[Qemu-devel] [PATCH 0/3] xen: XSA-197 follow-ups

[Qemu-devel] [PATCH 0/3] xen: XSA-197 follow-ups

[Jan Beulich]

1: fix quad word bufioreq handling
2: slightly simplify bufioreq handling
3: ignore direction in bufioreq handling

Signed-off-by: Jan Beulich <jbeulich@suse.com>

[PATCH 1/3] xen: fix quad word bufioreq handling

[Jan Beulich]

We should not consume the second slot if it didn't get written yet.
Normal writers - i.e. Xen - would not update write_pointer between the
two writes, but the page may get fiddled with by the guest itself, and
we're better off entering an infinite loop in that case.

Reported-by: yanghongke <yanghongke@huawei.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
---
TBD: Alternatively we could call e.g. hw_error() instead.

--- a/xen-hvm.c
+++ b/xen-hvm.c
@@ -1021,6 +1021,9 @@ static int handle_buffered_iopage(XenIOS
         xen_rmb();
         qw = (req.size == 8);
         if (qw) {
+            if (rdptr + 1 == wrptr) {
+                break;
+            }
             buf_req = &buf_page->buf_ioreq[(rdptr + 1) %
                                            IOREQ_BUFFER_SLOT_NUM];
             req.data |= ((uint64_t)buf_req->data) << 32;




_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

[Qemu-devel] [PATCH 1/3] xen: fix quad word bufioreq handling

[Jan Beulich]

We should not consume the second slot if it didn't get written yet.
Normal writers - i.e. Xen - would not update write_pointer between the
two writes, but the page may get fiddled with by the guest itself, and
we're better off entering an infinite loop in that case.

Reported-by: yanghongke <yanghongke@huawei.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
---
TBD: Alternatively we could call e.g. hw_error() instead.

--- a/xen-hvm.c
+++ b/xen-hvm.c
@@ -1021,6 +1021,9 @@ static int handle_buffered_iopage(XenIOS
         xen_rmb();
         qw = (req.size == 8);
         if (qw) {
+            if (rdptr + 1 == wrptr) {
+                break;
+            }
             buf_req = &buf_page->buf_ioreq[(rdptr + 1) %
                                            IOREQ_BUFFER_SLOT_NUM];
             req.data |= ((uint64_t)buf_req->data) << 32;

Re: [Qemu-devel] [PATCH 1/3] xen: fix quad word bufioreq handling

[Paul Durrant]

> -----Original Message-----
> From: Jan Beulich [mailto:JBeulich@suse.com]
> Sent: 23 November 2016 09:24
> To: qemu-devel@nongnu.org
> Cc: Anthony Perard <anthony.perard@citrix.com>; Paul Durrant
> <Paul.Durrant@citrix.com>; Stefano Stabellini <sstabellini@kernel.org>; xen-
> devel <xen-devel@lists.xenproject.org>
> Subject: [PATCH 1/3] xen: fix quad word bufioreq handling
> 
> We should not consume the second slot if it didn't get written yet.
> Normal writers - i.e. Xen - would not update write_pointer between the
> two writes, but the page may get fiddled with by the guest itself, and
> we're better off entering an infinite loop in that case.
> 

Xen would never put QEMU in this situation and the guest can't actually modify the page whilst it's in use, since activation of the IOREQ server removes the page from the guest's p2m so the premise of the patch is not correct.

  Paul

> Reported-by: yanghongke <yanghongke@huawei.com>
> Signed-off-by: Jan Beulich <jbeulich@suse.com>
> ---
> TBD: Alternatively we could call e.g. hw_error() instead.
> 
> --- a/xen-hvm.c
> +++ b/xen-hvm.c
> @@ -1021,6 +1021,9 @@ static int handle_buffered_iopage(XenIOS
>          xen_rmb();
>          qw = (req.size == 8);
>          if (qw) {
> +            if (rdptr + 1 == wrptr) {
> +                break;
> +            }
>              buf_req = &buf_page->buf_ioreq[(rdptr + 1) %
>                                             IOREQ_BUFFER_SLOT_NUM];
>              req.data |= ((uint64_t)buf_req->data) << 32;
> 
> 

Re: [PATCH 1/3] xen: fix quad word bufioreq handling

[Jan Beulich]

>>> On 23.11.16 at 10:48, <Paul.Durrant@citrix.com> wrote:
>> From: Jan Beulich [mailto:JBeulich@suse.com]
>> Sent: 23 November 2016 09:24
>> 
>> We should not consume the second slot if it didn't get written yet.
>> Normal writers - i.e. Xen - would not update write_pointer between the
>> two writes, but the page may get fiddled with by the guest itself, and
>> we're better off entering an infinite loop in that case.
>> 
> 
> Xen would never put QEMU in this situation and the guest can't actually 
> modify the page whilst it's in use, since activation of the IOREQ server 
> removes the page from the guest's p2m so the premise of the patch is not 
> correct.

Is that the case even for pre-ioreq-server Xen versions? The issue
here was reported together with what became XSA-197, and it's
not been assigned its own XSA just because there are other ways
for a guest to place high load on its qemu process (and there are
ways to deal with such high load situations).

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [Qemu-devel] [PATCH 1/3] xen: fix quad word bufioreq handling

[Jan Beulich]

>>> On 23.11.16 at 10:48, <Paul.Durrant@citrix.com> wrote:
>> From: Jan Beulich [mailto:JBeulich@suse.com]
>> Sent: 23 November 2016 09:24
>> 
>> We should not consume the second slot if it didn't get written yet.
>> Normal writers - i.e. Xen - would not update write_pointer between the
>> two writes, but the page may get fiddled with by the guest itself, and
>> we're better off entering an infinite loop in that case.
>> 
> 
> Xen would never put QEMU in this situation and the guest can't actually 
> modify the page whilst it's in use, since activation of the IOREQ server 
> removes the page from the guest's p2m so the premise of the patch is not 
> correct.

