Re: Integrating sysfs memory shrinks

Re: Integrating sysfs memory shrinks

[Adam J. Richter]

On Fri, 7 Jan 2005 10:22:10 +0530, Maneesh Soni wrote:
>On Thu, Jan 06, 2005 at 03:14:11PM -0800, Greg KH wrote:
>> On Wed, Dec 22, 2004 at 11:32:44PM -0800, Adam J. Richter wrote:
>> > On Thu, 23 Dec 2004 12:44:10 +0530, Maneesh Soni wrote:
>> > >On Wed, Dec 22, 2004 at 10:50:17PM -0800, Greg KH wrote:
>> > >> On Wed, Dec 22, 2004 at 06:41:59PM -0800, Adam J. Richter wrote:
>> > >> > Hi guys,
>> > >> > 
>> > >> > 	I'm back from vacation today, and would like to know what needs
>> > >> > to be done to integrate the changes for sysfs memory reduction that
>> > >> > I posted in the past couple of weeks.  Is there someone else I
>> > >> > should be submitting this to or some other mailing list?
>> > >> 
>> > >> Heh, I'm leaving for a few week vacation in a few hours, so I'm not
>> > >> ignoring you :)
>> > >> 
>> > >> Anyway, Maneesh is the one to get approval from on your ideas.
>> > >> 
>> > 
>> > >And I came back today from long vacations, give me a couple of days, by Monday 
>> > >I should be able to reply to your mails.
>> > 
>> > Thanks for the quick replies.
>> > 
>> > Greg, if you're still there, can you clarify whether you're OK with the
>> > sysfs changes getting integrated if Maneesh approves while you're away,
>> > so we don't have to wait a few weeks?
>> 
>> Well, I'm back now.  But which patches should I apply?  Maneesh, care to
>> forward them on, or just point them out to me?
>> 

>The following are ok to go in -mm. Probably Adam have to rediff them as
>code for separate slab for sysfs_dirent is already in -mm.

>  Delete sysfs_dirent.s_count
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110192821619641&w=2

>  Hide sysfs_dirent definition
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110204394818932&w=2

>  un-inline sysfs_get_kobject and release_sysfs_dirent
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110257081016215&w=2

>  Make sysfs_dirent.s_type an unsigned short
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110204704712806&w=2


>The other three patches will need success return from call viro(), IMHO as
>it involves some VFS issues. They also can be part of a set of patches
>implementing un-pinning of directories.

>  Do not allocate sysfs_dirent.s_children for non-directories
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110204491615294&w=2

>  Teach sysfs_get_name not to use a dentry
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110198678514492&w=2

>  unpin sysfs directories, saving ~0.5MB
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110257739705683&w=2
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110257894222719&w=2


>Thanks
>Maneesh


>-- 
>Maneesh Soni
>Linux Technology Center, 
>IBM India Software Labs,
>Bangalore, India
>email: maneesh@in.ibm.com
>Phone: 91-80-25044990

>From maneesh@in.ibm.com  Thu Jan  6 20:38:18 2005
>Date: Fri, 7 Jan 2005 10:22:10 +0530
>From: Maneesh Soni <maneesh@in.ibm.com>
>To: Greg KH <greg@kroah.com>
>Cc: "Adam J. Richter" <adam@yggdrasil.com>
>Subject: Re: Integrating sysfs memory shrinks
>Reply-To: maneesh@in.ibm.com
>References: <200412230732.iBN7Wib21211@adam.yggdrasil.com> <20050106231411.GB22174@kroah.com>
>Mime-Version: 1.0
>Content-Type: text/plain; charset=us-ascii
>Content-Disposition: inline
>In-Reply-To: <20050106231411.GB22174@kroah.com>
>User-Agent: Mutt/1.4.1i

>On Thu, Jan 06, 2005 at 03:14:11PM -0800, Greg KH wrote:
>> On Wed, Dec 22, 2004 at 11:32:44PM -0800, Adam J. Richter wrote:
>> > On Thu, 23 Dec 2004 12:44:10 +0530, Maneesh Soni wrote:
>> > >On Wed, Dec 22, 2004 at 10:50:17PM -0800, Greg KH wrote:
>> > >> On Wed, Dec 22, 2004 at 06:41:59PM -0800, Adam J. Richter wrote:
>> > >> > Hi guys,
>> > >> > 
>> > >> > 	I'm back from vacation today, and would like to know what needs
>> > >> > to be done to integrate the changes for sysfs memory reduction that
>> > >> > I posted in the past couple of weeks.  Is there someone else I
>> > >> > should be submitting this to or some other mailing list?
>> > >> 
>> > >> Heh, I'm leaving for a few week vacation in a few hours, so I'm not
>> > >> ignoring you :)
>> > >> 
>> > >> Anyway, Maneesh is the one to get approval from on your ideas.
>> > >> 
>> > 
>> > >And I came back today from long vacations, give me a couple of days, by Monday 
>> > >I should be able to reply to your mails.
>> > 
>> > Thanks for the quick replies.
>> > 
>> > Greg, if you're still there, can you clarify whether you're OK with the
>> > sysfs changes getting integrated if Maneesh approves while you're away,
>> > so we don't have to wait a few weeks?
>> 
>> Well, I'm back now.  But which patches should I apply?  Maneesh, care to
>> forward them on, or just point them out to me?
>> 

