[bug report] Patch "perf tools: Fix pattern matching for same substring in different pmu type" broken

[bug report] Patch "perf tools: Fix pattern matching for same substring in different pmu type" broken

[John Garry]

Hi guys,

The named patch has broken PMU alias matching on my arm64 system.

Specifically it is broken for when multiple tokens are used in the 
alias. For example, alias "hisi_sccl,l3c" would previously match for PMU 
"hisi_sccl3_l3c7", but that no longer works.

In my example, in looking at the code, the callchain 
pmu_uncore_alias_match("hisi_sccl,l3c", "hisi_sccl3_l3c7") -> 
per_pmu__valid_suffix("hisi_sccl3_l3c7", "hisi_sccl") fails in the 
following check:

static bool perf_pmu__valid_suffix(char *pmu_name, char *tok)
{
	char *p;
	...
	p = pmu_name + strlen(tok);
	...
	if (*p != '_') //here
		return false;
}

This check assumes the first token must be followed by a '_', but it is 
possibly a numeric.

Please let me know how this should work. Previously it would match on 
the tokens, ignoring numerics and '_'.

As an aside, I'll look at why our testcases don't cover this scenario 
and look to add a test if necessary.

Thanks,
john

Ps, please cc linux-perf-users@vger.kernel.org as in the MAINTAINERS 
file in future, as not all subscribe to the open kernel list (and so 
cannot easily reply directly).

Re: [bug report] Patch "perf tools: Fix pattern matching for same substring in different pmu type" broken

[Arnaldo Carvalho de Melo]

Em Mon, Jul 19, 2021 at 03:13:06PM +0100, John Garry escreveu:
> Hi guys,
> 
> The named patch has broken PMU alias matching on my arm64 system.

Hi Jin, can you please address this report? Otherwise I'll have to
revert the patch in my next pull req to Linus :-\

- Arnaldo
 
> Specifically it is broken for when multiple tokens are used in the alias.
> For example, alias "hisi_sccl,l3c" would previously match for PMU
> "hisi_sccl3_l3c7", but that no longer works.
> 
> In my example, in looking at the code, the callchain
> pmu_uncore_alias_match("hisi_sccl,l3c", "hisi_sccl3_l3c7") ->
> per_pmu__valid_suffix("hisi_sccl3_l3c7", "hisi_sccl") fails in the following
> check:
> 
> static bool perf_pmu__valid_suffix(char *pmu_name, char *tok)
> {
> 	char *p;
> 	...
> 	p = pmu_name + strlen(tok);
> 	...
> 	if (*p != '_') //here
> 		return false;
> }
> 
> This check assumes the first token must be followed by a '_', but it is
> possibly a numeric.
> 
> Please let me know how this should work. Previously it would match on the
> tokens, ignoring numerics and '_'.
> 
> As an aside, I'll look at why our testcases don't cover this scenario and
> look to add a test if necessary.
> 
> Thanks,
> john
> 
> Ps, please cc linux-perf-users@vger.kernel.org as in the MAINTAINERS file in
> future, as not all subscribe to the open kernel list (and so cannot easily
> reply directly).
> 

-- 

- Arnaldo

Re: [bug report] Patch "perf tools: Fix pattern matching for same substring in different pmu type" broken

[Jin, Yao]

Hi John, Hi Arnaldo,

On 7/19/2021 10:13 PM, John Garry wrote:
> Hi guys,
> 
> The named patch has broken PMU alias matching on my arm64 system.
> 
> Specifically it is broken for when multiple tokens are used in the alias. For example, alias 
> "hisi_sccl,l3c" would previously match for PMU "hisi_sccl3_l3c7", but that no longer works.
> 
> In my example, in looking at the code, the callchain pmu_uncore_alias_match("hisi_sccl,l3c", 
> "hisi_sccl3_l3c7") -> per_pmu__valid_suffix("hisi_sccl3_l3c7", "hisi_sccl") fails in the following 
> check:
> 
> static bool perf_pmu__valid_suffix(char *pmu_name, char *tok)
> {
>      char *p;
>      ...
>      p = pmu_name + strlen(tok);
>      ...
>      if (*p != '_') //here
>          return false;
> }
> 
> This check assumes the first token must be followed by a '_', but it is possibly a numeric.
> 

It looks that the PMU alias format on arm64 has big difference than the format on x86. My new idea 
is we create a x86 specific perf_pmu__valid_suffix, and for other arch, the weak function always 
returns true. That will not change original behavior.

What do you think?

> Please let me know how this should work. Previously it would match on the tokens, ignoring numerics 
> and '_'.
> 
> As an aside, I'll look at why our testcases don't cover this scenario and look to add a test if 
> necessary.
> 
> Thanks,
> john
> 
> Ps, please cc linux-perf-users@vger.kernel.org as in the MAINTAINERS file in future, as not all 
> subscribe to the open kernel list (and so cannot easily reply directly).
> 

Sorry about that! I will remember that, cc linux-perf-users@vger.kernel.org in next patch.

Thanks
Jin Yao