Is that the case even for pre-ioreq-server Xen versions? The issue
here was reported together with what became XSA-197, and it's
not been assigned its own XSA just because there are other ways
for a guest to place high load on its qemu process (and there are
ways to deal with such high load situations).

Jan

Re: [Qemu-devel] [PATCH 1/3] xen: fix quad word bufioreq handling

[Paul Durrant]

> -----Original Message-----
> From: Jan Beulich [mailto:JBeulich@suse.com]
> Sent: 23 November 2016 10:36
> To: Paul Durrant <Paul.Durrant@citrix.com>
> Cc: Anthony Perard <anthony.perard@citrix.com>; Stefano Stabellini
> <sstabellini@kernel.org>; xen-devel <xen-devel@lists.xenproject.org>;
> qemu-devel@nongnu.org
> Subject: RE: [PATCH 1/3] xen: fix quad word bufioreq handling
> 
> >>> On 23.11.16 at 10:48, <Paul.Durrant@citrix.com> wrote:
> >> From: Jan Beulich [mailto:JBeulich@suse.com]
> >> Sent: 23 November 2016 09:24
> >>
> >> We should not consume the second slot if it didn't get written yet.
> >> Normal writers - i.e. Xen - would not update write_pointer between the
> >> two writes, but the page may get fiddled with by the guest itself, and
> >> we're better off entering an infinite loop in that case.
> >>
> >
> > Xen would never put QEMU in this situation and the guest can't actually
> > modify the page whilst it's in use, since activation of the IOREQ server
> > removes the page from the guest's p2m so the premise of the patch is not
> > correct.
> 
> Is that the case even for pre-ioreq-server Xen versions? The issue
> here was reported together with what became XSA-197, and it's
> not been assigned its own XSA just because there are other ways
> for a guest to place high load on its qemu process (and there are
> ways to deal with such high load situations).
> 

No, if QEMU is using a default ioreq server (i.e. the legacy way of doing things) then it's vulnerable to the guest messing with the rings and I'd forgotten that migrated-in guests from old QEMUs also end up using the default server, so I guess this is a worthy checkt to make... although maybe it's best to just bail if the check fails, since it would indicate a malicious guest.

  Paul

> Jan

Re: [PATCH 1/3] xen: fix quad word bufioreq handling

[Jan Beulich]

>>> On 23.11.16 at 11:45, <Paul.Durrant@citrix.com> wrote:
> No, if QEMU is using a default ioreq server (i.e. the legacy way of doing 
> things) then it's vulnerable to the guest messing with the rings and I'd 
> forgotten that migrated-in guests from old QEMUs also end up using the default 
> server, so I guess this is a worthy checkt to make... although maybe it's 
> best to just bail if the check fails, since it would indicate a malicious 
> guest.

Okay, that's basically the TBD note I have in the patch; I'll wait for
at least one of the qemu maintainers to voice their preference.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [Qemu-devel] [PATCH 1/3] xen: fix quad word bufioreq handling

[Jan Beulich]

>>> On 23.11.16 at 11:45, <Paul.Durrant@citrix.com> wrote:
> No, if QEMU is using a default ioreq server (i.e. the legacy way of doing 
> things) then it's vulnerable to the guest messing with the rings and I'd 
> forgotten that migrated-in guests from old QEMUs also end up using the default 
> server, so I guess this is a worthy checkt to make... although maybe it's 
> best to just bail if the check fails, since it would indicate a malicious 
> guest.

Okay, that's basically the TBD note I have in the patch; I'll wait for
at least one of the qemu maintainers to voice their preference.

Jan

Re: [Qemu-devel] [PATCH 1/3] xen: fix quad word bufioreq handling

[Stefano Stabellini]

On Wed, 23 Nov 2016, Jan Beulich wrote:
> >>> On 23.11.16 at 11:45, <Paul.Durrant@citrix.com> wrote:
> > No, if QEMU is using a default ioreq server (i.e. the legacy way of doing 
> > things) then it's vulnerable to the guest messing with the rings and I'd 
> > forgotten that migrated-in guests from old QEMUs also end up using the default 
> > server, so I guess this is a worthy checkt to make... although maybe it's 
> > best to just bail if the check fails, since it would indicate a malicious 
> > guest.
> 
> Okay, that's basically the TBD note I have in the patch; I'll wait for
> at least one of the qemu maintainers to voice their preference.
 
I think we should just print an error and destroy_hvm_domain(false) or
hw_error if the check fails.

Re: [PATCH 1/3] xen: fix quad word bufioreq handling

[Stefano Stabellini]

On Wed, 23 Nov 2016, Jan Beulich wrote:
> >>> On 23.11.16 at 11:45, <Paul.Durrant@citrix.com> wrote:
> > No, if QEMU is using a default ioreq server (i.e. the legacy way of doing 
> > things) then it's vulnerable to the guest messing with the rings and I'd 
> > forgotten that migrated-in guests from old QEMUs also end up using the default 
> > server, so I guess this is a worthy checkt to make... although maybe it's 
> > best to just bail if the check fails, since it would indicate a malicious 
> > guest.
> 
> Okay, that's basically the TBD note I have in the patch; I'll wait for
> at least one of the qemu maintainers to voice their preference.
 
