x86 Hyper-V used to essentially always overwrite the effective cache type
of guest memory accesses to WB. This was problematic in cases where there
is a physical device assigned to the VM, since that often requires that
the VM should have control over cache types. Thus, on newer Hyper-V since
2018, Hyper-V always honors the VM's cache type, but unexpectedly Linux VM
users start to complain that Linux VM's VRAM becomes very slow, and it
turns out that Linux VM should not map the VRAM uncacheable by ioremap().
Fix this slowness issue by using ioremap_cache().
On ARM64, ioremap_cache() is also required as the host also maps the VRAM
cacheable, otherwise VM Connect can't display properly with ioremap() or
ioremap_wc().
With this change, the VRAM on new Hyper-V is as fast as regular RAM, so
it's no longer necessary to use the hacks we added to mitigate the
slowness, i.e. we no longer need to allocate physical memory and use
it to back up the VRAM in Generation-1 VM, and we also no longer need to
allocate physical memory to back up the framebuffer in a Generation-2 VM
and copy the framebuffer to the real VRAM. A further big change will
address these for v5.11.
Fixes: 68a2d20b79b1 ("drivers/video: add Hyper-V Synthetic Video Frame Buffer Driver")
Tested-by: Boqun Feng <[email protected]>
Signed-off-by: Dexuan Cui <[email protected]>
---
Hi Wei Liu, can you please pick this up into the hyperv/linux.git tree's
hyperv-fixes branch? I really hope this patch can be in v5.10 since it
fixes a longstanding issue: https://github.com/LIS/lis-next/issues/655
drivers/video/fbdev/hyperv_fb.c | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)
diff --git a/drivers/video/fbdev/hyperv_fb.c b/drivers/video/fbdev/hyperv_fb.c
index 5bc86f481a78..c8b0ae676809 100644
--- a/drivers/video/fbdev/hyperv_fb.c
+++ b/drivers/video/fbdev/hyperv_fb.c
@@ -1093,7 +1093,12 @@ static int hvfb_getmem(struct hv_device *hdev, struct fb_info *info)
goto err1;
}
- fb_virt = ioremap(par->mem->start, screen_fb_size);
+ /*
+ * Map the VRAM cacheable for performance. This is also required for
+ * VM Connect to display properly for ARM64 Linux VM, as the host also
+ * maps the VRAM cacheable.
+ */
+ fb_virt = ioremap_cache(par->mem->start, screen_fb_size);
if (!fb_virt)
goto err2;
--
2.19.1
From: Dexuan Cui <[email protected]> Sent: Tuesday, November 17, 2020 4:03 PM
>
> x86 Hyper-V used to essentially always overwrite the effective cache type
> of guest memory accesses to WB. This was problematic in cases where there
> is a physical device assigned to the VM, since that often requires that
> the VM should have control over cache types. Thus, on newer Hyper-V since
> 2018, Hyper-V always honors the VM's cache type, but unexpectedly Linux VM
> users start to complain that Linux VM's VRAM becomes very slow, and it
> turns out that Linux VM should not map the VRAM uncacheable by ioremap().
> Fix this slowness issue by using ioremap_cache().
>
> On ARM64, ioremap_cache() is also required as the host also maps the VRAM
> cacheable, otherwise VM Connect can't display properly with ioremap() or
> ioremap_wc().
>
> With this change, the VRAM on new Hyper-V is as fast as regular RAM, so
> it's no longer necessary to use the hacks we added to mitigate the
> slowness, i.e. we no longer need to allocate physical memory and use
> it to back up the VRAM in Generation-1 VM, and we also no longer need to
> allocate physical memory to back up the framebuffer in a Generation-2 VM
> and copy the framebuffer to the real VRAM. A further big change will
> address these for v5.11.
>
> Fixes: 68a2d20b79b1 ("drivers/video: add Hyper-V Synthetic Video Frame Buffer Driver")
> Tested-by: Boqun Feng <[email protected]>
> Signed-off-by: Dexuan Cui <[email protected]>
> ---
>
> Hi Wei Liu, can you please pick this up into the hyperv/linux.git tree's
> hyperv-fixes branch? I really hope this patch can be in v5.10 since it
> fixes a longstanding issue: https://github.com/LIS/lis-next/issues/655
>
> drivers/video/fbdev/hyperv_fb.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/video/fbdev/hyperv_fb.c b/drivers/video/fbdev/hyperv_fb.c
> index 5bc86f481a78..c8b0ae676809 100644
> --- a/drivers/video/fbdev/hyperv_fb.c
> +++ b/drivers/video/fbdev/hyperv_fb.c
> @@ -1093,7 +1093,12 @@ static int hvfb_getmem(struct hv_device *hdev, struct fb_info
> *info)
> goto err1;
> }
>
> - fb_virt = ioremap(par->mem->start, screen_fb_size);
> + /*
> + * Map the VRAM cacheable for performance. This is also required for
> + * VM Connect to display properly for ARM64 Linux VM, as the host also
> + * maps the VRAM cacheable.
