Re: [PATCH] Eclair: scan the last xen-syms linking pass on Arm

[Jan Beulich <jbeulich@xxxxxxxx>]

On 07.10.2026 07:45, Nicola Vetrini wrote:
> On 2026-10-06 18:09, Dmytro Prokopchuk1 wrote:
>> On 10/6/26 18:03, Jan Beulich wrote:
>>> On 06.10.2026 16:23, Dmytro Prokopchuk1 wrote:
>>>> Hiding .xen-syms.[013] assumed the almost-final image is always pass 
>>>> 2.
>>>> On x86 LAST_LINKING_PASS is 2, so .xen-syms.2 stays visible. On Arm
>>>> LAST_LINKING_PASS is 3 and the final image is .xen-syms.3, which was
>>>> hidden. Pass 2 is usually only a symlink, so Eclair saw no program 
>>>> link
>>>> and reported every translation unit as unreachable.
>>>
>>> When I did that work, rule 2.1 ended up clean for both ARM64-* 
>>> analysis
>>> jobs (iirc). Upon re-checking I now see that ARM64-amd has 177 
>>> violations
>>> (apparently that's what you talk about above), but ARM64-allcode has 
>>> none
>>> (https://gitlab.com/xen-project/hardware/xen/-/pipelines/2916569232).
>>>
>>>> Stop hiding pass 3. Passes 0 and 1 remain excluded.
>>>
>>> This will cause scanning of both passes 2 and 3, i.e. consume more 
>>> time
>>> and report duplicate violations when both passes are real linking 
>>> steps.
>>> I think we want to be smarter than that, to really scan exactly once. 
>>> I
>>> don't have a good idea just yet how to achieve that, though (and I was
>>> wondering already when putting together the original patch).
>>>
>>> Jan
>>
>> Hmm
>>
>> At the moment
>> (https://gitlab.com/xen-project/hardware/xen/-/pipelines/2917375773) I
>> see ARM64 R2.1 violations:
>> - 177 in AMD
>> - 249 in allcode
>>
>> Anyway, we can have if-else like this (copied from
>> automation/eclair_analysis/ECLAIR/tagging.ecl):
>>
>> -setq=target,getenv("XEN_TARGET_ARCH")
>> if(string_equal(target,"x86_64"),
>>      file_tag+({"xen_syms_tmp", "^xen/\\.xen-syms\\.[013]$"}),
>>      if(string_equal(target,"arm64"),
>>          file_tag+({"xen_syms_tmp", "^xen/\\.xen-syms\\.[012]$"})
>>      )
>> )
>>
>> ^^^ test pipeline is here:
>> https://gitlab.com/xen-project/people/dimaprkp4k/xen/-/pipelines/2918571749
>>
>> BR, Dmytro.
> 
> In theory we could even generate a configuration file on the fly just 
> before invoking make. See e.g. [1] and [2]. In this way, we can 
> determine programmatically which is the linking step to consider, since 
> the logic involves some conditionals in the Makefile rules.

Hmm, yes, that looks promising.

> In any case 
> intercepting one more command that is just a symlink does not incur in a 
> big cost: only if the intercepted step is another linking command ~5 
> more minutes are needed.

Yet that's an issue, at least as long as the analysis jobs generally
determine pipeline overall latency. Plus with code size (almost) only ever
growing, that time is also only going to (almost) ever grow.

Jan

> [1] 
> https://gitlab.com/xen-project/hardware/xen/-/blob/staging/automation/eclair_analysis/ECLAIR/generate_ecl.sh?ref_type=heads
> [2] 
> https://gitlab.com/xen-project/hardware/xen/-/blob/staging/automation/eclair_analysis/ECLAIR/generate-linker-symbols.sh?ref_type=heads
> 
>