Re: [PATCH RFC 4/4] vhost-vdpa: introduce IOTLB_PERSIST backend feature bit

From: Si-Wei Liu
Date: Mon Aug 28 2023 - 19:47:10 EST



On 8/22/2023 1:54 AM, Jason Wang wrote:
On Thu, Aug 17, 2023 at 7:44 AM Si-Wei Liu <si-wei.liu@xxxxxxxxxx> wrote:


On 8/15/2023 6:48 PM, Jason Wang wrote:
On Wed, Aug 16, 2023 at 6:31 AM Si-Wei Liu <si-wei.liu@xxxxxxxxxx> wrote:

On 8/14/2023 7:25 PM, Jason Wang wrote:
On Tue, Aug 15, 2023 at 9:45 AM Si-Wei Liu <si-wei.liu@xxxxxxxxxx> wrote:
Signed-off-by: Si-Wei Liu <si-wei.liu@xxxxxxxxxx>
---
drivers/vhost/vdpa.c | 16 +++++++++++++++-
include/uapi/linux/vhost_types.h | 2 ++
2 files changed, 17 insertions(+), 1 deletion(-)

diff --git a/drivers/vhost/vdpa.c b/drivers/vhost/vdpa.c
index 62b0a01..75092a7 100644
--- a/drivers/vhost/vdpa.c
+++ b/drivers/vhost/vdpa.c
@@ -406,6 +406,14 @@ static bool vhost_vdpa_can_resume(const struct vhost_vdpa *v)
return ops->resume;
}

+static bool vhost_vdpa_has_persistent_map(const struct vhost_vdpa *v)
+{
+ struct vdpa_device *vdpa = v->vdpa;
+ const struct vdpa_config_ops *ops = vdpa->config;
+
+ return (!ops->set_map && !ops->dma_map) || ops->reset_map;
So this means the IOTLB/IOMMU mappings have already been decoupled
from the vdpa reset.
Not in the sense of API, it' been coupled since day one from the
implementations of every on-chip IOMMU parent driver, namely mlx5_vdpa
and vdpa_sim. Because of that, later on the (improper) support for
virtio-vdpa, from commit 6f5312f80183 ("vdpa/mlx5: Add support for
running with virtio_vdpa") and 6c3d329e6486 ("vdpa_sim: get rid of DMA
ops") misused the .reset() op to realize 1:1 mapping, rendering strong
coupling between device reset and reset of iotlb mappings. This series
try to rectify that implementation deficiency, while keep userspace
continuing to work with older kernel behavior.

So it should have been noticed by the userspace.
Yes, userspace had noticed this no-chip IOMMU discrepancy since day one
I suppose. Unfortunately there's already code in userspace with this
assumption in mind that proactively tears down and sets up iotlb mapping
around vdpa device reset...
I guess we can just fix the simulator and mlx5 then we are fine?
Only IF we don't care about running new QEMU on older kernels with
flawed on-chip iommu behavior around reset. But that's a big IF...
So what I meant is:

Userspace doesn't know whether the vendor specific mappings (set_map)
are required or not. And in the implementation of vhost_vdpa, if
platform IOMMU is used, the mappings are decoupled from the reset. So
if the Qemu works with parents with platform IOMMU it means Qemu can
work if we just decouple vendor specific mappings from the parents
that uses set_map.
I was aware of this, and if you may notice I don't even offer a way
backward to retain/emulate the flawed vhost-iotlb reset behavior for
older userspace - I consider it more of a bug in .set_map driver
implementation of its own rather than what the vhost-vdpa iotlb
abstraction wishes to expose to userspace in the first place.
That's my understanding as well.

If you ever look into QEMU's vhost_vdpa_reset_status() function, you may
see memory_listener_unregister() will be called to evict all of the
existing iotlb mappings right after vhost_vdpa_reset_device() across
device reset, and later on at vhost_vdpa_dev_start(),
memory_listener_register() will set up all iotlb mappings again. In an
ideal world without this on-chip iommu deficiency QEMU should not have
to behave this way - this is what I mentioned earlier that userspace had
already noticed the discrepancy and it has to "proactively tear down and
set up iotlb mapping around vdpa device reset". Apparently from
functionality perspective this trick works completely fine with platform
IOMMU, however, it's sub-optimal in the performance perspective.
Right.

We can't simply fix QEMU by moving this memory_listener_unregister()
call out of the reset path unconditionally, as we don't want to break
the already-functioning older kernel even though it's suboptimal in
performance.
I'm not sure how things can be broken in this case?
Things won't be broken if we don't care about performance, for example reboot a large memory VM (translated to device reset internally) will freeze the guest and introduce extra reboot delay unnecessarily. If we want to fix the performance by remove memory_listener_unregister() unconditionally and we don't have such a flag to distinguish, we will break network connectivity entirely after reset - as all mappings are purged during reset on older parent driver.

Or why it is specific to parent with set_map.
As if without the .reset_map op and corresponding driver implementation (in correct way), there's no appropriate means for on-chip iommu parent driver to persist iotlb mappings across reset, isn't it? If the driver deliberately removes it from .reset, they don't support 1:1 DMA mapping for virtio-vdpa on the other hand, for instance.


Instead, to keep new QEMU continuing to work on top of the
existing or older kernels, QEMU has to check this IOTLB_PERSIST feature
flag to decide whether it is safe not to bother flushing and setting up
iotlb across reset. For the platform IOMMU case, vdpa parent driver
won't implement either the .set_map or .dma_map op, so it should be
covered in the vhost_vdpa_has_persistent_map() check I suppose.
Just to make sure we are at the same page.

From the userspace point of view, the IOTLB persists and vhost-vDPA
doesn't reset the IOTLB during vDPA reset. But we have are two levels
of the coupling in other places:

1) Qemu level: memory listener is coupled with DRIVER_OK/reset
2) vDPA parent level: mlx5 build/destroy MR during DRIVER_OK/reset

If I understand you correctly, since we've coupled in 1), Qemu can't
be aware of whether the mapping is coupled with reset or not?
I suspect it had been already noticed by someone who wrote this QEMU code since day one, just that there's lack of a comment documenting it. Or any other reason why QEMU had to decouple it in the first place? It affects performance across the board for platform IOMMU vdpa providers as well.

If we
simply decouple in 1), memory mappigns might be lost during vDPA rset.
I would tend to say 1) is an inadvertent artifact or side effect of 2), as I do not see memory listeners are used like this in other QEMU subsystems, e.g. vhost, vfio. Consider this coupling in 1) had been in the play since day one with neither advanced vDPA features like SVQ nor equivalent deficiency in platform IOMMU vdpa providers, it's suspicious that mlx5 build/destroy MR during reset was the curlprit then.

Regards,
-Siwei


Thanks


Thanks,
-Siwei
Thanks

Regards,
-Siwei
Thanks