> + */
> + fb_virt = ioremap_cache(par->mem->start, screen_fb_size);
> if (!fb_virt)
> goto err2;
>
> --
> 2.19.1
Reviewed-by: Michael Kelley <[email protected]>
On Wed, Nov 18, 2020 at 12:20:11AM +0000, Michael Kelley wrote:
> From: Dexuan Cui <[email protected]> Sent: Tuesday, November 17, 2020 4:03 PM
> >
> > x86 Hyper-V used to essentially always overwrite the effective cache type
> > of guest memory accesses to WB. This was problematic in cases where there
> > is a physical device assigned to the VM, since that often requires that
> > the VM should have control over cache types. Thus, on newer Hyper-V since
> > 2018, Hyper-V always honors the VM's cache type, but unexpectedly Linux VM
> > users start to complain that Linux VM's VRAM becomes very slow, and it
> > turns out that Linux VM should not map the VRAM uncacheable by ioremap().
> > Fix this slowness issue by using ioremap_cache().
> >
> > On ARM64, ioremap_cache() is also required as the host also maps the VRAM
> > cacheable, otherwise VM Connect can't display properly with ioremap() or
> > ioremap_wc().
> >
> > With this change, the VRAM on new Hyper-V is as fast as regular RAM, so
> > it's no longer necessary to use the hacks we added to mitigate the
> > slowness, i.e. we no longer need to allocate physical memory and use
> > it to back up the VRAM in Generation-1 VM, and we also no longer need to
> > allocate physical memory to back up the framebuffer in a Generation-2 VM
> > and copy the framebuffer to the real VRAM. A further big change will
> > address these for v5.11.
> >
> > Fixes: 68a2d20b79b1 ("drivers/video: add Hyper-V Synthetic Video Frame Buffer Driver")
> > Tested-by: Boqun Feng <[email protected]>
> > Signed-off-by: Dexuan Cui <[email protected]>
> Reviewed-by: Michael Kelley <[email protected]>
>
Applied to hyperv-fixes. Thanks.
> -----Original Message-----
> From: Dexuan Cui <[email protected]>
> Sent: Tuesday, November 17, 2020 7:03 PM
> To: KY Srinivasan <[email protected]>; Haiyang Zhang
> <[email protected]>; Stephen Hemminger
> <[email protected]>; [email protected];
> [email protected]; [email protected]; dri-
> [email protected]; [email protected]; linux-
> [email protected]; Michael Kelley <[email protected]>
> Cc: Wei Hu <[email protected]>; Dexuan Cui <[email protected]>
> Subject: [PATCH] video: hyperv_fb: Fix the cache type when mapping the
> VRAM
>
> x86 Hyper-V used to essentially always overwrite the effective cache type of
> guest memory accesses to WB. This was problematic in cases where there is
> a physical device assigned to the VM, since that often requires that the VM
> should have control over cache types. Thus, on newer Hyper-V since 2018,
> Hyper-V always honors the VM's cache type, but unexpectedly Linux VM
> users start to complain that Linux VM's VRAM becomes very slow, and it
> turns out that Linux VM should not map the VRAM uncacheable by ioremap().
> Fix this slowness issue by using ioremap_cache().
>
> On ARM64, ioremap_cache() is also required as the host also maps the VRAM
> cacheable, otherwise VM Connect can't display properly with ioremap() or
> ioremap_wc().
>
> With this change, the VRAM on new Hyper-V is as fast as regular RAM, so it's
> no longer necessary to use the hacks we added to mitigate the slowness, i.e.
> we no longer need to allocate physical memory and use it to back up the
> VRAM in Generation-1 VM, and we also no longer need to allocate physical
> memory to back up the framebuffer in a Generation-2 VM and copy the
> framebuffer to the real VRAM. A further big change will address these for
> v5.11.
>
> Fixes: 68a2d20b79b1 ("drivers/video: add Hyper-V Synthetic Video Frame
> Buffer Driver")
> Tested-by: Boqun Feng <[email protected]>
> Signed-off-by: Dexuan Cui <[email protected]>
> ---
>
> Hi Wei Liu, can you please pick this up into the hyperv/linux.git tree's hyperv-
> fixes branch? I really hope this patch can be in v5.10 since it fixes a
> longstanding issue:
> https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgith
> ub.com%2FLIS%2Flis-
> next%2Fissues%2F655&data=04%7C01%7Chaiyangz%40microsoft.com%
> 7C7e371bb6f79f41aae12208d88b556c85%7C72f988bf86f141af91ab2d7cd011d
> b47%7C1%7C0%7C637412546297591335%7CUnknown%7CTWFpbGZsb3d8eyJ
> WIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%
> 7C1000&sdata=StqnT%2Fx1XVoVWUZbJz5BNjaCIdtuNmSf2JoyLSt0c%2B
> Q%3D&reserved=0
>
> drivers/video/fbdev/hyperv_fb.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/video/fbdev/hyperv_fb.c
> b/drivers/video/fbdev/hyperv_fb.c index 5bc86f481a78..c8b0ae676809
> 100644
> --- a/drivers/video/fbdev/hyperv_fb.c
> +++ b/drivers/video/fbdev/hyperv_fb.c
> @@ -1093,7 +1093,12 @@ static int hvfb_getmem(struct hv_device *hdev,
> struct fb_info *info)
> goto err1;
> }
>
> - fb_virt = ioremap(par->mem->start, screen_fb_size);
> + /*
> + * Map the VRAM cacheable for performance. This is also required
> for
> + * VM Connect to display properly for ARM64 Linux VM, as the host
> also
> + * maps the VRAM cacheable.
> + */
> + fb_virt = ioremap_cache(par->mem->start, screen_fb_size);
> if (!fb_virt)
> goto err2;
Reviewed-by: Haiyang Zhang <[email protected]>
Thank you!