);

    for (i = 0; i < 10000; ++i) {
      c_arg.bpp = 32;
      c_arg.width = 1024;
      c_arg.height = 1024;      
      ret = drmIoctl(fd, DRM_IOCTL_MODE_CREATE_DUMB, &c_arg);
      checkerr(fd, argv[0]);

      m_arg.handle = c_arg.handle;
      ret = drmIoctl(fd, DRM_IOCTL_MODE_MAP_DUMB, &m_arg);
      checkerr(fd, argv[0]);
      
      map = mmap(NULL, c_arg.size, PROT_READ | PROT_WRITE, MAP_SHARED, fd,
	       m_arg.offset);
      checkerr(map == MAP_FAILED ? -1 : 0, argv[0]);

      (void) madvise((void *) map, c_arg.size, MADV_HUGEPAGE);
      memset(map, 0x67, c_arg.size);
      munmap(map, c_arg.size);

      d_arg.handle = c_arg.handle;
      ret = drmIoctl(fd, DRM_IOCTL_MODE_DESTROY_DUMB, &d_arg);
      checkerr(ret, argv[0]);
    }
    
    close(fd);
}

Cc: Andrew Morton <[e-mail]>
Cc: Michal Hocko <[e-mail]>
Cc: "Matthew Wilcox (Oracle)" <[e-mail]>
Cc: "Kirill A. Shutemov" <[e-mail]>
Cc: Ralph Campbell <[e-mail]>
Cc: "Jérôme Glisse" <[e-mail]>
Cc: "Christian König" <[e-mail]>
Cc: Dan Williams <[e-mail]>

Comments

Thomas Hellström (Intel) March 16, 2020, 12:32 p.m. UTC | #1

On 3/4/20 11:28 AM, Thomas Hellström (VMware) wrote:
> In order to reduce CPU usage [1] and in theory TLB misses this patchset enables
> huge- and giant page-table entries for TTM and TTM-enabled graphics drivers.
>
> Patch 1 and 2 introduce a vma_is_special_huge() function to make the mm code
> take the same path as DAX when splitting huge- and giant page table entries,
> (which currently means zapping the page-table entry and rely on re-faulting).
>
> Patch 3 makes the mm code split existing huge page-table entries
> on huge_fault fallbacks. Typically on COW or on buffer-objects that want
> write-notify. COW and write-notification is always done on the lowest
> page-table level. See the patch log message for additional considerations.
>
> Patch 4 introduces functions to allow the graphics drivers to manipulate
> the caching- and encryption flags of huge page-table entries without ugly
> hacks.
>
> Patch 5 implements the huge_fault handler in TTM.
> This enables huge page-table entries, provided that the kernel is configured
> to support transhuge pages, either by default or using madvise().
> However, they are unlikely to be inserted unless the kernel buffer object
> pfns and user-space addresses align perfectly. There are various options
> here, but since buffer objects that reside in system pages typically start
> at huge page boundaries if they are backed by huge pages, we try to enforce
> buffer object starting pfns and user-space addresses to be huge page-size
> aligned if their size exceeds a huge page-size. If pud-size transhuge
> ("giant") pages are enabled by the arch, the same holds for those.
>
> Patch 6 implements a specialized huge_fault handler for vmwgfx.
> The vmwgfx driver may perform dirty-tracking and needs some special code
> to handle that correctly.
>
> Patch 7 implements a drm helper to align user-space addresses according
> to the above scheme, if possible.
>
> Patch 8 implements a TTM range manager for vmwgfx that does the same for
> graphics IO memory. This may later be reused by other graphics drivers
> if necessary.
>
> Patch 9 finally hooks up the helpers of patch 7 and 8 to the vmwgfx driver.
> A similar change is needed for graphics drivers that want a reasonable
> likelyhood of actually using huge page-table entries.
>
> If a buffer object size is not huge-page or giant-page aligned,
> its size will NOT be inflated by this patchset. This means that the buffer
> object tail will use smaller size page-table entries and thus no memory
> overhead occurs. Drivers that want to pay the memory overhead price need to
> implement their own scheme to inflate buffer-object sizes.
>
> PMD size huge page-table-entries have been tested with vmwgfx and found to
> work well both with system memory backed and IO memory backed buffer objects.
>
> PUD size giant page-table-entries have seen limited (fault and COW) testing
> using a modified kernel (to support 1GB page allocations) and a fake vmwgfx
> TTM memory type. The vmwgfx driver does otherwise not support 1GB-size IO
> memory resources.
>
> Comments and suggestions welcome.
> Thomas
>
> Changes since RFC:
> * Check for buffer objects present in contigous IO Memory (Christian König)
> * Rebased on the vmwgfx emulated coherent memory functionality. That rebase
>    adds patch 5.
> Changes since v1:
> * Make the new TTM range manager vmwgfx-specific. (Christian König)
> * Minor fixes for configs that don't support or only partially support
>    transhuge pages.
> Changes since v2:
> * Minor coding style and doc fixes in patch 5/9 (Christian König)
> * Patch 5/9 doesn't touch mm. Remove from the patch title.
> Changes since v3:
> * Added reviews and acks
> * Implemented ugly but generic ttm_pgprot_is_wrprotecting() instead of arch
>    specific code.
> Changes since v4:
> * Added timings (Andrew Morton)
> * Updated function documentation (Andrew Morton)
> Changes since v6:
> * Fix drm build error with !CONFIG_MMU
>
> [1]
> The below test program generates the following gnu time output when run on a
> vmwgfx-enabled kernel without the patch series:
>
> 4.78user 6.02system 0:10.91elapsed 99%CPU (0avgtext+0avgdata 1624maxresident)k
> 0inputs+0outputs (0major+640077minor)pagefaults 0swaps
>
> and with the patch series:
>
> 1.71user 3.60system 0:05.40elapsed 98%CPU (0avgtext+0avgdata 1656maxresident)k
> 0inputs+0outputs (0major+20079minor)pagefaults 0swaps
>
> A consistent number of reduced graphics page-faults can be seen with normal
> graphics applications, but due to the aggressive buffer object caching in
> vmwgfx user-space drivers the CPU time reduction is within the error marginal.
>
> #include <unistd.h>
> #include <string.h>
> #include <sys/mman.h>
> #include <xf86drm.h>
>
> static void checkerr(int ret, const char *name)
> {
>    if (ret < 0) {
>      perror(name);
>      exit(-1);
>    }
> }
>
> int main(int agc, const char *argv[])
> {
>      struct drm_mode_create_dumb c_arg = {0};
>      struct drm_mode_map_dumb m_arg = {0};
>      struct drm_mode_destroy_dumb d_arg = {0};
>      int ret, i, fd;
>      void *map;
>
>      fd = open("/dev/dri/card0", O_RDWR);
>      checkerr(fd, argv[0]);
>
>      for (i = 0; i < 10000; ++i) {
>        c_arg.bpp = 32;
>        c_arg.width = 1024;
>        c_arg.height = 1024;
>        ret = drmIoctl(fd, DRM_IOCTL_MODE_CREATE_DUMB, &c_arg);
>        checkerr(fd, argv[0]);
>
>        m_arg.handle = c_arg.handle;
>        ret = drmIoctl(fd, DRM_IOCTL_MODE_MAP_DUMB, &m_arg);
>        checkerr(fd, argv[0]);
>        
>        map = mmap(NULL, c_arg.size, PROT_READ | PROT_WRITE, MAP_SHARED, fd,
> 	       m_arg.offset);
>        checkerr(map == MAP_FAILED ? -1 : 0, argv[0]);
>
>        (void) madvise((void *) map, c_arg.size, MADV_HUGEPAGE);
>        memset(map, 0x67, c_arg.size);
>        munmap(map, c_arg.size);
>
>        d_arg.handle = c_arg.handle;
>        ret = drmIoctl(fd, DRM_IOCTL_MODE_DESTROY_DUMB, &d_arg);
>        checkerr(ret, argv[0]);
>      }
>      
>      close(fd);
> }
>
> Cc: Andrew Morton <[e-mail]>
> Cc: Michal Hocko <[e-mail]>
> Cc: "Matthew Wilcox (Oracle)" <[e-mail]>
> Cc: "Kirill A. Shutemov" <[e-mail]>
> Cc: Ralph Campbell <[e-mail]>
> Cc: "Jérôme Glisse" <[e-mail]>
> Cc: "Christian König" <[e-mail]>
> Cc: Dan Williams <[e-mail]>
>
>
> _______________________________________________
> dri-devel mailing list
> [e-mail]
> https://lists.freedesktop.org/mailman/listinfo/dri-devel

