Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
/*-
|
2017-11-27 15:20:12 +00:00
|
|
|
* SPDX-License-Identifier: BSD-2-Clause-FreeBSD
|
|
|
|
|
*
|
2005-02-16 21:45:59 +00:00
|
|
|
* Copyright (c) 2004, 2005,
|
2012-12-07 22:19:41 +00:00
|
|
|
* Bosko Milekic <bmilekic@FreeBSD.org>. All rights reserved.
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
*
|
|
|
|
|
* Redistribution and use in source and binary forms, with or without
|
|
|
|
|
* modification, are permitted provided that the following conditions
|
|
|
|
|
* are met:
|
|
|
|
|
* 1. Redistributions of source code must retain the above copyright
|
|
|
|
|
* notice unmodified, this list of conditions and the following
|
|
|
|
|
* disclaimer.
|
|
|
|
|
* 2. Redistributions in binary form must reproduce the above copyright
|
|
|
|
|
* notice, this list of conditions and the following disclaimer in the
|
|
|
|
|
* documentation and/or other materials provided with the distribution.
|
|
|
|
|
*
|
|
|
|
|
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
|
|
|
|
|
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
|
|
|
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
|
|
|
|
|
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
|
|
|
|
|
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
|
|
|
|
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
|
|
|
|
|
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
|
|
|
|
|
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
|
|
|
|
|
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
|
|
|
|
|
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
|
|
|
|
|
* SUCH DAMAGE.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
#include <sys/cdefs.h>
|
|
|
|
|
__FBSDID("$FreeBSD$");
|
|
|
|
|
|
|
|
|
|
#include "opt_param.h"
|
|
|
|
|
|
|
|
|
|
#include <sys/param.h>
|
2018-10-20 22:24:58 +00:00
|
|
|
#include <sys/conf.h>
|
2018-10-30 18:26:34 +00:00
|
|
|
#include <sys/domainset.h>
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
#include <sys/malloc.h>
|
|
|
|
|
#include <sys/systm.h>
|
|
|
|
|
#include <sys/mbuf.h>
|
|
|
|
|
#include <sys/domain.h>
|
|
|
|
|
#include <sys/eventhandler.h>
|
|
|
|
|
#include <sys/kernel.h>
|
2018-05-06 00:19:48 +00:00
|
|
|
#include <sys/limits.h>
|
2015-12-20 02:05:33 +00:00
|
|
|
#include <sys/lock.h>
|
|
|
|
|
#include <sys/mutex.h>
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
#include <sys/protosw.h>
|
|
|
|
|
#include <sys/smp.h>
|
|
|
|
|
#include <sys/sysctl.h>
|
|
|
|
|
|
|
|
|
|
#include <vm/vm.h>
|
2009-06-18 17:59:04 +00:00
|
|
|
#include <vm/vm_extern.h>
|
|
|
|
|
#include <vm/vm_kern.h>
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
#include <vm/vm_page.h>
|
2013-01-17 21:28:31 +00:00
|
|
|
#include <vm/vm_map.h>
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
#include <vm/uma.h>
|
2005-06-23 04:33:39 +00:00
|
|
|
#include <vm/uma_dbg.h>
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
|
|
|
|
|
* Zones.
|
|
|
|
|
*
|
|
|
|
|
* Mbuf Clusters (2K, contiguous) are allocated from the Cluster
|
|
|
|
|
* Zone. The Zone can be capped at kern.ipc.nmbclusters, if the
|
|
|
|
|
* administrator so desires.
|
|
|
|
|
*
|
|
|
|
|
* Mbufs are allocated from a UMA Master Zone called the Mbuf
|
|
|
|
|
* Zone.
|
|
|
|
|
*
|
|
|
|
|
* Additionally, FreeBSD provides a Packet Zone, which it
|
|
|
|
|
* configures as a Secondary Zone to the Mbuf Master Zone,
|
|
|
|
|
* thus sharing backend Slab kegs with the Mbuf Master Zone.
