diff mbox series

[v4,06/14] iommu: flush I/O TLB if iommu_map() or iommu_unmap() fail

Message ID 20200804134209.8717-7-paul@xen.org
State Superseded
Headers show
Series IOMMU cleanup | expand

Commit Message

Paul Durrant Aug. 4, 2020, 1:42 p.m. UTC
From: Paul Durrant <pdurrant@amazon.com>

This patch adds a full I/O TLB flush to the error paths of iommu_map() and
iommu_unmap().

Without this change callers need constructs such as:

rc = iommu_map/unmap(...)
err = iommu_flush(...)
if ( !rc )
  rc = err;

With this change, it can be simplified to:

rc = iommu_map/unmap(...)
if ( !rc )
  rc = iommu_flush(...)

because, if the map or unmap fails the flush will be unnecessary. This saves
a stack variable and generally makes the call sites tidier.

Signed-off-by: Paul Durrant <pdurrant@amazon.com>
---
Cc: Jan Beulich <jbeulich@suse.com>

v2:
 - New in v2
---
 xen/drivers/passthrough/iommu.c | 28 ++++++++++++----------------
 1 file changed, 12 insertions(+), 16 deletions(-)

Comments

Jan Beulich Aug. 5, 2020, 4:06 p.m. UTC | #1
On 04.08.2020 15:42, Paul Durrant wrote:
> From: Paul Durrant <pdurrant@amazon.com>
> 
> This patch adds a full I/O TLB flush to the error paths of iommu_map() and
> iommu_unmap().
> 
> Without this change callers need constructs such as:
> 
> rc = iommu_map/unmap(...)
> err = iommu_flush(...)
> if ( !rc )
>   rc = err;
> 
> With this change, it can be simplified to:
> 
> rc = iommu_map/unmap(...)
> if ( !rc )
>   rc = iommu_flush(...)
> 
> because, if the map or unmap fails the flush will be unnecessary. This saves
> a stack variable and generally makes the call sites tidier.

I appreciate the intent of tidier code, but I wonder whether this
flushing doesn't go a little too far: There's a need to flush in
general when multiple pages were to be (un)mapped, and there was
at least partial success. Hence e.g. in the order == 0 case I
don't see why any flushing would be needed. Granted errors aren't
commonly expected, but anyway.

> --- a/xen/drivers/passthrough/iommu.c
> +++ b/xen/drivers/passthrough/iommu.c
> @@ -274,6 +274,10 @@ int iommu_map(struct domain *d, dfn_t dfn, mfn_t mfn,
>          break;
>      }
>  
> +    /* Something went wrong so flush everything and clear flush flags */
> +    if ( unlikely(rc) && iommu_iotlb_flush_all(d, *flush_flags) )

Both here and in the unmap path, did you get the return value
of iommu_iotlb_flush_all() the wrong way round (i.e. isn't there
a missing ! )?

Jan
Paul Durrant Aug. 5, 2020, 4:18 p.m. UTC | #2
> -----Original Message-----
> From: Jan Beulich <jbeulich@suse.com>
> Sent: 05 August 2020 17:06
> To: Paul Durrant <paul@xen.org>
> Cc: xen-devel@lists.xenproject.org; Paul Durrant <pdurrant@amazon.com>
> Subject: Re: [PATCH v4 06/14] iommu: flush I/O TLB if iommu_map() or iommu_unmap() fail
> 
> On 04.08.2020 15:42, Paul Durrant wrote:
> > From: Paul Durrant <pdurrant@amazon.com>
> >
> > This patch adds a full I/O TLB flush to the error paths of iommu_map() and
> > iommu_unmap().
> >
> > Without this change callers need constructs such as:
> >
> > rc = iommu_map/unmap(...)
> > err = iommu_flush(...)
> > if ( !rc )
> >   rc = err;
> >
> > With this change, it can be simplified to:
> >
> > rc = iommu_map/unmap(...)
> > if ( !rc )
> >   rc = iommu_flush(...)
> >
> > because, if the map or unmap fails the flush will be unnecessary. This saves
> > a stack variable and generally makes the call sites tidier.
> 
> I appreciate the intent of tidier code, but I wonder whether this
> flushing doesn't go a little too far: There's a need to flush in
> general when multiple pages were to be (un)mapped, and there was
> at least partial success. Hence e.g. in the order == 0 case I
> don't see why any flushing would be needed. Granted errors aren't
> commonly expected, but anyway.
> 

Yes, I wasn't really worried about optimizing the error case, but I can avoid unnecessary flushing in the order 0 case.