I think we should just print an error and destroy_hvm_domain(false) or
hw_error if the check fails.

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [PATCH 1/3] xen: fix quad word bufioreq handling

[Paul Durrant]

> -----Original Message-----
> From: Jan Beulich [mailto:JBeulich@suse.com]
> Sent: 23 November 2016 10:36
> To: Paul Durrant <Paul.Durrant@citrix.com>
> Cc: Anthony Perard <anthony.perard@citrix.com>; Stefano Stabellini
> <sstabellini@kernel.org>; xen-devel <xen-devel@lists.xenproject.org>;
> qemu-devel@nongnu.org
> Subject: RE: [PATCH 1/3] xen: fix quad word bufioreq handling
> 
> >>> On 23.11.16 at 10:48, <Paul.Durrant@citrix.com> wrote:
> >> From: Jan Beulich [mailto:JBeulich@suse.com]
> >> Sent: 23 November 2016 09:24
> >>
> >> We should not consume the second slot if it didn't get written yet.
> >> Normal writers - i.e. Xen - would not update write_pointer between the
> >> two writes, but the page may get fiddled with by the guest itself, and
> >> we're better off entering an infinite loop in that case.
> >>
> >
> > Xen would never put QEMU in this situation and the guest can't actually
> > modify the page whilst it's in use, since activation of the IOREQ server
> > removes the page from the guest's p2m so the premise of the patch is not
> > correct.
> 
> Is that the case even for pre-ioreq-server Xen versions? The issue
> here was reported together with what became XSA-197, and it's
> not been assigned its own XSA just because there are other ways
> for a guest to place high load on its qemu process (and there are
> ways to deal with such high load situations).
> 

No, if QEMU is using a default ioreq server (i.e. the legacy way of doing things) then it's vulnerable to the guest messing with the rings and I'd forgotten that migrated-in guests from old QEMUs also end up using the default server, so I guess this is a worthy checkt to make... although maybe it's best to just bail if the check fails, since it would indicate a malicious guest.

  Paul

> Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [PATCH 1/3] xen: fix quad word bufioreq handling

[Paul Durrant]

> -----Original Message-----
> From: Jan Beulich [mailto:JBeulich@suse.com]
> Sent: 23 November 2016 09:24
> To: qemu-devel@nongnu.org
> Cc: Anthony Perard <anthony.perard@citrix.com>; Paul Durrant
> <Paul.Durrant@citrix.com>; Stefano Stabellini <sstabellini@kernel.org>; xen-
> devel <xen-devel@lists.xenproject.org>
> Subject: [PATCH 1/3] xen: fix quad word bufioreq handling
> 
> We should not consume the second slot if it didn't get written yet.
> Normal writers - i.e. Xen - would not update write_pointer between the
> two writes, but the page may get fiddled with by the guest itself, and
> we're better off entering an infinite loop in that case.
> 

Xen would never put QEMU in this situation and the guest can't actually modify the page whilst it's in use, since activation of the IOREQ server removes the page from the guest's p2m so the premise of the patch is not correct.

  Paul

> Reported-by: yanghongke <yanghongke@huawei.com>
> Signed-off-by: Jan Beulich <jbeulich@suse.com>
> ---
> TBD: Alternatively we could call e.g. hw_error() instead.
> 
> --- a/xen-hvm.c
> +++ b/xen-hvm.c
> @@ -1021,6 +1021,9 @@ static int handle_buffered_iopage(XenIOS
>          xen_rmb();
>          qw = (req.size == 8);
>          if (qw) {
> +            if (rdptr + 1 == wrptr) {
> +                break;
> +            }
>              buf_req = &buf_page->buf_ioreq[(rdptr + 1) %
>                                             IOREQ_BUFFER_SLOT_NUM];
>              req.data |= ((uint64_t)buf_req->data) << 32;
> 
> 


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

[Qemu-devel] [PATCH 2/3] xen: slightly simplify bufioreq handling

[Jan Beulich]

There's no point setting fields always receiving the same value on each
iteration, as handle_ioreq() doesn't alter them anyway. Set state and
count once ahead of the loop, drop the redundant clearing of
data_is_ptr, and avoid the meaningless setting of df altogether.

Also avoid doing an unsigned long calculation of size when the field to
be initialized is only 32 bits wide (and the shift value in the range
0...3).

Signed-off-by: Jan Beulich <jbeulich@suse.com>

--- a/xen-hvm.c
+++ b/xen-hvm.c
@@ -995,6 +995,8 @@ static int handle_buffered_iopage(XenIOS
     }
 
     memset(&req, 0x00, sizeof(req));
+    req.state = STATE_IOREQ_READY;
+    req.count = 1;
 
     for (;;) {
         uint32_t rdptr = buf_page->read_pointer, wrptr;
@@ -1009,15 +1011,11 @@ static int handle_buffered_iopage(XenIOS
             break;
         }
         buf_req = &buf_page->buf_ioreq[rdptr % IOREQ_BUFFER_SLOT_NUM];
-        req.size = 1UL << buf_req->size;
-        req.count = 1;
+        req.size = 1U << buf_req->size;
         req.addr = buf_req->addr;
         req.data = buf_req->data;
-        req.state = STATE_IOREQ_READY;
         req.dir = buf_req->dir;
-        req.df = 1;
         req.type = buf_req->type;
-        req.data_is_ptr = 0;
         xen_rmb();
         qw = (req.size == 8);
         if (qw) {
@@ -1032,6 +1030,13 @@ static int handle_buffered_iopage(XenIOS
 
         handle_ioreq(state, &req);
 
+        /* Only req.data may get updated by handle_ioreq(), albeit even that
+         * should not happen as such data would never make it to the guest.
+         */
+        assert(req.state == STATE_IOREQ_READY);
+        assert(req.count == 1);
+        assert(!req.data_is_ptr);
+
         atomic_add(&buf_page->read_pointer, qw + 1);
     }
 

[PATCH 2/3] xen: slightly simplify bufioreq handling

[Jan Beulich]

There's no point setting fields always receiving the same value on each
iteration, as handle_ioreq() doesn't alter them anyway. Set state and
count once ahead of the loop, drop the redundant clearing of
data_is_ptr, and avoid the meaningless setting of df altogether.