|
|
|
|
|
*
|
|
|
|
|
* Thus common-case allocations and locking are simplified:
|
|
|
|
|
*
|
|
|
|
|
* m_clget() m_getcl()
|
|
|
|
|
* | |
|
|
|
|
|
* | .------------>[(Packet Cache)] m_get(), m_gethdr()
|
|
|
|
|
* | | [ Packet ] |
|
|
|
|
|
* [(Cluster Cache)] [ Secondary ] [ (Mbuf Cache) ]
|
|
|
|
|
* [ Cluster Zone ] [ Zone ] [ Mbuf Master Zone ]
|
|
|
|
|
* | \________ |
|
|
|
|
|
* [ Cluster Keg ] \ /
|
2012-12-07 22:19:41 +00:00
|
|
|
* | [ Mbuf Keg ]
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
* [ Cluster Slabs ] |
|
|
|
|
|
* | [ Mbuf Slabs ]
|
|
|
|
|
* \____________(VM)_________________/
|
2005-11-02 16:20:36 +00:00
|
|
|
*
|
|
|
|
|
*
|
2006-02-26 11:44:28 +00:00
|
|
|
* Whenever an object is allocated with uma_zalloc() out of
|
2005-11-02 16:20:36 +00:00
|
|
|
* one of the Zones its _ctor_ function is executed. The same
|
2006-02-26 11:44:28 +00:00
|
|
|
* for any deallocation through uma_zfree() the _dtor_ function
|
2005-11-02 16:20:36 +00:00
|
|
|
* is executed.
|
2006-02-26 11:44:28 +00:00
|
|
|
*
|
2005-11-02 16:20:36 +00:00
|
|
|
* Caches are per-CPU and are filled from the Master Zone.
|
|
|
|
|
*
|
2006-02-26 11:44:28 +00:00
|
|
|
* Whenever an object is allocated from the underlying global
|
2005-11-02 16:20:36 +00:00
|
|
|
* memory pool it gets pre-initialized with the _zinit_ functions.
|
2016-04-29 22:15:33 +00:00
|
|
|
* When the Keg's are overfull objects get decommissioned with
|
2005-11-02 16:20:36 +00:00
|
|
|
* _zfini_ functions and free'd back to the global memory pool.
|
|
|
|
|
*
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
*/
|
|
|
|
|
|
Base the mbuf related limits on the available physical memory or
kernel memory, whichever is lower. The overall mbuf related memory
limit must be set so that mbufs (and clusters of various sizes)
can't exhaust physical RAM or KVM.
The limit is set to half of the physical RAM or KVM (whichever is
lower) as the baseline. In any normal scenario we want to leave
at least half of the physmem/kvm for other kernel functions and
userspace to prevent it from swapping too easily. Via a tunable
kern.maxmbufmem the limit can be upped to at most 3/4 of physmem/kvm.
At the same time divorce maxfiles from maxusers and set maxfiles to
physpages / 8 with a floor based on maxusers. This way busy servers
can make use of the significantly increased mbuf limits with a much
larger number of open sockets.
Tidy up ordering in init_param2() and check up on some users of
those values calculated here.
Out of the overall mbuf memory limit 2K clusters and 4K (page size)
clusters to get 1/4 each because these are the most heavily used mbuf
sizes. 2K clusters are used for MTU 1500 ethernet inbound packets.
4K clusters are used whenever possible for sends on sockets and thus
outbound packets. The larger cluster sizes of 9K and 16K are limited
to 1/6 of the overall mbuf memory limit. When jumbo MTU's are used
these large clusters will end up only on the inbound path. They are
not used on outbound, there it's still 4K. Yes, that will stay that
way because otherwise we run into lots of complications in the
stack. And it really isn't a problem, so don't make a scene.
Normal mbufs (256B) weren't limited at all previously. This was
problematic as there are certain places in the kernel that on
allocation failure of clusters try to piece together their packet
from smaller mbufs.
The mbuf limit is the number of all other mbuf sizes together plus
some more to allow for standalone mbufs (ACK for example) and to
send off a copy of a cluster. Unfortunately there isn't a way to
set an overall limit for all mbuf memory together as UMA doesn't
support such a limiting.