> > --- a/xen/drivers/passthrough/iommu.c
> > +++ b/xen/drivers/passthrough/iommu.c
> > @@ -274,6 +274,10 @@ int iommu_map(struct domain *d, dfn_t dfn, mfn_t mfn,
> >          break;
> >      }
> >
> > +    /* Something went wrong so flush everything and clear flush flags */
> > +    if ( unlikely(rc) && iommu_iotlb_flush_all(d, *flush_flags) )
> 
> Both here and in the unmap path, did you get the return value
> of iommu_iotlb_flush_all() the wrong way round (i.e. isn't there
> a missing ! )?
> 

Yes, I think you're right. I'll need to re-work anyway to avoid the flush in the order 0 case.

  Paul

> Jan
Jan Beulich Aug. 6, 2020, 11:41 a.m. UTC | #3
On 04.08.2020 15:42, Paul Durrant wrote:
> --- a/xen/drivers/passthrough/iommu.c
> +++ b/xen/drivers/passthrough/iommu.c
> @@ -274,6 +274,10 @@ int iommu_map(struct domain *d, dfn_t dfn, mfn_t mfn,
>          break;
>      }
>  
> +    /* Something went wrong so flush everything and clear flush flags */
> +    if ( unlikely(rc) && iommu_iotlb_flush_all(d, *flush_flags) )
> +        flush_flags = 0;

Noticed only while looking at patch 9: There's also an indirection
missing both here and ...

> @@ -330,6 +328,10 @@ int iommu_unmap(struct domain *d, dfn_t dfn, unsigned int page_order,
>          }
>      }
>  
> +    /* Something went wrong so flush everything and clear flush flags */
> +    if ( unlikely(rc) && iommu_iotlb_flush_all(d, *flush_flags) )
> +        flush_flags = 0;

... here.

Jan
Tian, Kevin Aug. 14, 2020, 6:53 a.m. UTC | #4
> From: Paul Durrant
> Sent: Tuesday, August 4, 2020 9:42 PM
> 
> From: Paul Durrant <pdurrant@amazon.com>
> 
> This patch adds a full I/O TLB flush to the error paths of iommu_map() and
> iommu_unmap().
> 
> Without this change callers need constructs such as:
> 
> rc = iommu_map/unmap(...)
> err = iommu_flush(...)
> if ( !rc )
>   rc = err;
> 
> With this change, it can be simplified to:
> 
> rc = iommu_map/unmap(...)
> if ( !rc )
>   rc = iommu_flush(...)
> 
> because, if the map or unmap fails the flush will be unnecessary. This saves

this statement is different from change in iommu_map...

> a stack variable and generally makes the call sites tidier.
> 
> Signed-off-by: Paul Durrant <pdurrant@amazon.com>
> ---
> Cc: Jan Beulich <jbeulich@suse.com>
> 
> v2:
>  - New in v2
> ---
>  xen/drivers/passthrough/iommu.c | 28 ++++++++++++----------------
>  1 file changed, 12 insertions(+), 16 deletions(-)
> 
> diff --git a/xen/drivers/passthrough/iommu.c
> b/xen/drivers/passthrough/iommu.c
> index 660dc5deb2..e2c0193a09 100644
> --- a/xen/drivers/passthrough/iommu.c
> +++ b/xen/drivers/passthrough/iommu.c
> @@ -274,6 +274,10 @@ int iommu_map(struct domain *d, dfn_t dfn, mfn_t
> mfn,
>          break;
>      }
> 
> +    /* Something went wrong so flush everything and clear flush flags */
> +    if ( unlikely(rc) && iommu_iotlb_flush_all(d, *flush_flags) )
> +        flush_flags = 0;
> +

... earlier you said flush is unnecessary if map fails. But here actually you
still need to flush everything so it's just sort of moving error-path flush
within the map function?

