Message-ID: <20121029191659.GB21864@Krystal>
Date:	Mon, 29 Oct 2012 15:16:59 -0400
From:	Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
To:	Sasha Levin <levinsasha928@...il.com>
Cc:	Josh Triplett <josh@...htriplett.org>,
	torvalds@...ux-foundation.org, tj@...nel.org,
	akpm@...ux-foundation.org, linux-kernel@...r.kernel.org,
	linux-mm@...ck.org, paul.gortmaker@...driver.com,
	davem@...emloft.net, rostedt@...dmis.org, mingo@...e.hu,
	ebiederm@...ssion.com, aarcange@...hat.com, ericvh@...il.com,
	netdev@...r.kernel.org, eric.dumazet@...il.com, axboe@...nel.dk,
	agk@...hat.com, dm-devel@...hat.com, neilb@...e.de,
	ccaulfie@...hat.com, teigland@...hat.com,
	Trond.Myklebust@...app.com, bfields@...ldses.org,
	fweisbec@...il.com, jesse@...ira.com,
	venkat.x.venkatsubra@...cle.com, ejt@...hat.com,
	snitzer@...hat.com, edumazet@...gle.com, linux-nfs@...r.kernel.org,
	dev@...nvswitch.org, rds-devel@....oracle.com, lw@...fujitsu.com
Subject: Re: [PATCH v7 06/16] tracepoint: use new hashtable implementation