NB: Every cluster also has an mbuf associated with it.
Two examples on the revised mbuf sizing limits:
1GB KVM:
512MB limit for mbufs
419,430 mbufs
65,536 2K mbuf clusters
32,768 4K mbuf clusters
9,709 9K mbuf clusters
5,461 16K mbuf clusters
16GB RAM:
8GB limit for mbufs
33,554,432 mbufs
1,048,576 2K mbuf clusters
524,288 4K mbuf clusters
155,344 9K mbuf clusters
87,381 16K mbuf clusters
These defaults should be sufficient for even the most demanding
network loads.
MFC after: 1 month
2012-11-27 21:19:58 +00:00
|
|
|
int nmbufs; /* limits number of mbufs */
|
2005-11-02 16:20:36 +00:00
|
|
|
int nmbclusters; /* limits number of mbuf clusters */
|
2006-02-17 14:14:15 +00:00
|
|
|
int nmbjumbop; /* limits number of page size jumbo clusters */
|
2005-11-02 16:20:36 +00:00
|
|
|
int nmbjumbo9; /* limits number of 9k jumbo clusters */
|
|
|
|
|
int nmbjumbo16; /* limits number of 16k jumbo clusters */
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
|
2013-07-11 12:46:35 +00:00
|
|
|
static quad_t maxmbufmem; /* overall real memory limit for all mbufs */
|
|
|
|
|
|
2014-06-28 03:56:17 +00:00
|
|
|
SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
|
2013-09-05 18:18:23 +00:00
|
|
|
"Maximum real memory allocatable to various mbuf types");
|
2013-07-11 12:46:35 +00:00
|
|
|
|
2008-12-10 22:17:09 +00:00
|
|
|
/*
|
2013-01-17 21:28:31 +00:00
|
|
|
* tunable_mbinit() has to be run before any mbuf allocations are done.
|
2008-12-10 22:17:09 +00:00
|
|
|
*/
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
static void
|
|
|
|
|
tunable_mbinit(void *dummy)
|
|
|
|
|
{
|
2013-07-11 12:46:35 +00:00
|
|
|
quad_t realmem;
|
2013-01-17 21:28:31 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The default limit for all mbuf related memory is 1/2 of all
|
|
|
|
|
* available kernel memory (physical or kmem).
|
|
|
|
|
* At most it can be 3/4 of available kernel memory.
|
|
|
|
|
*/
|
2013-08-07 06:21:20 +00:00
|
|
|
realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
|
2013-01-17 21:28:31 +00:00
|
|
|
maxmbufmem = realmem / 2;
|
2013-07-11 12:46:35 +00:00
|
|
|
TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
|
2013-01-17 21:28:31 +00:00
|
|
|
if (maxmbufmem > realmem / 4 * 3)
|
|
|
|
|
maxmbufmem = realmem / 4 * 3;
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
|
2012-08-23 21:32:02 +00:00
|
|
|
TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
|
2012-11-27 23:16:56 +00:00
|
|
|
if (nmbclusters == 0)
|
|
|
|
|
nmbclusters = maxmbufmem / MCLBYTES / 4;
|
2012-08-23 21:32:02 +00:00
|
|
|
|
|
|
|
|
TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
|
|
|
|
|
if (nmbjumbop == 0)
|
2012-11-27 23:16:56 +00:00
|
|
|
nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
|
2012-08-23 21:32:02 +00:00
|
|
|
|
|
|
|
|
TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
|
|
|
|
|
if (nmbjumbo9 == 0)
|
2012-11-27 23:16:56 +00:00
|
|
|
nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
|
2012-08-23 21:32:02 +00:00
|
|
|
|
|
|
|
|
TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
|
|
|
|
|
if (nmbjumbo16 == 0)
|
2012-11-27 23:16:56 +00:00
|
|
|
nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We need at least as many mbufs as we have clusters of
|
|
|
|
|
* the various types added together.
