We are using test_and_* operations on the status and flag fields of
struct sock_mapping. However, these functions require the operand to be
64-bit aligned on arm64. Currently, only status is 64-bit aligned.
Make flags 64-bit aligned by introducing an explicit padding field.
Signed-off-by: Stefano Stabellini <[email protected]>
diff --git a/drivers/xen/pvcalls-front.c b/drivers/xen/pvcalls-front.c
index ca5b773..aa07b2a 100644
--- a/drivers/xen/pvcalls-front.c
+++ b/drivers/xen/pvcalls-front.c
@@ -78,6 +78,7 @@ struct sock_mapping {
#define PVCALLS_STATUS_BIND 1
#define PVCALLS_STATUS_LISTEN 2
uint8_t status;
+ uint8_t pad[7];
/*
* Internal state-machine flags.
* Only one accept operation can be inflight for a socket.
On 02/27/2018 02:54 PM, Stefano Stabellini wrote:
> We are using test_and_* operations on the status and flag fields of
> struct sock_mapping. However, these functions require the operand to be
> 64-bit aligned on arm64. Currently, only status is 64-bit aligned.
>
> Make flags 64-bit aligned by introducing an explicit padding field.
>
> Signed-off-by: Stefano Stabellini <[email protected]>
>
> diff --git a/drivers/xen/pvcalls-front.c b/drivers/xen/pvcalls-front.c
> index ca5b773..aa07b2a 100644
> --- a/drivers/xen/pvcalls-front.c
> +++ b/drivers/xen/pvcalls-front.c
> @@ -78,6 +78,7 @@ struct sock_mapping {
> #define PVCALLS_STATUS_BIND 1
> #define PVCALLS_STATUS_LISTEN 2
> uint8_t status;
> + uint8_t pad[7];
Does this guarantee alignment (for either status or flag)?
-boris
> /*
> * Internal state-machine flags.
> * Only one accept operation can be inflight for a socket.
On Tue, 27 Feb 2018, Boris Ostrovsky wrote:
> On 02/27/2018 02:54 PM, Stefano Stabellini wrote:
> > We are using test_and_* operations on the status and flag fields of
> > struct sock_mapping. However, these functions require the operand to be
> > 64-bit aligned on arm64. Currently, only status is 64-bit aligned.
> >
> > Make flags 64-bit aligned by introducing an explicit padding field.
> >
> > Signed-off-by: Stefano Stabellini <[email protected]>
> >
> > diff --git a/drivers/xen/pvcalls-front.c b/drivers/xen/pvcalls-front.c
> > index ca5b773..aa07b2a 100644
> > --- a/drivers/xen/pvcalls-front.c
> > +++ b/drivers/xen/pvcalls-front.c
> > @@ -78,6 +78,7 @@ struct sock_mapping {
> > #define PVCALLS_STATUS_BIND 1
> > #define PVCALLS_STATUS_LISTEN 2
> > uint8_t status;
> > + uint8_t pad[7];
>
> Does this guarantee alignment (for either status or flag)?
Yes: status is part of a struct and a union. Unions and structs have the
alignment of their most aligned type. In this case they are 64-bit
aligned, as some of the fields are pointers.
The padding makes sure that flags is 1+7 bytes from it.
> > /*
> > * Internal state-machine flags.
> > * Only one accept operation can be inflight for a socket.
>
On 02/27/2018 04:32 PM, Stefano Stabellini wrote:
> On Tue, 27 Feb 2018, Boris Ostrovsky wrote:
>> On 02/27/2018 02:54 PM, Stefano Stabellini wrote:
>>> We are using test_and_* operations on the status and flag fields of
>>> struct sock_mapping. However, these functions require the operand to be
>>> 64-bit aligned on arm64. Currently, only status is 64-bit aligned.
>>>
>>> Make flags 64-bit aligned by introducing an explicit padding field.
>>>
>>> Signed-off-by: Stefano Stabellini <[email protected]>
>>>
>>> diff --git a/drivers/xen/pvcalls-front.c b/drivers/xen/pvcalls-front.c
>>> index ca5b773..aa07b2a 100644
>>> --- a/drivers/xen/pvcalls-front.c
>>> +++ b/drivers/xen/pvcalls-front.c
>>> @@ -78,6 +78,7 @@ struct sock_mapping {
>>> #define PVCALLS_STATUS_BIND 1
>>> #define PVCALLS_STATUS_LISTEN 2
>>> uint8_t status;
>>> + uint8_t pad[7];
>> Does this guarantee alignment (for either status or flag)?
> Yes: status is part of a struct and a union. Unions and structs have the
> alignment of their most aligned type. In this case they are 64-bit
> aligned, as some of the fields are pointers.
>
> The padding makes sure that flags is 1+7 bytes from it.
OK.
What about adding __attribute__((aligned(8))) to both (with a comment
explaining reasoning)?
-boris
>
>
>>> /*
>>> * Internal state-machine flags.
>>> * Only one accept operation can be inflight for a socket.
On Tue, 27 Feb 2018, Boris Ostrovsky wrote:
> On 02/27/2018 04:32 PM, Stefano Stabellini wrote:
> > On Tue, 27 Feb 2018, Boris Ostrovsky wrote:
> >> On 02/27/2018 02:54 PM, Stefano Stabellini wrote:
> >>> We are using test_and_* operations on the status and flag fields of
> >>> struct sock_mapping. However, these functions require the operand to be
> >>> 64-bit aligned on arm64. Currently, only status is 64-bit aligned.
> >>>
> >>> Make flags 64-bit aligned by introducing an explicit padding field.
> >>>
> >>> Signed-off-by: Stefano Stabellini <[email protected]>
> >>>
> >>> diff --git a/drivers/xen/pvcalls-front.c b/drivers/xen/pvcalls-front.c
> >>> index ca5b773..aa07b2a 100644
> >>> --- a/drivers/xen/pvcalls-front.c
> >>> +++ b/drivers/xen/pvcalls-front.c
> >>> @@ -78,6 +78,7 @@ struct sock_mapping {
> >>> #define PVCALLS_STATUS_BIND 1
> >>> #define PVCALLS_STATUS_LISTEN 2
> >>> uint8_t status;
> >>> + uint8_t pad[7];
> >> Does this guarantee alignment (for either status or flag)?
> > Yes: status is part of a struct and a union. Unions and structs have the
> > alignment of their most aligned type. In this case they are 64-bit
> > aligned, as some of the fields are pointers.
> >
> > The padding makes sure that flags is 1+7 bytes from it.
>
> OK.
>
> What about adding __attribute__((aligned(8))) to both (with a comment
> explaining reasoning)?
That's fine by me
> >
> >
> >>> /*
> >>> * Internal state-machine flags.
> >>> * Only one accept operation can be inflight for a socket.
>
From: Stefano Stabellini
> Sent: 27 February 2018 19:55
>
> We are using test_and_* operations on the status and flag fields of
> struct sock_mapping. However, these functions require the operand to be
> 64-bit aligned on arm64. Currently, only status is 64-bit aligned.
That sounds like a big accident just waiting to happen...
If the operand is required to be 64bit aligned why isn't it a type
that requires 64bit alignment?
> Make flags 64-bit aligned by introducing an explicit padding field.
At some point such padding will always go wrong...
David