Thanks
Kevin

>      return rc;
>  }
> 
> @@ -283,14 +287,8 @@ int iommu_legacy_map(struct domain *d, dfn_t dfn,
> mfn_t mfn,
>      unsigned int flush_flags = 0;
>      int rc = iommu_map(d, dfn, mfn, page_order, flags, &flush_flags);
> 
> -    if ( !this_cpu(iommu_dont_flush_iotlb) )
> -    {
> -        int err = iommu_iotlb_flush(d, dfn, (1u << page_order),
> -                                    flush_flags);
> -
> -        if ( !rc )
> -            rc = err;
> -    }
> +    if ( !this_cpu(iommu_dont_flush_iotlb) && !rc )
> +        rc = iommu_iotlb_flush(d, dfn, (1u << page_order), flush_flags);
> 
>      return rc;
>  }
> @@ -330,6 +328,10 @@ int iommu_unmap(struct domain *d, dfn_t dfn,
> unsigned int page_order,
>          }
>      }
> 
> +    /* Something went wrong so flush everything and clear flush flags */
> +    if ( unlikely(rc) && iommu_iotlb_flush_all(d, *flush_flags) )
> +        flush_flags = 0;
> +
>      return rc;
>  }
> 
> @@ -338,14 +340,8 @@ int iommu_legacy_unmap(struct domain *d, dfn_t
> dfn, unsigned int page_order)
>      unsigned int flush_flags = 0;
>      int rc = iommu_unmap(d, dfn, page_order, &flush_flags);
> 
> -    if ( !this_cpu(iommu_dont_flush_iotlb) )
> -    {
> -        int err = iommu_iotlb_flush(d, dfn, (1u << page_order),
> -                                    flush_flags);
> -
> -        if ( !rc )
> -            rc = err;
> -    }
> +    if ( !this_cpu(iommu_dont_flush_iotlb) && ! rc )
> +        rc = iommu_iotlb_flush(d, dfn, (1u << page_order), flush_flags);
> 
>      return rc;
>  }
> --
> 2.20.1
>
Durrant, Paul Aug. 14, 2020, 7:19 a.m. UTC | #5
> -----Original Message-----
> From: Tian, Kevin <kevin.tian@intel.com>
> Sent: 14 August 2020 07:53
> To: Paul Durrant <paul@xen.org>; xen-devel@lists.xenproject.org
> Cc: Durrant, Paul <pdurrant@amazon.co.uk>; Jan Beulich <jbeulich@suse.com>
> Subject: RE: [EXTERNAL] [PATCH v4 06/14] iommu: flush I/O TLB if iommu_map() or iommu_unmap() fail
> 
> CAUTION: This email originated from outside of the organization. Do not click links or open
> attachments unless you can confirm the sender and know the content is safe.
> 
> 
> 
> > From: Paul Durrant
> > Sent: Tuesday, August 4, 2020 9:42 PM
> >
> > From: Paul Durrant <pdurrant@amazon.com>
> >
> > This patch adds a full I/O TLB flush to the error paths of iommu_map() and
> > iommu_unmap().
> >
> > Without this change callers need constructs such as:
> >
> > rc = iommu_map/unmap(...)
> > err = iommu_flush(...)
> > if ( !rc )
> >   rc = err;
> >
> > With this change, it can be simplified to:
> >
> > rc = iommu_map/unmap(...)
> > if ( !rc )
> >   rc = iommu_flush(...)
> >
> > because, if the map or unmap fails the flush will be unnecessary. This saves
> 
> this statement is different from change in iommu_map...
> 
> > a stack variable and generally makes the call sites tidier.
> >
> > Signed-off-by: Paul Durrant <pdurrant@amazon.com>
> > ---
> > Cc: Jan Beulich <jbeulich@suse.com>
> >
> > v2:
> >  - New in v2
> > ---
> >  xen/drivers/passthrough/iommu.c | 28 ++++++++++++----------------
> >  1 file changed, 12 insertions(+), 16 deletions(-)
> >
> > diff --git a/xen/drivers/passthrough/iommu.c
> > b/xen/drivers/passthrough/iommu.c
> > index 660dc5deb2..e2c0193a09 100644
> > --- a/xen/drivers/passthrough/iommu.c
> > +++ b/xen/drivers/passthrough/iommu.c
> > @@ -274,6 +274,10 @@ int iommu_map(struct domain *d, dfn_t dfn, mfn_t
> > mfn,
> >          break;
> >      }
> >
> > +    /* Something went wrong so flush everything and clear flush flags */
> > +    if ( unlikely(rc) && iommu_iotlb_flush_all(d, *flush_flags) )
> > +        flush_flags = 0;
> > +
> 
> ... earlier you said flush is unnecessary if map fails. But here actually you
> still need to flush everything so it's just sort of moving error-path flush
> within the map function?

Yes, that's actually what's happening. The language in the comment is ambiguous I guess. I'll modify it to say

"because, if the map or unmap fails an explicit flush will be unnecessary."

Hopefully that is clearer.