|
|
|
|
|
*/
|
|
|
|
|
TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
|
|
|
|
|
if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
|
|
|
|
|
nmbufs = lmax(maxmbufmem / MSIZE / 5,
|
2012-12-07 22:22:04 +00:00
|
|
|
nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
}
|
2013-01-17 21:28:31 +00:00
|
|
|
SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
|
2006-04-21 09:25:40 +00:00
|
|
|
static int
|
|
|
|
|
sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
|
|
|
|
|
{
|
|
|
|
|
int error, newnmbclusters;
|
|
|
|
|
|
|
|
|
|
newnmbclusters = nmbclusters;
|
2012-12-07 22:19:41 +00:00
|
|
|
error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
|
2013-10-10 16:11:34 +00:00
|
|
|
if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
|
Base the mbuf related limits on the available physical memory or
kernel memory, whichever is lower. The overall mbuf related memory
limit must be set so that mbufs (and clusters of various sizes)
can't exhaust physical RAM or KVM.
The limit is set to half of the physical RAM or KVM (whichever is
lower) as the baseline. In any normal scenario we want to leave
at least half of the physmem/kvm for other kernel functions and
userspace to prevent it from swapping too easily. Via a tunable
kern.maxmbufmem the limit can be upped to at most 3/4 of physmem/kvm.
At the same time divorce maxfiles from maxusers and set maxfiles to
physpages / 8 with a floor based on maxusers. This way busy servers
can make use of the significantly increased mbuf limits with a much
larger number of open sockets.
Tidy up ordering in init_param2() and check up on some users of
those values calculated here.
Out of the overall mbuf memory limit 2K clusters and 4K (page size)
clusters to get 1/4 each because these are the most heavily used mbuf
sizes. 2K clusters are used for MTU 1500 ethernet inbound packets.
4K clusters are used whenever possible for sends on sockets and thus
outbound packets. The larger cluster sizes of 9K and 16K are limited
to 1/6 of the overall mbuf memory limit. When jumbo MTU's are used
these large clusters will end up only on the inbound path. They are
not used on outbound, there it's still 4K. Yes, that will stay that
way because otherwise we run into lots of complications in the
stack. And it really isn't a problem, so don't make a scene.
Normal mbufs (256B) weren't limited at all previously. This was
problematic as there are certain places in the kernel that on
allocation failure of clusters try to piece together their packet
from smaller mbufs.
The mbuf limit is the number of all other mbuf sizes together plus
some more to allow for standalone mbufs (ACK for example) and to
send off a copy of a cluster. Unfortunately there isn't a way to
set an overall limit for all mbuf memory together as UMA doesn't
support such a limiting.
NB: Every cluster also has an mbuf associated with it.
Two examples on the revised mbuf sizing limits:
1GB KVM:
512MB limit for mbufs
419,430 mbufs
65,536 2K mbuf clusters
32,768 4K mbuf clusters
9,709 9K mbuf clusters
5,461 16K mbuf clusters
16GB RAM:
8GB limit for mbufs
33,554,432 mbufs
1,048,576 2K mbuf clusters
524,288 4K mbuf clusters
155,344 9K mbuf clusters
87,381 16K mbuf clusters
These defaults should be sufficient for even the most demanding
network loads.
