Skip to contentSkip to conversation

x86/resctrl: Fix potential lockdep warning

PATCH Applied Xiaochen Shen 3 participants 6 messages Nov 6, 2019 – Nov 19, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/resctrl: Fix potential lockdep warning

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 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.

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.

Fixes 94457b36e8a5 ("x86/intel_rdt: Add diagnostics when writing the cpus file")
Fixes cfd0f34e4cd5 ("x86/intel_rdt: Add diagnostics when making directories")
---
 arch/x86/kernel/cpu/resctrl/rdtgroup.c | 4 ----
 1 file changed, 4 deletions(-)
arch/x86/kernel/cpu/resctrl/rdtgroup.c +0 -4
index a46dee8e78db..2e3b06d6bbc6 100644
@@ -461,10 +461,8 @@ static ssize_t rdtgroup_cpus_write(struct kernfs_open_file *of,
 	}
 
 	rdtgrp = rdtgroup_kn_lock_live(of->kn);
-	rdt_last_cmd_clear();
 	if (!rdtgrp) {
 		ret = -ENOENT;
-		rdt_last_cmd_puts("Directory was removed\n");
 		goto unlock;
 	}
 
@@ -2648,10 +2646,8 @@ static int mkdir_rdt_prepare(struct kernfs_node *parent_kn,
 	int ret;
 
 	prdtgrp = rdtgroup_kn_lock_live(prgrp_kn);
-	rdt_last_cmd_clear();
 	if (!prdtgrp) {
 		ret = -ENODEV;
-		rdt_last_cmd_puts("Directory was removed\n");
 		goto out_unlock;
 	}
 
Signature · 1 line
1.8.3.1
maintainer
#1 lore raw
Quoted 3 lines · Xiaochen Shen

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.

Quoted 4 lines

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...

Quoted 3 lines

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.

Thx.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
author Reviewed-by Tony Luck Reviewed-by Fenghua Yu Reviewed-by Reinette Chatre
#2 lore raw

Hi Boris,

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

Quoted 9 lines · Borislav Petkov
Quoted 3 lines · Xiaochen Shen

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}().

Quoted 9 lines
Quoted 5 lines

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.

Quoted 9 lines
Quoted 4 lines

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")

-------------------
Updated commit comment to provide additional context on how these were found.

Quoted 3 lines

Thx.

Signature · 2 lines
Best regards,
Xiaochen
maintainer
#3 lore raw
Quoted 3 lines · Xiaochen Shen

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.

That's kinda what I suggested, isn't it?

All I meant was, not to have a

rdtgroup_kn_lock_live()

call in the code as this function does *not* unconditionally grab the rdtgroup_mutex. And then call a function which unconditionally checks whether the mutex is held.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
#4 lore raw
Quoted 7 lines · Borislav Petkov
Quoted 3 lines · Xiaochen Shen

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.

That's kinda what I suggested, isn't it?

Quoted 8 lines

All I meant was, not to have a

rdtgroup_kn_lock_live()

call in the code as this function does *not* unconditionally grab the rdtgroup_mutex. And then call a function which unconditionally checks whether the mutex is held.

Hi Boris,

Thank you for your good suggestion. I will try to follow up if we could improve the code in call sites of rdtgroup_kn_lock_live() in separate patch.

In my opinion, the potential lockdep issues in all call sites of rdt_last_cmd_{clear,puts,...}() have been fixed in this patch.

Thank you.

Signature · 2 lines
Best regards,
Xiaochen
merge bot Applied
#5 lore raw
In reply to Xiaochen Shen
The following commit has been merged into the x86/urgent branch of tip:

The following commit has been merged into the x86/urgent branch of tip:

Commit-ID:     c8eafe1495303bfd0eedaa8156b1ee9082ee9642
Gitweb:        https://git.kernel.org/tip/c8eafe1495303bfd0eedaa8156b1ee9082ee9642
Author:        Xiaochen Shen <[email protected]>
AuthorDate:    Thu, 07 Nov 2019 06:36:36 +08:00
Committer:     Borislav Petkov <[email protected]>
CommitterDate: Wed, 13 Nov 2019 12:34:44 +01:00

x86/resctrl: Fix potential lockdep warning

rdtgroup_cpus_write() and mkdir_rdt_prepare() call rdtgroup_kn_lock_live() -> kernfs_to_rdtgroup() to get 'rdtgrp', and then call the rdt_last_cmd_{clear,puts,...}() functions which will check if 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_{clear,puts,...}() will result in a self-incurred, potential lockdep warning.

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

[ bp: Massage. ]

Fixes 94457b36e8a5 ("x86/intel_rdt: Add diagnostics when writing the cpus file")
Fixes cfd0f34e4cd5 ("x86/intel_rdt: Add diagnostics when making directories")
Cc "H. Peter Anvin" <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc x86-ml <[email protected]>
---
 arch/x86/kernel/cpu/resctrl/rdtgroup.c | 4 ----
 1 file changed, 4 deletions(-)
arch/x86/kernel/cpu/resctrl/rdtgroup.c +0 -4
index a46dee8..2e3b06d 100644
@@ -461,10 +461,8 @@ static ssize_t rdtgroup_cpus_write(struct kernfs_open_file *of,
 	}
 
 	rdtgrp = rdtgroup_kn_lock_live(of->kn);
-	rdt_last_cmd_clear();
 	if (!rdtgrp) {
 		ret = -ENOENT;
-		rdt_last_cmd_puts("Directory was removed\n");
 		goto unlock;
 	}
 
@@ -2648,10 +2646,8 @@ static int mkdir_rdt_prepare(struct kernfs_node *parent_kn,
 	int ret;
 
 	prdtgrp = rdtgroup_kn_lock_live(prgrp_kn);
-	rdt_last_cmd_clear();
 	if (!prdtgrp) {
 		ret = -ENODEV;
-		rdt_last_cmd_puts("Directory was removed\n");
 		goto out_unlock;
 	}