Also avoid doing an unsigned long calculation of size when the field to
be initialized is only 32 bits wide (and the shift value in the range
0...3).

Signed-off-by: Jan Beulich <jbeulich@suse.com>

--- a/xen-hvm.c
+++ b/xen-hvm.c
@@ -995,6 +995,8 @@ static int handle_buffered_iopage(XenIOS
     }
 
     memset(&req, 0x00, sizeof(req));
+    req.state = STATE_IOREQ_READY;
+    req.count = 1;
 
     for (;;) {
         uint32_t rdptr = buf_page->read_pointer, wrptr;
@@ -1009,15 +1011,11 @@ static int handle_buffered_iopage(XenIOS
             break;
         }
         buf_req = &buf_page->buf_ioreq[rdptr % IOREQ_BUFFER_SLOT_NUM];
-        req.size = 1UL << buf_req->size;
-        req.count = 1;
+        req.size = 1U << buf_req->size;
         req.addr = buf_req->addr;
         req.data = buf_req->data;
-        req.state = STATE_IOREQ_READY;
         req.dir = buf_req->dir;
-        req.df = 1;
         req.type = buf_req->type;
-        req.data_is_ptr = 0;
         xen_rmb();
         qw = (req.size == 8);
         if (qw) {
@@ -1032,6 +1030,13 @@ static int handle_buffered_iopage(XenIOS
 
         handle_ioreq(state, &req);
 
+        /* Only req.data may get updated by handle_ioreq(), albeit even that
+         * should not happen as such data would never make it to the guest.
+         */
+        assert(req.state == STATE_IOREQ_READY);
+        assert(req.count == 1);
+        assert(!req.data_is_ptr);
+
         atomic_add(&buf_page->read_pointer, qw + 1);
     }
 




_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [Qemu-devel] [PATCH 2/3] xen: slightly simplify bufioreq handling

[Paul Durrant]

> -----Original Message-----
> From: Jan Beulich [mailto:JBeulich@suse.com]
> Sent: 23 November 2016 09:25
> To: qemu-devel@nongnu.org
> Cc: Anthony Perard <anthony.perard@citrix.com>; Paul Durrant
> <Paul.Durrant@citrix.com>; Stefano Stabellini <sstabellini@kernel.org>; xen-
> devel <xen-devel@lists.xenproject.org>
> Subject: [PATCH 2/3] xen: slightly simplify bufioreq handling
> 
> There's no point setting fields always receiving the same value on each
> iteration, as handle_ioreq() doesn't alter them anyway. Set state and
> count once ahead of the loop, drop the redundant clearing of
> data_is_ptr, and avoid the meaningless setting of df altogether.
> 
> Also avoid doing an unsigned long calculation of size when the field to
> be initialized is only 32 bits wide (and the shift value in the range
> 0...3).
> 
> Signed-off-by: Jan Beulich <jbeulich@suse.com>

Reviewed-by: Paul Durrant <paul.durrant@citrix.com>

> 
> --- a/xen-hvm.c
> +++ b/xen-hvm.c
> @@ -995,6 +995,8 @@ static int handle_buffered_iopage(XenIOS
>      }
> 
>      memset(&req, 0x00, sizeof(req));
> +    req.state = STATE_IOREQ_READY;
> +    req.count = 1;
> 
>      for (;;) {
>          uint32_t rdptr = buf_page->read_pointer, wrptr;
> @@ -1009,15 +1011,11 @@ static int handle_buffered_iopage(XenIOS
>              break;
>          }
>          buf_req = &buf_page->buf_ioreq[rdptr % IOREQ_BUFFER_SLOT_NUM];
> -        req.size = 1UL << buf_req->size;
> -        req.count = 1;
> +        req.size = 1U << buf_req->size;
>          req.addr = buf_req->addr;
>          req.data = buf_req->data;
> -        req.state = STATE_IOREQ_READY;
>          req.dir = buf_req->dir;
> -        req.df = 1;
>          req.type = buf_req->type;
> -        req.data_is_ptr = 0;
>          xen_rmb();
>          qw = (req.size == 8);
>          if (qw) {
> @@ -1032,6 +1030,13 @@ static int handle_buffered_iopage(XenIOS
> 
>          handle_ioreq(state, &req);
> 
> +        /* Only req.data may get updated by handle_ioreq(), albeit even that
> +         * should not happen as such data would never make it to the guest.
> +         */
> +        assert(req.state == STATE_IOREQ_READY);
> +        assert(req.count == 1);
> +        assert(!req.data_is_ptr);
> +
>          atomic_add(&buf_page->read_pointer, qw + 1);
>      }
> 
> 
> 

Re: [PATCH 2/3] xen: slightly simplify bufioreq handling

[Paul Durrant]

