On Wed, May 28, 2025 at 02:17:54PM -0300, Jason Gunthorpe wrote:
> On Sat, May 17, 2025 at 08:21:27PM -0700, Nicolin Chen wrote:
> > The new HW QUEUE object will be added for HW to access the guest queue for
> > HW-accelerated virtualization feature. Some of HW QUEUEs are designed in a
> > way of accessing the guest queue via a host physical address without doing
> > a translation using the nesting parent IO page table, while others can use
> > the guest physical address. For the former case, kernel working with a VMM
> > needs to pin the physical pages backing the guest memory to lock them when
> > HW QUEUE is accessing, and to ensure those physical pages to be contiguous
> > in the physical address space.
> >
> > This is very like the existing iommufd_access_pin_pages() that outputs the
> > pinned page list for the caller to test its contiguity.
> >
> > Move those code from iommufd_access_pin/unpin_pages() and related function
> > for a pair of iopt helpers that can be shared with the HW QUEUE allocator.
> >
> > Rename check_area_prot() to align with the existing iopt_area helpers, and
> > inline it to the header since iommufd_access_rw() still uses it.
> >
> > Reviewed-by: Pranjal Shrivastava <pr...@google.com>
> > Reviewed-by: Kevin Tian <kevin.t...@intel.com>
> > Reviewed-by: Jason Gunthorpe <j...@nvidia.com>
> > Signed-off-by: Nicolin Chen <nicol...@nvidia.com>
> > ---
> > drivers/iommu/iommufd/io_pagetable.h | 8 ++
> > drivers/iommu/iommufd/iommufd_private.h | 6 ++
> > drivers/iommu/iommufd/device.c | 119 ++----------------------
> > drivers/iommu/iommufd/io_pagetable.c | 97 +++++++++++++++++++
> > 4 files changed, 119 insertions(+), 111 deletions(-)
>
> And if you do what was suggested do we need this patch at all? Just
> use the normal access sequence:
>
> iommufd_access_create(ops=NULL)
> iommufd_access_attach(viommu->hwpt->ioas)
> iommufd_access_pin_pages()
>
> And store a viommu->access pointer to undo it all.
I found the entire ictx would be locked by iommufd_access_create(),
then the release fop couldn't even get invoked to destroy objects.
I added a new flag to address this:
-----------------------------------------------------------------
diff --git a/drivers/iommu/iommufd/device.c b/drivers/iommu/iommufd/device.c
index f25e272ae378c..a3e0ace583a66 100644
--- a/drivers/iommu/iommufd/device.c
+++ b/drivers/iommu/iommufd/device.c
@@ -1085,7 +1085,8 @@ void iommufd_access_destroy_object(struct iommufd_object
*obj)
if (access->ioas)
WARN_ON(iommufd_access_change_ioas(access, NULL));
mutex_unlock(&access->ioas_lock);
- iommufd_ctx_put(access->ictx);
+ if (!access->ops->internal_use)
+ iommufd_ctx_put(access->ictx);
}
/**
@@ -1126,7 +1127,8 @@ iommufd_access_create(struct iommufd_ctx *ictx,
/* The calling driver is a user until iommufd_access_destroy() */
refcount_inc(&access->obj.users);
access->ictx = ictx;
- iommufd_ctx_get(ictx);
+ if (!ops->internal_use)
+ iommufd_ctx_get(ictx);
iommufd_object_finalize(ictx, &access->obj);
*id = access->obj.id;
mutex_init(&access->ioas_lock);
-----------------------------------------------------------------
Btw, I think we can have an ops but only set unmap to NULL:
static const struct iommufd_access_ops hw_queue_access_ops = {
.needs_pin_pages = 1,
+ .internal_use = 1,
/* NULL unmap to reject IOMMUFD_CMD_IOAS_UNMAP */
};
Having two flags makes the code slightly more readable. After all,
HW queue does need to pin pages.
Thanks
Nicolin