Andrew, would it be possible to have an ack for merge using a DRM tree 
for the -mm patches?

Thanks,

Thomas

Andrew Morton March 18, 2020, 11:27 p.m. UTC | #2

On Mon, 16 Mar 2020 13:32:08 +0100 Thomas Hellström (VMware) <[e-mail]> wrote:

> > _______________________________________________
> > dri-devel mailing list
> > [e-mail]
> > https://lists.freedesktop.org/mailman/listinfo/dri-devel
> 
> Andrew, would it be possible to have an ack for merge using a DRM tree 
> for the -mm patches?

Yes, please do.  It's all pretty straightforward addition of new
functionality which won't affect existing code.

Thomas Hellström (Intel) March 19, 2020, 10:20 a.m. UTC | #3

On 3/19/20 12:27 AM, Andrew Morton wrote:
> On Mon, 16 Mar 2020 13:32:08 +0100 Thomas Hellström (VMware) <[e-mail]> wrote:
>
>>> _______________________________________________
>>> dri-devel mailing list
>>> [e-mail]
>>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>> Andrew, would it be possible to have an ack for merge using a DRM tree
>> for the -mm patches?
> Yes, please do.  It's all pretty straightforward addition of new
> functionality which won't affect existing code.

Thanks Andrew. Can I add your Acked-by: To the mm patches for Linus' 
reference?

Thanks,

Thomas

Andrew Morton March 21, 2020, 1:58 a.m. UTC | #4

On Thu, 19 Mar 2020 11:20:44 +0100 Thomas Hellström (VMware) <[e-mail]> wrote:

> On 3/19/20 12:27 AM, Andrew Morton wrote:
> > On Mon, 16 Mar 2020 13:32:08 +0100 Thomas Hellström (VMware) <[e-mail]> wrote:
> >
> >>> _______________________________________________
> >>> dri-devel mailing list
> >>> [e-mail]
> >>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
> >> Andrew, would it be possible to have an ack for merge using a DRM tree
> >> for the -mm patches?
> > Yes, please do.  It's all pretty straightforward addition of new
> > functionality which won't affect existing code.
> 
> Thanks Andrew. Can I add your Acked-by: To the mm patches for Linus' 
> reference?
> 

Please do.