> -----Original Message-----
> From: Jan Beulich [mailto:JBeulich@suse.com]
> Sent: 23 November 2016 09:25
> To: qemu-devel@nongnu.org
> Cc: Anthony Perard <anthony.perard@citrix.com>; Paul Durrant
> <Paul.Durrant@citrix.com>; Stefano Stabellini <sstabellini@kernel.org>; xen-
> devel <xen-devel@lists.xenproject.org>
> Subject: [PATCH 2/3] xen: slightly simplify bufioreq handling
> 
> There's no point setting fields always receiving the same value on each
> iteration, as handle_ioreq() doesn't alter them anyway. Set state and
> count once ahead of the loop, drop the redundant clearing of
> data_is_ptr, and avoid the meaningless setting of df altogether.
> 
> Also avoid doing an unsigned long calculation of size when the field to
> be initialized is only 32 bits wide (and the shift value in the range
> 0...3).
> 
> Signed-off-by: Jan Beulich <jbeulich@suse.com>

Reviewed-by: Paul Durrant <paul.durrant@citrix.com>

> 
> --- a/xen-hvm.c
> +++ b/xen-hvm.c
> @@ -995,6 +995,8 @@ static int handle_buffered_iopage(XenIOS
>      }
> 
>      memset(&req, 0x00, sizeof(req));
> +    req.state = STATE_IOREQ_READY;
> +    req.count = 1;
> 
>      for (;;) {
>          uint32_t rdptr = buf_page->read_pointer, wrptr;
> @@ -1009,15 +1011,11 @@ static int handle_buffered_iopage(XenIOS
>              break;
>          }
>          buf_req = &buf_page->buf_ioreq[rdptr % IOREQ_BUFFER_SLOT_NUM];
> -        req.size = 1UL << buf_req->size;
> -        req.count = 1;
> +        req.size = 1U << buf_req->size;
>          req.addr = buf_req->addr;
>          req.data = buf_req->data;
> -        req.state = STATE_IOREQ_READY;
>          req.dir = buf_req->dir;
> -        req.df = 1;
>          req.type = buf_req->type;
> -        req.data_is_ptr = 0;
>          xen_rmb();
>          qw = (req.size == 8);
>          if (qw) {
> @@ -1032,6 +1030,13 @@ static int handle_buffered_iopage(XenIOS
> 
>          handle_ioreq(state, &req);
> 
> +        /* Only req.data may get updated by handle_ioreq(), albeit even that
> +         * should not happen as such data would never make it to the guest.
> +         */
> +        assert(req.state == STATE_IOREQ_READY);
> +        assert(req.count == 1);
> +        assert(!req.data_is_ptr);
> +
>          atomic_add(&buf_page->read_pointer, qw + 1);
>      }
> 
> 
> 


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [PATCH 2/3] xen: slightly simplify bufioreq handling

[Stefano Stabellini]

On Wed, 23 Nov 2016, Jan Beulich wrote:
> There's no point setting fields always receiving the same value on each
> iteration, as handle_ioreq() doesn't alter them anyway. Set state and
> count once ahead of the loop, drop the redundant clearing of
> data_is_ptr, and avoid the meaningless setting of df altogether.

Why setting df is meaningless?


> Also avoid doing an unsigned long calculation of size when the field to
> be initialized is only 32 bits wide (and the shift value in the range
> 0...3).
> 
> Signed-off-by: Jan Beulich <jbeulich@suse.com>
> 
> --- a/xen-hvm.c
> +++ b/xen-hvm.c
> @@ -995,6 +995,8 @@ static int handle_buffered_iopage(XenIOS
>      }
>  
>      memset(&req, 0x00, sizeof(req));
> +    req.state = STATE_IOREQ_READY;
> +    req.count = 1;
>  
>      for (;;) {
>          uint32_t rdptr = buf_page->read_pointer, wrptr;
> @@ -1009,15 +1011,11 @@ static int handle_buffered_iopage(XenIOS
>              break;
>          }
>          buf_req = &buf_page->buf_ioreq[rdptr % IOREQ_BUFFER_SLOT_NUM];
> -        req.size = 1UL << buf_req->size;
> -        req.count = 1;
> +        req.size = 1U << buf_req->size;
>          req.addr = buf_req->addr;
>          req.data = buf_req->data;
> -        req.state = STATE_IOREQ_READY;
>          req.dir = buf_req->dir;
> -        req.df = 1;
>          req.type = buf_req->type;
> -        req.data_is_ptr = 0;
>          xen_rmb();
>          qw = (req.size == 8);
>          if (qw) {
> @@ -1032,6 +1030,13 @@ static int handle_buffered_iopage(XenIOS
>  
>          handle_ioreq(state, &req);
>  
> +        /* Only req.data may get updated by handle_ioreq(), albeit even that
> +         * should not happen as such data would never make it to the guest.
> +         */
> +        assert(req.state == STATE_IOREQ_READY);
> +        assert(req.count == 1);
> +        assert(!req.data_is_ptr);
> +
>          atomic_add(&buf_page->read_pointer, qw + 1);
>      }
>  
> 
> 
> 

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [Qemu-devel] [PATCH 2/3] xen: slightly simplify bufioreq handling

[Stefano Stabellini]

On Wed, 23 Nov 2016, Jan Beulich wrote:
> There's no point setting fields always receiving the same value on each
> iteration, as handle_ioreq() doesn't alter them anyway. Set state and
> count once ahead of the loop, drop the redundant clearing of
> data_is_ptr, and avoid the meaningless setting of df altogether.

Why setting df is meaningless?


> Also avoid doing an unsigned long calculation of size when the field to
> be initialized is only 32 bits wide (and the shift value in the range
> 0...3).
> 
> Signed-off-by: Jan Beulich <jbeulich@suse.com>
> 
> --- a/xen-hvm.c
> +++ b/xen-hvm.c
> @@ -995,6 +995,8 @@ static int handle_buffered_iopage(XenIOS
>      }
>  
>      memset(&req, 0x00, sizeof(req));
> +    req.state = STATE_IOREQ_READY;
> +    req.count = 1;
>  
>      for (;;) {
>          uint32_t rdptr = buf_page->read_pointer, wrptr;
> @@ -1009,15 +1011,11 @@ static int handle_buffered_iopage(XenIOS
>              break;
>          }
>          buf_req = &buf_page->buf_ioreq[rdptr % IOREQ_BUFFER_SLOT_NUM];
> -        req.size = 1UL << buf_req->size;
> -        req.count = 1;
> +        req.size = 1U << buf_req->size;
>          req.addr = buf_req->addr;
>          req.data = buf_req->data;
> -        req.state = STATE_IOREQ_READY;
>          req.dir = buf_req->dir;
> -        req.df = 1;
>          req.type = buf_req->type;
> -        req.data_is_ptr = 0;
>          xen_rmb();
>          qw = (req.size == 8);
>          if (qw) {
> @@ -1032,6 +1030,13 @@ static int handle_buffered_iopage(XenIOS
>  
>          handle_ioreq(state, &req);
>  
> +        /* Only req.data may get updated by handle_ioreq(), albeit even that
> +         * should not happen as such data would never make it to the guest.
> +         */
> +        assert(req.state == STATE_IOREQ_READY);
> +        assert(req.count == 1);
> +        assert(!req.data_is_ptr);
> +
>          atomic_add(&buf_page->read_pointer, qw + 1);
>      }
>  
> 
> 
> 

Re: [Qemu-devel] [PATCH 2/3] xen: slightly simplify bufioreq handling

[Jan Beulich]

>>> On 23.11.16 at 19:13, <sstabellini@kernel.org> wrote:
> On Wed, 23 Nov 2016, Jan Beulich wrote:
>> There's no point setting fields always receiving the same value on each
>> iteration, as handle_ioreq() doesn't alter them anyway. Set state and
>> count once ahead of the loop, drop the redundant clearing of
>> data_is_ptr, and avoid the meaningless setting of df altogether.
> 
> Why setting df is meaningless?

With count being fixed to one there's no need to update addresses,
and hence no use for knowing which direction the updates should go.

Jan

Re: [PATCH 2/3] xen: slightly simplify bufioreq handling

[Jan Beulich]

>>> On 23.11.16 at 19:13, <sstabellini@kernel.org> wrote:
> On Wed, 23 Nov 2016, Jan Beulich wrote:
>> There's no point setting fields always receiving the same value on each
>> iteration, as handle_ioreq() doesn't alter them anyway. Set state and
>> count once ahead of the loop, drop the redundant clearing of
>> data_is_ptr, and avoid the meaningless setting of df altogether.
> 
> Why setting df is meaningless?

With count being fixed to one there's no need to update addresses,
and hence no use for knowing which direction the updates should go.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

[Qemu-devel] [PATCH 3/3] xen: ignore direction in bufioreq handling

[Jan Beulich]

There's no way to communicate back read data, so only writes can ever
be usefully specified. Ignore the field, paving the road for eventually
re-using the bit for something else in a few (many?) years time.

Signed-off-by: Jan Beulich <jbeulich@suse.com>

--- a/xen-hvm.c
+++ b/xen-hvm.c
@@ -997,6 +997,7 @@ static int handle_buffered_iopage(XenIOS
     memset(&req, 0x00, sizeof(req));
     req.state = STATE_IOREQ_READY;
     req.count = 1;
+    req.dir = IOREQ_WRITE;
 
     for (;;) {
         uint32_t rdptr = buf_page->read_pointer, wrptr;
@@ -1014,7 +1015,6 @@ static int handle_buffered_iopage(XenIOS
         req.size = 1U << buf_req->size;
         req.addr = buf_req->addr;
         req.data = buf_req->data;
-        req.dir = buf_req->dir;
         req.type = buf_req->type;
         xen_rmb();
         qw = (req.size == 8);
@@ -1031,10 +1031,12 @@ static int handle_buffered_iopage(XenIOS
         handle_ioreq(state, &req);
 
         /* Only req.data may get updated by handle_ioreq(), albeit even that
-         * should not happen as such data would never make it to the guest.
+         * should not happen as such data would never make it to the guest (we
+         * can only usefully see writes here after all).
          */
         assert(req.state == STATE_IOREQ_READY);
         assert(req.count == 1);
+        assert(req.dir == IOREQ_WRITE);
         assert(!req.data_is_ptr);
 
         atomic_add(&buf_page->read_pointer, qw + 1);

[PATCH 3/3] xen: ignore direction in bufioreq handling

[Jan Beulich]

There's no way to communicate back read data, so only writes can ever
be usefully specified. Ignore the field, paving the road for eventually
re-using the bit for something else in a few (many?) years time.

