Skip to contentSkip to conversation

Staging: lustre: sparse static warning fix

PATCH In review Adrian Remonda 5 participants 15 messages May 18, 2015 – May 22, 2015 1 message not in the archive Staging

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw

PATCH 1/4 Staging: lustre: sparse static warning fix

This patch clears the warning given by sparse that the function should be static. Done by moving the declaration of the structure 'nrs_conf_fifo' into the header 'ptlrpc_internal.h'

modified: drivers/staging/lustre/lustre/ptlrpc/nrs.c
modified: drivers/staging/lustre/lustre/ptlrpc/ptlrpc_internal.h

---
 drivers/staging/lustre/lustre/ptlrpc/nrs.c             | 4 ----
 drivers/staging/lustre/lustre/ptlrpc/ptlrpc_internal.h | 4 ++++
 2 files changed, 4 insertions(+), 4 deletions(-)
drivers/staging/lustre/lustre/ptlrpc/nrs.c +0 -4
index 63a05f4a902d..159db41c53a0 100644
@@ -1698,10 +1698,6 @@ out:
 	return rc;
 }
 
-
-/* ptlrpc/nrs_fifo.c */
-extern struct ptlrpc_nrs_pol_conf nrs_conf_fifo;
-
 /**
  * Adds all policies that ship with the ptlrpc module, to NRS core's list of
  * policies \e nrs_core.nrs_policies.
drivers/staging/lustre/lustre/ptlrpc/ptlrpc_internal.h +4 -0
index a66dc3c6da41..e07e2aacc62c 100644
@@ -309,4 +309,8 @@ static inline void ptlrpc_reqset_put(struct ptlrpc_request_set *set)
 	if (atomic_dec_and_test(&set->set_refcount))
 		OBD_FREE_PTR(set);
 }
+
+/* ptlrpc/nrs_fifo.c */
+extern struct ptlrpc_nrs_pol_conf nrs_conf_fifo;
+
 #endif /* PTLRPC_INTERNAL_H */
Signature · 1 line
2.1.4
#2 lore raw

PATCH 2/4 Staging: lustre: sparse static warning fix

This patch clears the warning given by sparse that the function should be static. Done by moving the declaration of the structure 'nrs_core' into the header 'ptlrpc_internal.h'

modified: drivers/staging/lustre/lustre/ptlrpc/lproc_ptlrpc.c
modified: drivers/staging/lustre/lustre/ptlrpc/ptlrpc_internal.h

---
 drivers/staging/lustre/lustre/ptlrpc/lproc_ptlrpc.c    | 6 ------
 drivers/staging/lustre/lustre/ptlrpc/ptlrpc_internal.h | 7 +++++++
 2 files changed, 7 insertions(+), 6 deletions(-)
drivers/staging/lustre/lustre/ptlrpc/lproc_ptlrpc.c +0 -6
index aeceef5152ac..300310e064f5 100644
@@ -417,12 +417,6 @@ ptlrpc_lprocfs_threads_max_seq_write(struct file *file,
 LPROC_SEQ_FOPS(ptlrpc_lprocfs_threads_max);
 
 /**
- * \addtogoup nrs
- * @{
- */
-extern struct nrs_core nrs_core;
-
-/**
  * Translates \e ptlrpc_nrs_pol_state values to human-readable strings.
  *
  * \param[in] state The policy state
drivers/staging/lustre/lustre/ptlrpc/ptlrpc_internal.h +7 -0
index e07e2aacc62c..f9bfcff2ac9b 100644
@@ -313,4 +313,11 @@ static inline void ptlrpc_reqset_put(struct ptlrpc_request_set *set)
 /* ptlrpc/nrs_fifo.c */
 extern struct ptlrpc_nrs_pol_conf nrs_conf_fifo;
 
+/**
+ * \addtogoup nrs
+ * @{
+ */
+extern struct nrs_core nrs_core;
+
+
 #endif /* PTLRPC_INTERNAL_H */
Signature · 1 line
2.1.4
#3 lore raw

PATCH 3/4 Staging: lustre: Fixed typo

In the explanation of the function the name of the function was incorrect

---
 drivers/staging/lustre/lustre/ptlrpc/nrs.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/staging/lustre/lustre/ptlrpc/nrs.c +1 -1
index 159db41c53a0..43da95f0bce2 100644
@@ -478,7 +478,7 @@ static void nrs_resource_get_safe(struct ptlrpc_nrs *nrs,
  *
  * \param resp	the resource hierarchy that is being released
  *
- * \see ptlrpcnrs_req_hp_move()
+ * \see ptlrpc_nrs_req_hp_move()
  * \see ptlrpc_nrs_req_finalize()
  */
 static void nrs_resource_put_safe(struct ptlrpc_nrs_resource **resp)
Signature · 1 line
2.1.4
#4 lore raw

PATCH 4/4 Staging: lustre: sparse lock warning fix

Fixed sparse warning: context imbalance in 'nrs_resource_put_safe' - 'different lock contexts for basic block' by releasing the lock on each iteration of the for loop.