  Paul

> 
> Thanks
> Kevin
> 
> >      return rc;
> >  }
> >
> > @@ -283,14 +287,8 @@ int iommu_legacy_map(struct domain *d, dfn_t dfn,
> > mfn_t mfn,
> >      unsigned int flush_flags = 0;
> >      int rc = iommu_map(d, dfn, mfn, page_order, flags, &flush_flags);
> >
> > -    if ( !this_cpu(iommu_dont_flush_iotlb) )
> > -    {
> > -        int err = iommu_iotlb_flush(d, dfn, (1u << page_order),
> > -                                    flush_flags);
> > -
> > -        if ( !rc )
> > -            rc = err;
> > -    }
> > +    if ( !this_cpu(iommu_dont_flush_iotlb) && !rc )
> > +        rc = iommu_iotlb_flush(d, dfn, (1u << page_order), flush_flags);
> >
> >      return rc;
> >  }
> > @@ -330,6 +328,10 @@ int iommu_unmap(struct domain *d, dfn_t dfn,
> > unsigned int page_order,
> >          }
> >      }
> >
> > +    /* Something went wrong so flush everything and clear flush flags */
> > +    if ( unlikely(rc) && iommu_iotlb_flush_all(d, *flush_flags) )
> > +        flush_flags = 0;
> > +
> >      return rc;
> >  }
> >
> > @@ -338,14 +340,8 @@ int iommu_legacy_unmap(struct domain *d, dfn_t
> > dfn, unsigned int page_order)
> >      unsigned int flush_flags = 0;
> >      int rc = iommu_unmap(d, dfn, page_order, &flush_flags);
> >
> > -    if ( !this_cpu(iommu_dont_flush_iotlb) )
> > -    {
> > -        int err = iommu_iotlb_flush(d, dfn, (1u << page_order),
> > -                                    flush_flags);
> > -
> > -        if ( !rc )
> > -            rc = err;
> > -    }
> > +    if ( !this_cpu(iommu_dont_flush_iotlb) && ! rc )
> > +        rc = iommu_iotlb_flush(d, dfn, (1u << page_order), flush_flags);
> >
> >      return rc;
> >  }
> > --
> > 2.20.1
> >
diff mbox series

Patch

diff --git a/xen/drivers/passthrough/iommu.c b/xen/drivers/passthrough/iommu.c
index 660dc5deb2..e2c0193a09 100644
--- a/xen/drivers/passthrough/iommu.c
+++ b/xen/drivers/passthrough/iommu.c
@@ -274,6 +274,10 @@  int iommu_map(struct domain *d, dfn_t dfn, mfn_t mfn,
         break;
     }
 
+    /* Something went wrong so flush everything and clear flush flags */
+    if ( unlikely(rc) && iommu_iotlb_flush_all(d, *flush_flags) )
+        flush_flags = 0;
+
     return rc;
 }
 
@@ -283,14 +287,8 @@  int iommu_legacy_map(struct domain *d, dfn_t dfn, mfn_t mfn,
     unsigned int flush_flags = 0;
     int rc = iommu_map(d, dfn, mfn, page_order, flags, &flush_flags);
 
-    if ( !this_cpu(iommu_dont_flush_iotlb) )
-    {
-        int err = iommu_iotlb_flush(d, dfn, (1u << page_order),
-                                    flush_flags);
-
-        if ( !rc )
-            rc = err;
-    }
+    if ( !this_cpu(iommu_dont_flush_iotlb) && !rc )
+        rc = iommu_iotlb_flush(d, dfn, (1u << page_order), flush_flags);
 
     return rc;
 }
@@ -330,6 +328,10 @@  int iommu_unmap(struct domain *d, dfn_t dfn, unsigned int page_order,
         }
     }
 
+    /* Something went wrong so flush everything and clear flush flags */
+    if ( unlikely(rc) && iommu_iotlb_flush_all(d, *flush_flags) )
+        flush_flags = 0;
+
     return rc;
 }
 
@@ -338,14 +340,8 @@  int iommu_legacy_unmap(struct domain *d, dfn_t dfn, unsigned int page_order)
     unsigned int flush_flags = 0;
     int rc = iommu_unmap(d, dfn, page_order, &flush_flags);
 
-    if ( !this_cpu(iommu_dont_flush_iotlb) )
-    {
-        int err = iommu_iotlb_flush(d, dfn, (1u << page_order),
-                                    flush_flags);
-
-        if ( !rc )
-            rc = err;
-    }
+    if ( !this_cpu(iommu_dont_flush_iotlb) && ! rc )
+        rc = iommu_iotlb_flush(d, dfn, (1u << page_order), flush_flags);
 
     return rc;
 }