Signed-off-by: Jan Beulich <jbeulich@suse.com>

--- a/xen-hvm.c
+++ b/xen-hvm.c
@@ -997,6 +997,7 @@ static int handle_buffered_iopage(XenIOS
     memset(&req, 0x00, sizeof(req));
     req.state = STATE_IOREQ_READY;
     req.count = 1;
+    req.dir = IOREQ_WRITE;
 
     for (;;) {
         uint32_t rdptr = buf_page->read_pointer, wrptr;
@@ -1014,7 +1015,6 @@ static int handle_buffered_iopage(XenIOS
         req.size = 1U << buf_req->size;
         req.addr = buf_req->addr;
         req.data = buf_req->data;
-        req.dir = buf_req->dir;
         req.type = buf_req->type;
         xen_rmb();
         qw = (req.size == 8);
@@ -1031,10 +1031,12 @@ static int handle_buffered_iopage(XenIOS
         handle_ioreq(state, &req);
 
         /* Only req.data may get updated by handle_ioreq(), albeit even that
-         * should not happen as such data would never make it to the guest.
+         * should not happen as such data would never make it to the guest (we
+         * can only usefully see writes here after all).
          */
         assert(req.state == STATE_IOREQ_READY);
         assert(req.count == 1);
+        assert(req.dir == IOREQ_WRITE);
         assert(!req.data_is_ptr);
 
         atomic_add(&buf_page->read_pointer, qw + 1);




_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [Qemu-devel] [PATCH 3/3] xen: ignore direction in bufioreq handling

[Paul Durrant]

> -----Original Message-----
> From: Jan Beulich [mailto:JBeulich@suse.com]
> Sent: 23 November 2016 09:25
> To: qemu-devel@nongnu.org
> Cc: Anthony Perard <anthony.perard@citrix.com>; Paul Durrant
> <Paul.Durrant@citrix.com>; Stefano Stabellini <sstabellini@kernel.org>; xen-
> devel <xen-devel@lists.xenproject.org>
> Subject: [PATCH 3/3] xen: ignore direction in bufioreq handling
> 
> There's no way to communicate back read data, so only writes can ever
> be usefully specified. Ignore the field, paving the road for eventually
> re-using the bit for something else in a few (many?) years time.
> 
> Signed-off-by: Jan Beulich <jbeulich@suse.com>

Reviewed-by: Paul Durrant <paul.durrant@citrix.com>

> 
> --- a/xen-hvm.c
> +++ b/xen-hvm.c
> @@ -997,6 +997,7 @@ static int handle_buffered_iopage(XenIOS
>      memset(&req, 0x00, sizeof(req));
>      req.state = STATE_IOREQ_READY;
>      req.count = 1;
> +    req.dir = IOREQ_WRITE;
> 
>      for (;;) {
>          uint32_t rdptr = buf_page->read_pointer, wrptr;
> @@ -1014,7 +1015,6 @@ static int handle_buffered_iopage(XenIOS
>          req.size = 1U << buf_req->size;
>          req.addr = buf_req->addr;
>          req.data = buf_req->data;
> -        req.dir = buf_req->dir;
>          req.type = buf_req->type;
>          xen_rmb();
>          qw = (req.size == 8);
> @@ -1031,10 +1031,12 @@ static int handle_buffered_iopage(XenIOS
>          handle_ioreq(state, &req);
> 
>          /* Only req.data may get updated by handle_ioreq(), albeit even that
> -         * should not happen as such data would never make it to the guest.
> +         * should not happen as such data would never make it to the guest (we
> +         * can only usefully see writes here after all).
>           */
>          assert(req.state == STATE_IOREQ_READY);
>          assert(req.count == 1);
> +        assert(req.dir == IOREQ_WRITE);
>          assert(!req.data_is_ptr);
> 
>          atomic_add(&buf_page->read_pointer, qw + 1);
> 
> 

Re: [PATCH 3/3] xen: ignore direction in bufioreq handling

[Paul Durrant]

> -----Original Message-----
> From: Jan Beulich [mailto:JBeulich@suse.com]
> Sent: 23 November 2016 09:25
> To: qemu-devel@nongnu.org
> Cc: Anthony Perard <anthony.perard@citrix.com>; Paul Durrant
> <Paul.Durrant@citrix.com>; Stefano Stabellini <sstabellini@kernel.org>; xen-
> devel <xen-devel@lists.xenproject.org>
> Subject: [PATCH 3/3] xen: ignore direction in bufioreq handling
> 
> There's no way to communicate back read data, so only writes can ever
> be usefully specified. Ignore the field, paving the road for eventually
> re-using the bit for something else in a few (many?) years time.
> 
> Signed-off-by: Jan Beulich <jbeulich@suse.com>

Reviewed-by: Paul Durrant <paul.durrant@citrix.com>

> 
> --- a/xen-hvm.c
> +++ b/xen-hvm.c
> @@ -997,6 +997,7 @@ static int handle_buffered_iopage(XenIOS
>      memset(&req, 0x00, sizeof(req));
>      req.state = STATE_IOREQ_READY;
>      req.count = 1;
> +    req.dir = IOREQ_WRITE;
> 
>      for (;;) {
>          uint32_t rdptr = buf_page->read_pointer, wrptr;
> @@ -1014,7 +1015,6 @@ static int handle_buffered_iopage(XenIOS
>          req.size = 1U << buf_req->size;
>          req.addr = buf_req->addr;
>          req.data = buf_req->data;
> -        req.dir = buf_req->dir;
>          req.type = buf_req->type;
>          xen_rmb();
>          qw = (req.size == 8);
> @@ -1031,10 +1031,12 @@ static int handle_buffered_iopage(XenIOS
>          handle_ioreq(state, &req);
> 
>          /* Only req.data may get updated by handle_ioreq(), albeit even that
> -         * should not happen as such data would never make it to the guest.
> +         * should not happen as such data would never make it to the guest (we
> +         * can only usefully see writes here after all).
>           */
>          assert(req.state == STATE_IOREQ_READY);
>          assert(req.count == 1);
> +        assert(req.dir == IOREQ_WRITE);
>          assert(!req.data_is_ptr);
> 
>          atomic_add(&buf_page->read_pointer, qw + 1);
> 
> 