---
 drivers/staging/lustre/lustre/ptlrpc/nrs.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)
drivers/staging/lustre/lustre/ptlrpc/nrs.c +2 -4
index 43da95f0bce2..3a1722437ca1 100644
@@ -503,13 +503,11 @@ static void nrs_resource_put_safe(struct ptlrpc_nrs_resource **resp)
 
 		if (nrs == NULL) {
 			nrs = pols[i]->pol_nrs;
-			spin_lock(&nrs->nrs_lock);
 		}
+		spin_lock(&nrs->nrs_lock);
 		nrs_policy_put_locked(pols[i]);
-	}
-
-	if (nrs != NULL)
 		spin_unlock(&nrs->nrs_lock);
+	}
 }
 
 /**
Signature · 1 line
2.1.4
#5 lore raw

Re: [PATCH 4/4] Staging: lustre: sparse lock warning fix

Quoted 4 lines · Adrian Remonda

Fixed sparse warning: context imbalance in 'nrs_resource_put_safe' - 'different lock contexts for basic block' by releasing the lock on each iteration of the for loop.

That changelog doesn't sound correct at all. That's not a correct motivation or explanation.

I reviewed the patch and it's likely going to cause dead locks. The code is trying to take the spinlock for the first pointer in the array and release it at the end. Now it takes the first pointer's spinlock a bunch of times (dead lock) and releases it once (will not happen because we are already dead).

regards,
dan carpenter

#6 lore raw

Re: [PATCH 4/4] Staging: lustre: sparse lock warning fix

Quoted 6 lines · Dan Carpenter
Quoted 4 lines · Adrian Remonda

Fixed sparse warning: context imbalance in 'nrs_resource_put_safe' - 'different lock contexts for basic block' by releasing the lock on each iteration of the for loop.

Quoted 8 lines

That changelog doesn't sound correct at all. That's not a correct motivation or explanation.

I reviewed the patch and it's likely going to cause dead locks. The code is trying to take the spinlock for the first pointer in the array and release it at the end. Now it takes the first pointer's spinlock a bunch of times (dead lock) and releases it once (will not happen because we are already dead).

It isn't clear to me what the checkpatch complaint actually means? Is it that the spin_lock() and spin_unlock() calls have different amounts of indentation?

Cheers, Andreas

Signature · 4 lines
Andreas Dilger

Lustre Software Architect
Intel High Performance Data Division
#7 lore raw

Re: [PATCH 4/4] Staging: lustre: sparse lock warning fix

Quoted 17 lines · Andreas
Quoted 6 lines · Dan Carpenter
Quoted 4 lines · Adrian Remonda

Fixed sparse warning: context imbalance in 'nrs_resource_put_safe' - 'different lock contexts for basic block' by releasing the lock on each iteration of the for loop.

Quoted 8 lines

That changelog doesn't sound correct at all. That's not a correct motivation or explanation.

I reviewed the patch and it's likely going to cause dead locks. The code is trying to take the spinlock for the first pointer in the array and release it at the end. Now it takes the first pointer's spinlock a bunch of times (dead lock) and releases it once (will not happen because we are already dead).

Quoted 4 lines

It isn't clear to me what the checkpatch complaint actually means? Is it that the spin_lock() and spin_unlock() calls have different amounts of indentation?

It's not a checkpatch.pl warning, it's a Sparse warning. Sparse is crappy at reporting locking bugs. It's mostly false positives.

I think it's saying that some paths lock and unlock some don't.

Smatch is also fairly crappy at finding locking bugs, unfortunately. I need to re-write it using modern features and cross function analysis.

regards,
dan carpenter

#8 lore raw

Re: [PATCH 4/4] Staging: lustre: sparse lock warning fix

In Smatch, it the equivalent warning is turned off by default because there are too many false positives, but you can enable it with the --spammy flag.

kchecker --spammy drivers/staging/lustre/lustre/ptlrpc/nrs.c

drivers/staging/lustre/lustre/ptlrpc/nrs.c:512 nrs_resource_put_safe()
warn: 'spin_lock:&nrs->nrs_lock' is sometimes locked here and sometimes unlocked.

regards,
dan carpenter

#9 lore raw

Re: [PATCH 4/4] Staging: lustre: sparse lock warning fix

Quoted 9 lines · Dan Carpenter

In Smatch, it the equivalent warning is turned off by default because there are too many false positives, but you can enable it with the --spammy flag.

kchecker --spammy drivers/staging/lustre/lustre/ptlrpc/nrs.c

drivers/staging/lustre/lustre/ptlrpc/nrs.c:512 nrs_resource_put_safe()
warn: 'spin_lock:&nrs->nrs_lock' is sometimes locked here and sometimes unlocked.