>The following are ok to go in -mm. Probably Adam have to rediff them as
>code for separate slab for sysfs_dirent is already in -mm.

	Can we get the "separate slab for sysfs_dirent" patch
into the stock tree if that is all that it would take for the patches
to apply to both trees?  It's a pretty conservative change (and
makes it easier to observe parts of sysfs memory usage).

>  Delete sysfs_dirent.s_count
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110192821619641&w=2

>  Hide sysfs_dirent definition
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110204394818932&w=2

>  un-inline sysfs_get_kobject and release_sysfs_dirent
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110257081016215&w=2

>  Make sysfs_dirent.s_type an unsigned short
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110204704712806&w=2

	I will be glad to see any progress toward integration, of
course, but I think a slightly faster approach for three of
the above patches is warranted.  I suggest that "hide sysfs_dirent
definition" (a patch which should generate _exactly_ the same object
code), "un-inline sysfs_get_kobject and release_sysfs_dirent" and
"Make sysfs_dirent.s_type an unsigned short" are each trivial enough
that they should go straight to mainline (and mm) for the sake of
simplifying diffs between mm and mainline for people who might want
to apply the subsequent patches to their own trees and also to avoid
gratuitous differences between mm and mainline when people are
tracking down real mm problems.


>The other three patches will need success return from call viro(), IMHO as
>it involves some VFS issues. They also can be part of a set of patches
>implementing un-pinning of directories.

>  Do not allocate sysfs_dirent.s_children for non-directories
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110204491615294&w=2

>  Teach sysfs_get_name not to use a dentry
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110198678514492&w=2

>  unpin sysfs directories, saving ~0.5MB
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110257739705683&w=2
>  http://marc.theaimsgroup.com/?l=linux-kernel&m=110257894222719&w=2


>Thanks
>Maneesh

	I cc'ed Al Viro on those messages and I haven't seen any further
comment from him since his message before this to the effect that
he doubted that I could come up with a workable locking scheme.
I would think that if he found a problem he would have said something
by now.  I'm cc'ing Al and linux-fs-devel in case anyone thinks they
see a problem or has futher questions.  At some point, silence should
be considered equivalent to a "success return."

	I suggest the following course of action:

	1a. Get the "separate slab for sysfs_dirent" into the main
-bk releases to reduce or eliminate differences between -mm and main
sysfs patches.

	1b,1c,1d (1a-d can be done in any order).  Apply the trivial
patches for "hide sysfs_dirent defintion", "un-inline sysfs_get_kobject"
and release_sysfs_dirent" and "make sysfs_dirent.s_type an unsigned short."
To both mm and mainline.  (Please let me know if you find that new
diffs are necessary for any of these.)

	2. After a bk or mm snapshot with thoese patches applied,
we can iterate through the other patches, starting, I suppose, with
"eliminate sysfs_dirent.s_count."  I'll make new diffs if that
will help, and I would prefer if the mm and mainline trees
for fs/sysfs at this start of this step matched so that more
people could more easily try the patches for themselves.  At
this phase, if nobody has identified any problems with the unpinning
of sysfs directories by that time, then I think we should proceed
with integration, especially for -mm.  I've been running these patches
for weeks now without problems.

	Finally, if any of 1a,1b,1c or 1d are agreeable, please
feel free to just do them without needing to settle on the rest
of the game plan.  Even if none of the other patches were ever
integrated in the future, the 1a-1d patches are improvements
by themselves and are quite simple.

                    __     ______________
Adam J. Richter        \ /
adam@yggdrasil.com      | g g d r a s i l

Re: Integrating sysfs memory shrinks

[Adam J. Richter]

	Sorry for including two copies of Maneesh's message in my
reply.  In case anybody missed it, I did reply in and after the
second copy of Maneesh's message.

                    __     ______________
Adam J. Richter        \ /
adam@yggdrasil.com      | g g d r a s i l