Intel i40e code finds smatch bug

Intel i40e code finds smatch bug

[Dan McDonald]

[-- Attachment #1: Type: text/plain, Size: 703 bytes --]

We in illumos are in the midst of updating our i40e code from Intel upstream. We found a bug in smatch while building it. It's documented here in:

	https://www.illumos.org/issues/13457

A colleague recommended I include a toy program, try the smatch `master` branch on a Linux distro, and submit the toy to this list.  Its source is attached, ready to compile (without incident) and smatch (with one incident, but not two).  You'll see two cases of potential failure, but only one (the one Intel source uses to confirm proper compilation) breaks smatch.  Look for the KEBE comments.

Once this bug gets fixed, I suspect we'll also be updating our branch of smatch as well.  :)

Thanks,
Dan


[-- Attachment #2: pack.c --]
[-- Type: application/octet-stream, Size: 1646 bytes --]

#include <sys/types.h>
#include <inttypes.h>
#include <stdio.h>

#define u8 uint8_t
#define __le16 uint16_t
#define __le32 uint32_t

#pragma pack(1)
/* Run PHY Activity (0x0626) */
struct i40e_aqc_run_phy_activity {
        u8      cmd_flags;
        __le16  activity_id;
#define I40E_AQ_RUN_PHY_ACT_ID_USR_DFND                 0x10
        u8      reserved;
        union {
                struct {
                        __le32  dnl_opcode;
#define I40E_AQ_RUN_PHY_ACT_DNL_OPCODE_GET_EEE_STAT_DUR 0x801a
#define I40E_AQ_RUN_PHY_ACT_DNL_OPCODE_GET_EEE_STAT     0x801b
#define I40E_AQ_RUN_PHY_ACT_DNL_OPCODE_GET_EEE_DUR      0x1801b
                        __le32  data;
                        u8      reserved2[4];
                } cmd;
                struct {
                        __le32  cmd_status;
#define I40E_AQ_RUN_PHY_ACT_CMD_STAT_SUCC               0x4
#define I40E_AQ_RUN_PHY_ACT_CMD_STAT_MASK               0xFFFF
                        __le32  data0;
                        __le32  data1;
                } resp;
   } params;
};
#pragma pack()

#define I40E_CHECK_STRUCT_LEN(n, X) enum i40e_static_assert_enum_##X \
        { i40e_static_assert_##X = (n)/((sizeof(struct X) == (n)) ? 1 : 0) }

/* XXX KEBE SAYS this one fails! */
I40E_CHECK_STRUCT_LEN(16, i40e_aqc_run_phy_activity);

int
main(void)
{
	struct i40e_aqc_run_phy_activity foo;

	printf("foo == 0x%p, foo.activity_id = 0x%p, size == %d, tripval = %d\n",
	    &foo, &foo.activity_id, sizeof (struct i40e_aqc_run_phy_activity),
	    /* XXX KEBE SAYS but this one doesn't. */
	    16 / ((sizeof (struct i40e_aqc_run_phy_activity) == 16) ? 1 : 0) );
	return (0);
}

Re: Intel i40e code finds smatch bug

[Dan Carpenter]

On Tue, Jan 19, 2021 at 12:22:50PM -0500, Dan McDonald wrote:
> We in illumos are in the midst of updating our i40e code from Intel upstream. We found a bug in smatch while building it. It's documented here in:
> 
> 	https://www.illumos.org/issues/13457
> 
> A colleague recommended I include a toy program, try the smatch `master` branch on a Linux distro, and submit the toy to this list.  Its source is attached, ready to compile (without incident) and smatch (with one incident, but not two).  You'll see two cases of potential failure, but only one (the one Intel source uses to confirm proper compilation) breaks smatch.  Look for the KEBE comments.
> 
> Once this bug gets fixed, I suspect we'll also be updating our branch of smatch as well.  :)
> 
> Thanks,
> Dan
> 

Unfortunately Sparse doesn't support #pragma pack.  This was reported to
the Sparse list last February and the response was that pragma pack is
disgusting so don't use it.

http://tw.spinics.net/lists/linux-sparse/msg10816.html

I wish I could give you better news than that...

Luk recently was working on better support of packed structs (which have
been a known issue for a while) but I doubt that will help here.

regards,
dan carpenter

Re: Intel i40e code finds smatch bug

[Dan McDonald]

On Jan 19, 2021, at 1:55 PM, Dan Carpenter <dan.carpenter@oracle.com> wrote:
> 
> On Tue, Jan 19, 2021 at 12:22:50PM -0500, Dan McDonald wrote:
>> We in illumos are in the midst of updating our i40e code from Intel upstream. We found a bug in smatch while building it. It's documented here in:
>> 
>> 	https://www.illumos.org/issues/13457
>> 
>> A colleague recommended I include a toy program, try the smatch `master` branch on a Linux distro, and submit the toy to this list. Its source is attached, ready to compile (without incident) and smatch (with one incident, but not two).  You'll see two cases of potential failure, but only one (the one Intel source uses to confirm proper compilation) breaks smatch.  Look for the KEBE comments.
>> 
>> Once this bug gets fixed, I suspect we'll also be updating our branch of smatch as well.  :)
>> 
>> Thanks,
>> Dan
>> 
> 
> Unfortunately Sparse doesn't support #pragma pack.  This was reported to
> the Sparse list last February and the response was that pragma pack is
> disgusting so don't use it.
> 
> http://tw.spinics.net/lists/linux-sparse/msg10816.html

Heh heh.