Would this be happier with something like:

        for (i = 0; i < NRS_RES_MAX; i++) {
		if (pols[i] == NULL)
			continue;
        

		if (nrs == NULL) {
			nrs = pols[i]->pol_nrs;
			if (likely(nrs != NULL)) /* make sparse happy */
				spin_lock(&nrs->nrs_lock);
		}
		nrs_policy_put_locked(pols[i]);
	}

	if (nrs != NULL)
	spin_unlock(&nrs->nrs_lock);

so that the "if" conditions are the same? The code definitely doesn't have a bug, because the lock is only locked once when nrs is first set, and only unlocked if it is set. Or is there a comment to put there that will quiet the static checker?

Cheers, Andreas

Signature · 4 lines
Andreas Dilger

Lustre Software Architect
Intel High Performance Data Division
#10 lore raw

Re: [PATCH 4/4] Staging: lustre: sparse lock warning fix

Quoted 30 lines · Andreas
Quoted 9 lines · Dan Carpenter

In Smatch, it the equivalent warning is turned off by default because there are too many false positives, but you can enable it with the --spammy flag.

kchecker --spammy drivers/staging/lustre/lustre/ptlrpc/nrs.c

drivers/staging/lustre/lustre/ptlrpc/nrs.c:512 nrs_resource_put_safe()
warn: 'spin_lock:&nrs->nrs_lock' is sometimes locked here and sometimes unlocked.

Would this be happier with something like:

        for (i = 0; i < NRS_RES_MAX; i++) {
		if (pols[i] == NULL)
			continue;
        

		if (nrs == NULL) {
			nrs = pols[i]->pol_nrs;
			if (likely(nrs != NULL)) /* make sparse happy */
				spin_lock(&nrs->nrs_lock);
		}
		nrs_policy_put_locked(pols[i]);
	}

	if (nrs != NULL)
	spin_unlock(&nrs->nrs_lock);
Quoted 4 lines

so that the "if" conditions are the same? The code definitely doesn't have a bug, because the lock is only locked once when nrs is first set, and only unlocked if it is set. Or is there a comment to put there that will quiet the static checker?

Forget about Sparse, it's good at some things and it's fast but it has crappy flow analysis compared to Smatch.

Adding that check does actually silence the warning in Smatch but it's a bad idea. Smatch is supposed to know that "nrs" is non-NULL because of the dereference on the next line. I think this is a recent regression. I'll look into it.

Smatch doesn't have a way to silence false positives. It's still developing, so many of these false positives can be solved by improving the flow analysis.

regards,
dan carpenter

#11 lore raw

Re: [PATCH 4/4] Staging: lustre: sparse lock warning fix

In reply to Dan Carpenter
Quoted 6 lines · Dan Carpenter
Quoted 4 lines · Adrian Remonda

Fixed sparse warning: context imbalance in 'nrs_resource_put_safe' - 'different lock contexts for basic block' by releasing the lock on each iteration of the for loop.

Quoted 10 lines

That changelog doesn't sound correct at all. That's not a correct motivation or explanation.

I reviewed the patch and it's likely going to cause dead locks. The code is trying to take the spinlock for the first pointer in the array and release it at the end. Now it takes the first pointer's spinlock a bunch of times (dead lock) and releases it once (will not happen because we are already dead).

Hello Dan,

thanks for the comments. The code would end up looking as next, I don't undertand where the deadlock would be.
I know the older code would work, I just changed it to keep the lock locked the less time as possible.

        for (i = 0; i < NRS_RES_MAX; i++) {
                if (pols[i] == NULL)
                        continue;

                if (nrs == NULL) {
                        nrs = pols[i]->pol_nrs;
                }
                spin_lock(&nrs->nrs_lock);
                nrs_policy_put_locked(pols[i]);
                spin_unlock(&nrs->nrs_lock);
        }

best regards,
Adrian

#12 lore raw

Re: [PATCH 4/4] Staging: lustre: sparse lock warning fix

Oh, sorry, I didn't read your patch very carefully. It won't cause a deadlock. But I'm going to assume it's still not right until lustre expert Acks it.

How well do you understand lustre locking?

regards,
dan carpenter

#13 lore raw

Re: [PATCH 4/4] Staging: lustre: sparse lock warning fix

Quoted 3 lines · Dan Carpenter

Oh, sorry, I didn't read your patch very carefully. It won't cause a deadlock. But I'm going to assume it's still not right until lustre expert Acks it.

I just took a closer look and it appears original code is buggy and the patch just propagates the bugginess.

