| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <f1164ae9-ebcf-41f0-8395-224cdb0f249d@default>
Date: Tue, 10 Dec 2019 10:56:18 -0800 (PST)
From: Alex Kogan <alex.kogan@...cle.com>
To: <longman@...hat.com>
Cc: <rahul.x.yadav@...cle.com>, <tglx@...utronix.de>,
<linux@...linux.org.uk>, <hpa@...or.com>, <dave.dice@...cle.com>,
<mingo@...hat.com>, <will.deacon@....com>, <arnd@...db.de>,
<jglauber@...vell.com>, <guohanjun@...wei.com>, <x86@...nel.org>,
<daniel.m.jordan@...cle.com>, <steven.sistare@...cle.com>,
<bp@...en8.de>, <linux-arm-kernel@...ts.infradead.org>,
<linux-kernel@...r.kernel.org>, <peterz@...radead.org>,
<linux-arch@...r.kernel.org>
Subject: Re: [PATCH v7 5/5] locking/qspinlock: Introduce the shuffle reduction
optimization into CNA
----- longman@...hat.com wrote:
> On 11/25/19 4:07 PM, Alex Kogan wrote:
> > @@ -234,12 +263,13 @@ __always_inline u32 cna_pre_scan(struct
> qspinlock *lock,
> > struct cna_node *cn = (struct cna_node *)node;
> >
> > /*
> > - * setting @pre_scan_result to 1 indicates that no post-scan
> > + * setting @pre_scan_result to 1 or 2 indicates that no post-scan
> > * should be made in cna_pass_lock()
> > */
> > cn->pre_scan_result =
> > - cn->intra_count == intra_node_handoff_threshold ?
> > - 1 : cna_scan_main_queue(node, node);
> > + (node->locked <= 1 && probably(SHUFFLE_REDUCTION_PROB_ARG)) ?
> > + 1 : cn->intra_count == intra_node_handoff_threshold ?
> > + 2 : cna_scan_main_queue(node, node);
> >
> > return 0;
> > }
> > @@ -253,12 +283,15 @@ static inline void cna_pass_lock(struct
> mcs_spinlock *node,
> >
> > u32 scan = cn->pre_scan_result;
> >
> > + if (scan == 1)
> > + goto pass_lock;
> > +
> > /*
> > * check if a successor from the same numa node has not been found
> in
> > * pre-scan, and if so, try to find it in post-scan starting from
> the
> > * node where pre-scan stopped (stored in @pre_scan_result)
> > */
> > - if (scan > 1)
> > + if (scan > 2)
> > scan = cna_scan_main_queue(node, decode_tail(scan));
> >
> > if (!scan) { /* if found a successor from the same numa node */
> > @@ -281,5 +314,6 @@ static inline void cna_pass_lock(struct
> mcs_spinlock *node,
> > tail_2nd->next = next;
> > }
> >
> > +pass_lock:
> > arch_mcs_pass_lock(&next_holder->locked, val);
> > }
>
> I think you might have mishandled the proper accounting of
> intra_count.
> How about something like:
>
> diff --git a/kernel/locking/qspinlock_cna.h
> b/kernel/locking/qspinlock_cna.h
> index f1eef6bece7b..03f8fdec2b80 100644
> --- a/kernel/locking/qspinlock_cna.h
> +++ b/kernel/locking/qspinlock_cna.h
> @@ -268,7 +268,7 @@ __always_inline u32 cna_pre_scan(struct qspinlock
> *lock,
> */
> cn->pre_scan_result =
> (node->locked <= 1 &&
> probably(SHUFFLE_REDUCTION_PROB_ARG)) ?
> - 1 : cn->intra_count ==
> intra_node_handoff_threshold ?
> + 1 : cn->intra_count >=
> intra_node_handoff_threshold ?
We reset ‘intra_count' in cna_init_node(), which is called before we enter
the slow path, and set it at most once when we pass the internal (CNA) lock
by taking the owner’s value + 1. Only after we get the internal lock, we
call this cna_pre_scan() function, where we check the threshold.
IOW, having 'intra_count > intra_node_handoff_threshold' would mean a bug,
and having “>=“ would mask it.
Perhaps I can add WARN_ON(cn->intra_count > intra_node_handoff_threshold)
here instead, although I'm not sure if that is a good idea performance-wise.
> 2 : cna_scan_main_queue(node, node);
>
> return 0;
> @@ -283,9 +283,6 @@ static inline void cna_pass_lock(struct
> mcs_spinlock
> *node,
>
> u32 scan = cn->pre_scan_result;
>
> - if (scan == 1)
> - goto pass_lock;
> -
The thing is that we want to avoid as much of the shuffling-related overhead
as we can when the spinlock is only lightly contended. That's why we have this
early exit here that avoids the rest of the logic of triaging through possible
'scan' values.
> /*
> * check if a successor from the same numa node has not been
> found in
> * pre-scan, and if so, try to find it in post-scan starting
> from the
> @@ -294,7 +291,13 @@ static inline void cna_pass_lock(struct
> mcs_spinlock *node,
> if (scan > 2)
> scan = cna_scan_main_queue(node, decode_tail(scan));
>
> - if (!scan) { /* if found a successor from the same numa node
> */
> + if (scan <= 1) { /* if found a successor from the same numa
> node */
> + /* inc @intra_count if the secondary queue is not
> empty */
> + ((struct cna_node *)next_holder)->intra_count =
> + cn->intra_count + (node->locked > 1);
> + if ((scan == 1)
> + goto pass_lock;
> +
Hmm, I am not sure this makes the code any better/more readable,
while this does add the overhead of going through 3 branches before
jumping to 'pass_lock'.
> next_holder = node->next;
> /*
> * we unlock successor by passing a non-zero value,
> @@ -302,9 +305,6 @@ static inline void cna_pass_lock(struct
> mcs_spinlock
> *node,
> * if we acquired the MCS lock when its queue was
> empty
> */
> val = node->locked ? node->locked : 1;
> - /* inc @intra_count if the secondary queue is not
> empty */
> - ((struct cna_node *)next_holder)->intra_count =
> - cn->intra_count + (node->locked > 1);
> } else if (node->locked > 1) { /* if secondary queue is
> not
> empty */
> /* next holder will be the first node in the
> secondary
> queue */
> tail_2nd = decode_tail(node->locked);
>
> The meaning of scan value:
>
> 0 - pass to next cna node, which is in the same numa node. Additional
> cna node may or may not be added to the secondary queue
>
> 1 - pass to next cna node, which may not be in the same numa node. No
> change to secondary queue
>
> 2 - exceed intra node handoff threshold, unconditionally merge back
> the
> secondary queue cna nodes, if available
>
> >2 no cna node of the same numa node found, unconditionally merge
> back
> the secondary queue cna nodes, if available
'scan' passes information from pre_scan to pass_lock.
The way I see its values is similar, but slightly different:
1 - pass to next cna node, which may not be in the same numa node. No
change to secondary queue.
2 - exceed intra node handoff threshold, unconditionally merge back
the secondary queue cna nodes, if available.
0 - pass to next cna node, which is in the same numa node. pre_scan found
that node, and no further changes to the secondary queue are necessary.
>2 pre_scan could not find cna node in the same numa node. Scan the main
queue from the point where pre_scan stopped, and pass the lock according
to the result of this scan.
>
> The code will be easier to read if symbolic names instead of just
> numbers.
I agree with that. I guess the challenge would be to find short enough symbols
that would convey the meaning of various values. I will think about that.
Best regards,
-- Alex
Powered by blists - more mailing lists