Re: [PATCH] x86/resctrl: Fix potential lockdep warning

[Xiaochen Shen]

Hi Boris,

Thank you for your kind code review. Please find my comments inline.

On 11/13/2019 19:44, Borislav Petkov wrote:
> On Thu, Nov 07, 2019 at 06:36:36AM +0800, Xiaochen Shen wrote:
> > rdtgroup_cpus_write() and mkdir_rdt_prepare() call
> > rdtgroup_kn_lock_live() -> kernfs_to_rdtgroup() to get 'rdtgrp', and
> > then call rdt_last_cmd_xxx() functions which will check if
>
> Write those names like this:
>
> rdt_last_cmd_{clear,puts,...} but not with an "xxx" which confuses
> people unfamiliar with the code.

OK. I got it. rdt_last_cmd_{clear,puts,printf}().

> > rdtgroup_mutex is held/requires its caller to hold rdtgroup_mutex.
> > But if 'rdtgrp' returned from kernfs_to_rdtgroup() is NULL,
> > rdtgroup_mutex is not held and calling rdt_last_cmd_xxx() will result
> > in a lockdep warning.
>
> That's more of a self-incurred lockdep warning. You can't be calling
> lockdep_assert_held() after a function which doesn't always grab the
> mutex. Looks like the design needs changing here...

Actually this fix covers all the cases of an audit of the calling paths
of rdt_last_cmd_{clear,puts,printf}(), to make sure we only have the
lockdep_assert_held() in places where we are sure that it must be held.
Please find more background details as below.

> > Remove rdt_last_cmd_xxx() in these two paths. Just returning error
> > should be sufficient to report to the user that the entry doesn't exist
> > any more.
>
> ... or that.
>
> In any case, you should consider fixing such patterns in the code as it
> looks sub-optimal from where I'm standing.

I would like to provide more of the background details in the commit
comment in v2 patch:

-------------------
x86/resctrl: Fix potential lockdep warning

rdt_last_cmd_{clear,puts,printf}() call lockdep_assert_held() to assert
that rdtgroup_mutex is held.

During internal review of some other changes we found that there are
code paths that call rdt_last_cmd_{clear,puts}() when the rdtgroup_mutex
is not held.

An audit of calling sequences identified two different cases in
rdtgroup_kn_lock_live() which both returning NULL:
1.'rdtgrp' returned from kernfs_to_rdtgroup() is NULL, rdtgroup_mutex
is not held.
2.'rdtgrp' is being deleted, rdtgroup_mutex is held.

Checking all call sites of rdt_last_cmd_{clear,puts,printf}() found two
code paths where rdtgroup_mutex is not held: rdtgroup_cpus_write() and
mkdir_rdt_prepare().

Fix by removing rdt_last_cmd_{clear,puts}() in these two paths. Just
returning error should be sufficient to report to the user that the
entry doesn't exist any more.

Fixes: 94457b36e8a5 ("x86/intel_rdt: Add diagnostics when writing the 
cpus file")
Fixes: cfd0f34e4cd5 ("x86/intel_rdt: Add diagnostics when making 
directories")
Signed-off-by: Xiaochen Shen <xiaochen.shen@xxxxxxxxx>
Reviewed-by: Tony Luck <tony.luck@xxxxxxxxx>
Reviewed-by: Fenghua Yu <fenghua.yu@xxxxxxxxx>
Reviewed-by: Reinette Chatre <reinette.chatre@xxxxxxxxx>
-------------------
Updated commit comment to provide additional context on how these were
found.

> Thx.

--
Best regards,
Xiaochen