* Sasha Levin (levinsasha928@...il.com) wrote:
> On Mon, Oct 29, 2012 at 2:53 PM, Mathieu Desnoyers
> <mathieu.desnoyers@...icios.com> wrote:
> > * Sasha Levin (levinsasha928@...il.com) wrote:
> >> On Mon, Oct 29, 2012 at 2:31 PM, Josh Triplett <josh@...htriplett.org> wrote:
> >> > On Mon, Oct 29, 2012 at 01:29:24PM -0400, Sasha Levin wrote:
> >> >> On Mon, Oct 29, 2012 at 7:35 AM, Mathieu Desnoyers
> >> >> <mathieu.desnoyers@...icios.com> wrote:
> >> >> > * Sasha Levin (levinsasha928@...il.com) wrote:
> >> >> >> Switch tracepoints to use the new hashtable implementation. This reduces the amount of
> >> >> >> generic unrelated code in the tracepoints.
> >> >> >>
> >> >> >> Signed-off-by: Sasha Levin <levinsasha928@...il.com>
> >> >> >> ---
> >> >> >>  kernel/tracepoint.c | 27 +++++++++++----------------
> >> >> >>  1 file changed, 11 insertions(+), 16 deletions(-)
> >> >> >>
> >> >> >> diff --git a/kernel/tracepoint.c b/kernel/tracepoint.c
> >> >> >> index d96ba22..854df92 100644
> >> >> >> --- a/kernel/tracepoint.c
> >> >> >> +++ b/kernel/tracepoint.c
> >> >> >> @@ -26,6 +26,7 @@
> >> >> >>  #include <linux/slab.h>
> >> >> >>  #include <linux/sched.h>
> >> >> >>  #include <linux/static_key.h>
> >> >> >> +#include <linux/hashtable.h>
> >> >> >>
> >> >> >>  extern struct tracepoint * const __start___tracepoints_ptrs[];
> >> >> >>  extern struct tracepoint * const __stop___tracepoints_ptrs[];
> >> >> >> @@ -49,8 +50,7 @@ static LIST_HEAD(tracepoint_module_list);
> >> >> >>   * Protected by tracepoints_mutex.
> >> >> >>   */
> >> >> >>  #define TRACEPOINT_HASH_BITS 6
> >> >> >> -#define TRACEPOINT_TABLE_SIZE (1 << TRACEPOINT_HASH_BITS)
> >> >> >> -static struct hlist_head tracepoint_table[TRACEPOINT_TABLE_SIZE];
> >> >> >> +static DEFINE_HASHTABLE(tracepoint_table, TRACEPOINT_HASH_BITS);
> >> >> >>
> >> >> > [...]
> >> >> >>
> >> >> >> @@ -722,6 +715,8 @@ struct notifier_block tracepoint_module_nb = {
> >> >> >>
> >> >> >>  static int init_tracepoints(void)
> >> >> >>  {
> >> >> >> +     hash_init(tracepoint_table);
> >> >> >> +
> >> >> >>       return register_module_notifier(&tracepoint_module_nb);
> >> >> >>  }
> >> >> >>  __initcall(init_tracepoints);
> >> >> >
> >> >> > So we have a hash table defined in .bss (therefore entirely initialized
> >> >> > to NULL), and you add a call to "hash_init", which iterates on the whole
> >> >> > array and initialize it to NULL (again) ?
> >> >> >
> >> >> > This extra initialization is redundant. I think it should be removed
> >> >> > from here, and hashtable.h should document that hash_init() don't need
> >> >> > to be called on zeroed memory (which includes static/global variables,
> >> >> > kzalloc'd memory, etc).
> >> >>
> >> >> This was discussed in the previous series, the conclusion was to call
> >> >> hash_init() either way to keep the encapsulation and consistency.
> >> >>
> >> >> It's cheap enough and happens only once, so why not?
> >> >
> >> > Unnecessary work adds up.  Better not to do it unnecessarily, even if by
> >> > itself it doesn't cost that much.
> >> >
> >> > It doesn't seem that difficult for future fields to have 0 as their
> >> > initialized state.
> >>
> >> Let's put it this way: hlist requires the user to initialize hlist
> >> head before usage, therefore as a hlist user, hashtable implementation
> >> must do that.
> >>
> >> We do it automatically when the hashtable user does
> >> DEFINE_HASHTABLE(), but we can't do that if he does
> >> DECLARE_HASHTABLE(). This means that the hashtable user must call
> >> hash_init() whenever he uses DECLARE_HASHTABLE() to create his
> >> hashtable.
> >>
> >> There are two options here, either we specify that hash_init() should
> >> only be called if DECLARE_HASHTABLE() was called, which is confusing,
> >> inconsistent and prone to errors, or we can just say that it should be
> >> called whenever a hashtable is used.
> >>
> >> The only way to work around it IMO is to get hlist to not require
> >> initializing before usage, and there are good reasons that that won't
> >> happen.
> >
> > Hrm, just a second here.
> >
> > The argument about hash_init being useful to add magic values in the
> > future only works for the cases where a hash table is declared with
> > DECLARE_HASHTABLE(). It's completely pointless with DEFINE_HASHTABLE(),
> > because we could initialize any debugging variables from within
> > DEFINE_HASHTABLE().
> >
> > So I take my "Agreed" back. I disagree with initializing the hash table
> > twice redundantly. There should be at least "DEFINE_HASHTABLE()" or a
> > hash_init() (for DECLARE_HASHTABLE()), but not useless execution
> > initialization on top of an already statically initialized hash table.
> 
> The "magic values" argument was used to point out that some sort of
> initialization *must* occur, either by hash_init() or by a proper
> initialization in DEFINE_HASHTABLE(), and we can't simply memset() it
> to 0. It appears that we all agree on that.

Yes.

> The other thing is whether hash_init() should be called for hashtables
> that were created with DEFINE_HASHTABLE(). That point was raised by
> Neil Brown last time this series went around, and it seems that no one
> objected to the point that it should be consistent across the code.

I was probably busy in the San Diego area at that time, or preparing for
it, sorry! :)

> 
> Even if we ignore hash_init() being mostly optimized out, is it really
> worth it taking the risk that some future patch would move a hashtable
> that user DEFINE_HASHTABLE() into a struct and will start using
> DECLARE_HASHTABLE() and forgetting to initialize it, for example?

There is a saying that with "if"s, we could put Paris in a bottle. ;)

Please have a look at "linux/wait.h", where if a wait queue is defined
with DEFINE_*(), there is just no need to initialize it at runtime.
There are plenty other kernel headers that do the same. I don't see why
hashtable.h should be different.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists