While the endiannes is being handled correctly as indicated by the comment
above the offending line - sparse was unhappy with the missing annotation
as be64_to_cpu() expects a __be64 argument. To mitigate this annotation
issue forced annotation is introduced. Note that this patch has no impact
on the generated binary.
Signed-off-by: Nicholas Mc Guire <[email protected]>
---
Problem located by an experimental coccinelle script to locate
patters that make sparse unhappy (false positives):
net/rds/ib_recv.c:827:23: warning: cast to restricted __le64
Patch was compile-tested with: x86_64_defconfig + RDS_RDMA=m
Patch was verified not to change the binary by diffing the
generated object code before and after applying the patch.
Patch is against 5.1-rc6 (localversion-next is 20190426)
net/rds/ib_recv.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
index 7055985..a070a2d 100644
--- a/net/rds/ib_recv.c
+++ b/net/rds/ib_recv.c
@@ -824,7 +824,7 @@ static void rds_ib_cong_recv(struct rds_connection *conn,
}
/* the congestion map is in little endian order */
- uncongested = le64_to_cpu(uncongested);
+ uncongested = le64_to_cpu((__force __le64)uncongested);
rds_cong_map_updated(map, uncongested);
}
--
2.1.4
On 29/04/2019 07:09, Nicholas Mc Guire wrote:
> diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
> index 7055985..a070a2d 100644
> --- a/net/rds/ib_recv.c
> +++ b/net/rds/ib_recv.c
> @@ -824,7 +824,7 @@ static void rds_ib_cong_recv(struct rds_connection *conn,
> }
>
> /* the congestion map is in little endian order */
> - uncongested = le64_to_cpu(uncongested);
> + uncongested = le64_to_cpu((__force __le64)uncongested);
>
> rds_cong_map_updated(map, uncongested);
> }
Again, a __force cast doesn't seem necessary here. It looks like the
code is just using the wrong types; if all of src, dst and uncongested
were __le64 instead of uint64_t, and the last two lines replaced with
rds_cong_map_updated(map, le64_to_cpu(uncongested)); then the semantics
would be kept with neither sparse errors nor __force.
__force is almost never necessary and mostly just masks other bugs or
endianness confusion in the surrounding code. Instead of adding a
__force, either fix the code to be sparse-clean or leave the sparse
warning in place so that future developers know there's something not
right.
-Ed
On Mon, Apr 29, 2019 at 12:00:06PM +0100, Edward Cree wrote:
> On 29/04/2019 07:09, Nicholas Mc Guire wrote:
> > diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
> > index 7055985..a070a2d 100644
> > --- a/net/rds/ib_recv.c
> > +++ b/net/rds/ib_recv.c
> > @@ -824,7 +824,7 @@ static void rds_ib_cong_recv(struct rds_connection *conn,
> > }
> >
> > /* the congestion map is in little endian order */
> > - uncongested = le64_to_cpu(uncongested);
> > + uncongested = le64_to_cpu((__force __le64)uncongested);
> >
> > rds_cong_map_updated(map, uncongested);
> > }
> Again, a __force cast doesn't seem necessary here.? It looks like the
> ?code is just using the wrong types; if all of src, dst and uncongested
> ?were __le64 instead of uint64_t, and the last two lines replaced with
> ?rds_cong_map_updated(map, le64_to_cpu(uncongested)); then the semantics
> ?would be kept with neither sparse errors nor __force.
>
> __force is almost never necessary and mostly just masks other bugs or
> ?endianness confusion in the surrounding code.? Instead of adding a
> ?__force, either fix the code to be sparse-clean or leave the sparse
> ?warning in place so that future developers know there's something not
> ?right.
>
changing uncongested to __le64 is not an option here - it would only move
the sparse warnings to those other locatoins where the ports that
became uncongested are being or'ed into uncongested.
I'm not using __force as the prime way to silence sparse - I try to find
an alternative first - the problem is in line 805
for (k = 0; k < to_copy; k += 8) {
/* Record ports that became uncongested, ie
* bits that changed from 0 to 1. */
uncongested |= ~(*src) & *dst;
*dst++ = *src++;
}
And in this case the endianness handling does seem right.
But ok with me to leave it in as it is - if you think that the __force
here is not justified.
thanks for your comments and notably the explainations !
thx!
hofrat
alternative
On 29/04/2019 12:18, Nicholas Mc Guire wrote:
> On Mon, Apr 29, 2019 at 12:00:06PM +0100, Edward Cree wrote:
>> Again, a __force cast doesn't seem necessary here. It looks like the
>> code is just using the wrong types; if all of src, dst and uncongested
>> were __le64 instead of uint64_t, and the last two lines replaced with
>> rds_cong_map_updated(map, le64_to_cpu(uncongested)); then the semantics
>> would be kept with neither sparse errors nor __force.
>>
>> __force is almost never necessary and mostly just masks other bugs or
>> endianness confusion in the surrounding code. Instead of adding a
>> __force, either fix the code to be sparse-clean or leave the sparse
>> warning in place so that future developers know there's something not
>> right.
>>
> changing uncongested to __le64 is not an option here - it would only move
> the sparse warnings to those other locatoins where the ports that
> became uncongested are being or'ed into uncongested.
That's why I say to change *src and *dst too. Sparse won't mind the
conversion from void * to __le64 * when they're assigned, and the only
operations we do on them...
> uncongested |= ~(*src) & *dst;
> *dst++ = *src++;
... are some bitwise ops on the values (bitwise ops are legal in any
endianness) and incrementation of the pointers (which cares only about
the pointee size, not type).
-Ed
On Mon, Apr 29, 2019 at 01:18:36PM +0200, Nicholas Mc Guire wrote:
> changing uncongested to __le64 is not an option here - it would only move
> the sparse warnings to those other locatoins where the ports that
> became uncongested are being or'ed into uncongested.
Than fix that a well. Either by throwing in a conversion, or
add {be,le}XX_{and,or} helpers.
On Mon, Apr 29, 2019 at 01:02:31PM +0100, Edward Cree wrote:
> ... are some bitwise ops on the values (bitwise ops are legal in any
> ?endianness) and incrementation of the pointers (which cares only about
> ?the pointee size, not type).
Oh, true. That is why the underlying annotation is called __bitwise :)
I'll take my previous comment back.
On Mon, Apr 29, 2019 at 05:21:32AM -0700, Christoph Hellwig wrote:
> On Mon, Apr 29, 2019 at 01:18:36PM +0200, Nicholas Mc Guire wrote:
> > changing uncongested to __le64 is not an option here - it would only move
> > the sparse warnings to those other locatoins where the ports that
> > became uncongested are being or'ed into uncongested.
>
> Than fix that a well. Either by throwing in a conversion, or
> add {be,le}XX_{and,or} helpers.
ok - that is an option in that case - will try that route
thx!
hofrat