If we look at the nrs_policy_put_locked, it eventually ends up in nrs_policy_stop0,
it would hold a lock on whatever happened to be the first policy in the array not NULL.
But nrs_policy_stop0 would unlock the lock on the policy it was called on (already likely a deadlock material) and then relock it.

The problems would arise only if there are more than one nrs policy registered which is theoretically possible, but certainly makes no sense a client (besides, none of the advanced NRS policies made it in anyway and I almost feel like they just add unnecessary complication in client-only code).

The code looks elaborate enough as if the first policy lock is to be always used as the guardian lock, but then stop0 behavior might be a bug then? Or it's possible we never end up in stop0 due to nrs state machine?
Let's see what Nikitas and Liang remember about any of this (one of them is the original author of this code, but I am not sure who.)

Nikitas, Liang: The code in question is in nrs_resource_put_safe:

        for (i = 0; i < NRS_RES_MAX; i++) {
                if (pols[i] == NULL)
                        continue;

                if (nrs == NULL) {
                        nrs = pols[i]->pol_nrs;
                        spin_lock(&nrs->nrs_lock);
                }
                nrs_policy_put_locked(pols[i]);
        }

        if (nrs != NULL)
                spin_unlock(&nrs->nrs_lock);
#14 lore raw

Re: [PATCH 4/4] Staging: lustre: sparse lock warning fix

(Apologies for the double-posting to individual recipients; resending in plain text to get through vger list filters.)

The current code does not have a bug in this path (until we find one, of course). nrs_resource_put_safe() releases references to the related NRS resource hierarchies and policies for an RPC; as Oleg pointed out, the server-side code which is included in the Intel/OpenSFS tree
(http://git.whamcloud.com/fs/lustre-release.git) has more than just the FIFO policy, and so up to NRS_RES_MAX (currently 2) policies can be related with an NRS head (struct ptlrpc_nrs). nrs_resource_put_safe() takes the NRS head spinlock (nrs_lock), releases references to the NRS head's policies, and releases the same spinlock. nrs_policy_stop0() can be entered from nrs_policy_put_locked(), but whichever policy is passed to it, it will release the same spinlock, as all policies for an RPC share the same NRS head, and thus the same nrs_lock.

Liang wrote this part; I believe the aim is to avoid a lock dance of nrs_lock when releasing policy references for the NRS head; the proposed patch would add the lock dance, so I am not sure it would help much; not that it would hurt, but I don't see a good reason to add it, other than to suppress the sparse warning.

Btw, I am guessing that the reason for having a part of both NRS and the server-side of the PTLRPC code in the kernel is only to satisfy build dependencies, and this code is unreachable; i.e. as the TODO file suggests, the client and server sides of PTLRPC will be decoupled, and so this code will eventually be removed from the kernel, likely before transitioning from staging.

Cheers,
Nikitas

Quoted 34 lines · Oleg
Quoted 3 lines · Dan Carpenter

Oh, sorry, I didn't read your patch very carefully. It won't cause a deadlock. But I'm going to assume it's still not right until lustre expert Acks it.

I just took a closer look and it appears original code is buggy and the patch just propagates the bugginess.

If we look at the nrs_policy_put_locked, it eventually ends up in nrs_policy_stop0,
it would hold a lock on whatever happened to be the first policy in the array not NULL.
But nrs_policy_stop0 would unlock the lock on the policy it was called on (already likely a deadlock material) and then relock it.

The problems would arise only if there are more than one nrs policy registered which is theoretically possible, but certainly makes no sense a client (besides, none of the advanced NRS policies made it in anyway and I almost feel like they just add unnecessary complication in client-only code).

The code looks elaborate enough as if the first policy lock is to be always used as the guardian lock, but then stop0 behavior might be a bug then? Or it's possible we never end up in stop0 due to nrs state machine?
Let's see what Nikitas and Liang remember about any of this (one of them is the original author of this code, but I am not sure who.)

Nikitas, Liang: The code in question is in nrs_resource_put_safe:

        for (i = 0; i < NRS_RES_MAX; i++) {
                if (pols[i] == NULL)
                        continue;

                if (nrs == NULL) {
                        nrs = pols[i]->pol_nrs;
                        spin_lock(&nrs->nrs_lock);
                }
                nrs_policy_put_locked(pols[i]);
        }

        if (nrs != NULL)
                spin_unlock(&nrs->nrs_lock);
#15 lore raw
In reply to Adrian Remonda
Quoted 6 lines · Adrian Remonda

This patch clears the warning given by sparse that the function should be static. Done by moving the declaration of the structure 'nrs_conf_fifo' into the header 'ptlrpc_internal.h'

modified: drivers/staging/lustre/lustre/ptlrpc/nrs.c
modified: drivers/staging/lustre/lustre/ptlrpc/ptlrpc_internal.h

Don't include these lines in the changelog.

regards,
dan carpenter