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [PATCH 3/3] xen: ignore direction in bufioreq handling

[Stefano Stabellini]

On Wed, 23 Nov 2016, Paul Durrant wrote:
> > -----Original Message-----
> > From: Jan Beulich [mailto:JBeulich@suse.com]
> > Sent: 23 November 2016 09:25
> > To: qemu-devel@nongnu.org
> > Cc: Anthony Perard <anthony.perard@citrix.com>; Paul Durrant
> > <Paul.Durrant@citrix.com>; Stefano Stabellini <sstabellini@kernel.org>; xen-
> > devel <xen-devel@lists.xenproject.org>
> > Subject: [PATCH 3/3] xen: ignore direction in bufioreq handling
> > 
> > There's no way to communicate back read data, so only writes can ever
> > be usefully specified. Ignore the field, paving the road for eventually
> > re-using the bit for something else in a few (many?) years time.
> > 
> > Signed-off-by: Jan Beulich <jbeulich@suse.com>
> 
> Reviewed-by: Paul Durrant <paul.durrant@citrix.com>

Acked-by: Stefano Stabellini <sstabellini@kernel.org>


> > 
> > --- a/xen-hvm.c
> > +++ b/xen-hvm.c
> > @@ -997,6 +997,7 @@ static int handle_buffered_iopage(XenIOS
> >      memset(&req, 0x00, sizeof(req));
> >      req.state = STATE_IOREQ_READY;
> >      req.count = 1;
> > +    req.dir = IOREQ_WRITE;
> > 
> >      for (;;) {
> >          uint32_t rdptr = buf_page->read_pointer, wrptr;
> > @@ -1014,7 +1015,6 @@ static int handle_buffered_iopage(XenIOS
> >          req.size = 1U << buf_req->size;
> >          req.addr = buf_req->addr;
> >          req.data = buf_req->data;
> > -        req.dir = buf_req->dir;
> >          req.type = buf_req->type;
> >          xen_rmb();
> >          qw = (req.size == 8);
> > @@ -1031,10 +1031,12 @@ static int handle_buffered_iopage(XenIOS
> >          handle_ioreq(state, &req);
> > 
> >          /* Only req.data may get updated by handle_ioreq(), albeit even that
> > -         * should not happen as such data would never make it to the guest.
> > +         * should not happen as such data would never make it to the guest (we
> > +         * can only usefully see writes here after all).
> >           */
> >          assert(req.state == STATE_IOREQ_READY);
> >          assert(req.count == 1);
> > +        assert(req.dir == IOREQ_WRITE);
> >          assert(!req.data_is_ptr);
> > 
> >          atomic_add(&buf_page->read_pointer, qw + 1);
> > 
> > 
> 

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [Qemu-devel] [PATCH 3/3] xen: ignore direction in bufioreq handling

[Stefano Stabellini]

On Wed, 23 Nov 2016, Paul Durrant wrote:
> > -----Original Message-----
> > From: Jan Beulich [mailto:JBeulich@suse.com]
> > Sent: 23 November 2016 09:25
> > To: qemu-devel@nongnu.org
> > Cc: Anthony Perard <anthony.perard@citrix.com>; Paul Durrant
> > <Paul.Durrant@citrix.com>; Stefano Stabellini <sstabellini@kernel.org>; xen-
> > devel <xen-devel@lists.xenproject.org>
> > Subject: [PATCH 3/3] xen: ignore direction in bufioreq handling
> > 
> > There's no way to communicate back read data, so only writes can ever
> > be usefully specified. Ignore the field, paving the road for eventually
> > re-using the bit for something else in a few (many?) years time.
> > 
> > Signed-off-by: Jan Beulich <jbeulich@suse.com>
> 
> Reviewed-by: Paul Durrant <paul.durrant@citrix.com>

Acked-by: Stefano Stabellini <sstabellini@kernel.org>


> > 
> > --- a/xen-hvm.c
> > +++ b/xen-hvm.c
> > @@ -997,6 +997,7 @@ static int handle_buffered_iopage(XenIOS
> >      memset(&req, 0x00, sizeof(req));
> >      req.state = STATE_IOREQ_READY;
> >      req.count = 1;
> > +    req.dir = IOREQ_WRITE;
> > 
> >      for (;;) {
> >          uint32_t rdptr = buf_page->read_pointer, wrptr;
> > @@ -1014,7 +1015,6 @@ static int handle_buffered_iopage(XenIOS
> >          req.size = 1U << buf_req->size;
> >          req.addr = buf_req->addr;
> >          req.data = buf_req->data;
> > -        req.dir = buf_req->dir;
> >          req.type = buf_req->type;
> >          xen_rmb();
> >          qw = (req.size == 8);
> > @@ -1031,10 +1031,12 @@ static int handle_buffered_iopage(XenIOS
> >          handle_ioreq(state, &req);
> > 
> >          /* Only req.data may get updated by handle_ioreq(), albeit even that
> > -         * should not happen as such data would never make it to the guest.
> > +         * should not happen as such data would never make it to the guest (we
> > +         * can only usefully see writes here after all).
> >           */
> >          assert(req.state == STATE_IOREQ_READY);
> >          assert(req.count == 1);
> > +        assert(req.dir == IOREQ_WRITE);
> >          assert(!req.data_is_ptr);
> > 
> >          atomic_add(&buf_page->read_pointer, qw + 1);
> > 
> > 
>