> I wish I could give you better news than that...
> 
> Luk recently was working on better support of packed structs (which have
> been a known issue for a while) but I doubt that will help here.

I saw his followup on the archive pointer you gave me (thanks for that, too, as I couldn't find it with some cursory Googling).

We have a workaround in place, since it's only in one (OEM-supplied) driver.

Thanks!
Dan

Re: Intel i40e code finds smatch bug

[Dan McDonald]

On Jan 19, 2021, at 1:55 PM, Dan Carpenter <dan.carpenter@oracle.com> wrote:
> 
> Luk recently was working on better support of packed structs (which have
> been a known issue for a while) but I doubt that will help here.

I tried using packed struct attributes instead of the pragma and no luck there either:

centos(~)[0]% diff -u pack.c pack2.c
--- pack.c	2021-01-19 17:16:13.258512270 +0000
+++ pack2.c	2021-01-19 20:49:45.886696658 +0000
@@ -6,9 +6,8 @@
 #define __le16 uint16_t
 #define __le32 uint32_t
 
-#pragma pack(1)
 /* Run PHY Activity (0x0626) */
-struct i40e_aqc_run_phy_activity {
+struct __attribute__((packed)) i40e_aqc_run_phy_activity {
         u8      cmd_flags;
         __le16  activity_id;
 #define I40E_AQ_RUN_PHY_ACT_ID_USR_DFND                 0x10
@@ -31,7 +30,6 @@
                 } resp;
    } params;
 };
-#pragma pack()
 
 #define I40E_CHECK_STRUCT_LEN(n, X) enum i40e_static_assert_enum_##X \
         { i40e_static_assert_##X = (n)/((sizeof(struct X) == (n)) ? 1 : 0) }
centos(~)[1]% 


Dan

Re: Intel i40e code finds smatch bug

[Luc Van Oostenryck]

On Tue, Jan 19, 2021 at 03:51:57PM -0500, Dan McDonald wrote:
> On Jan 19, 2021, at 1:55 PM, Dan Carpenter <dan.carpenter@oracle.com> wrote:
> > 
> > Luk recently was working on better support of packed structs (which have
> > been a known issue for a while) but I doubt that will help here.

Indeed, the support for packed structs (via the attribute) is now
complete in Sparse and was mainlined 2 weeks ago or so and ...
 
> I tried using packed struct attributes instead of the pragma and no luck there either:

now it just needs to be merged into smatch.
 
-- Luc

Re: Intel i40e code finds smatch bug

[Dan McDonald]

On Jan 19, 2021, at 4:42 PM, Luc Van Oostenryck <luc.vanoostenryck@gmail.com> wrote:
> 
> On Tue, Jan 19, 2021 at 03:51:57PM -0500, Dan McDonald wrote:
>> On Jan 19, 2021, at 1:55 PM, Dan Carpenter <dan.carpenter@oracle.com> wrote:
>>> 
>>> Luk recently was working on better support of packed structs (which have
>>> been a known issue for a while) but I doubt that will help here.
> 
> Indeed, the support for packed structs (via the attribute) is now
> complete in Sparse and was mainlined 2 weeks ago or so and ...
> 
>> I tried using packed struct attributes instead of the pragma and no luck there either:
> 
> now it just needs to be merged into smatch.

Well that's good news! Of course, one may have to mention this to Intel in regards to i40e, but that's not this mailing list's problem. Nor is having illumos update its downstream with this as well. :)

Thank you!
Dan

Re: Intel i40e code finds smatch bug

[Dan Carpenter]

On Tue, Jan 19, 2021 at 10:42:21PM +0100, Luc Van Oostenryck wrote:
> On Tue, Jan 19, 2021 at 03:51:57PM -0500, Dan McDonald wrote:
> > On Jan 19, 2021, at 1:55 PM, Dan Carpenter <dan.carpenter@oracle.com> wrote:
> > > 
> > > Luk recently was working on better support of packed structs (which have
> > > been a known issue for a while) but I doubt that will help here.
> 
> Indeed, the support for packed structs (via the attribute) is now
> complete in Sparse and was mainlined 2 weeks ago or so and ...
>  
> > I tried using packed struct attributes instead of the pragma and no luck there either:
> 
> now it just needs to be merged into smatch.
>  

Awesome!  I've merged it and will push the merge tomorrow if the tests
run well tonight.  Thanks!

regards,
dan carpenter

Re: Intel i40e code finds smatch bug

[Dan Carpenter]

Pushed!

Thanks, everyone.

regards,
dan carpenter

Re: Intel i40e code finds smatch bug

[Luc Van Oostenryck]

On Mon, Jan 25, 2021 at 06:10:19PM +0300, Dan Carpenter wrote:
> Pushed!

Great. Thank you.
- Luc 

Re: Intel i40e code finds smatch bug

[Dan McDonald]

> On Jan 25, 2021, at 10:10 AM, Dan Carpenter <dan.carpenter@oracle.com> wrote:
> 
> Pushed!
> 
> Thanks, everyone.

Thank you!

Dan