MFC after: 1 month
2012-11-27 21:19:58 +00:00
|
|
|
if (newnmbclusters > nmbclusters &&
|
|
|
|
|
nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
|
2006-04-21 09:25:40 +00:00
|
|
|
nmbclusters = newnmbclusters;
|
2013-07-11 12:53:13 +00:00
|
|
|
nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
|
2006-04-21 09:25:40 +00:00
|
|
|
EVENTHANDLER_INVOKE(nmbclusters_change);
|
|
|
|
|
} else
|
|
|
|
|
error = EINVAL;
|
|
|
|
|
}
|
|
|
|
|
return (error);
|
|
|
|
|
}
|
|
|
|
|
SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW,
|
|
|
|
|
&nmbclusters, 0, sysctl_nmbclusters, "IU",
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
"Maximum number of mbuf clusters allowed");
|
2007-12-05 15:29:44 +00:00
|
|
|
|
|
|
|
|
static int
|
|
|
|
|
sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
|
|
|
|
|
{
|
|
|
|
|
int error, newnmbjumbop;
|
|
|
|
|
|
|
|
|
|
newnmbjumbop = nmbjumbop;
|
2012-12-07 22:19:41 +00:00
|
|
|
error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
|
2013-10-10 16:11:34 +00:00
|
|
|
if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
|
Base the mbuf related limits on the available physical memory or
kernel memory, whichever is lower. The overall mbuf related memory
limit must be set so that mbufs (and clusters of various sizes)
can't exhaust physical RAM or KVM.
The limit is set to half of the physical RAM or KVM (whichever is
lower) as the baseline. In any normal scenario we want to leave
at least half of the physmem/kvm for other kernel functions and
userspace to prevent it from swapping too easily. Via a tunable
kern.maxmbufmem the limit can be upped to at most 3/4 of physmem/kvm.
At the same time divorce maxfiles from maxusers and set maxfiles to
physpages / 8 with a floor based on maxusers. This way busy servers
can make use of the significantly increased mbuf limits with a much
larger number of open sockets.
Tidy up ordering in init_param2() and check up on some users of
those values calculated here.
Out of the overall mbuf memory limit 2K clusters and 4K (page size)
clusters to get 1/4 each because these are the most heavily used mbuf
sizes. 2K clusters are used for MTU 1500 ethernet inbound packets.
4K clusters are used whenever possible for sends on sockets and thus
outbound packets. The larger cluster sizes of 9K and 16K are limited
to 1/6 of the overall mbuf memory limit. When jumbo MTU's are used
these large clusters will end up only on the inbound path. They are
not used on outbound, there it's still 4K. Yes, that will stay that
way because otherwise we run into lots of complications in the
stack. And it really isn't a problem, so don't make a scene.
Normal mbufs (256B) weren't limited at all previously. This was
problematic as there are certain places in the kernel that on
allocation failure of clusters try to piece together their packet
from smaller mbufs.
The mbuf limit is the number of all other mbuf sizes together plus
some more to allow for standalone mbufs (ACK for example) and to
send off a copy of a cluster. Unfortunately there isn't a way to
set an overall limit for all mbuf memory together as UMA doesn't
support such a limiting.
NB: Every cluster also has an mbuf associated with it.
Two examples on the revised mbuf sizing limits:
1GB KVM:
512MB limit for mbufs
419,430 mbufs
65,536 2K mbuf clusters
32,768 4K mbuf clusters
9,709 9K mbuf clusters
5,461 16K mbuf clusters
16GB RAM:
8GB limit for mbufs
33,554,432 mbufs
1,048,576 2K mbuf clusters
524,288 4K mbuf clusters
155,344 9K mbuf clusters
87,381 16K mbuf clusters
These defaults should be sufficient for even the most demanding
network loads.
MFC after: 1 month
2012-11-27 21:19:58 +00:00
|
|
|
if (newnmbjumbop > nmbjumbop &&
|
|
|
|
|
nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
|
2007-12-05 15:29:44 +00:00
|
|
|
nmbjumbop = newnmbjumbop;
|
2013-07-11 12:53:13 +00:00
|
|
|
nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
|
2007-12-05 15:29:44 +00:00
|
|
|
} else
|
|
|
|
|
error = EINVAL;
|
|
|
|
|
}
|
|
|
|
|
return (error);
|
|
|
|
|
}
|
|
|
|
|
SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop, CTLTYPE_INT|CTLFLAG_RW,
|
|
|
|
|
&nmbjumbop, 0, sysctl_nmbjumbop, "IU",
|
2012-12-07 22:22:04 +00:00
|
|
|
"Maximum number of mbuf page size jumbo clusters allowed");
|
2007-12-05 15:29:44 +00:00
|
|
|
|
|
|
|
|
static int
|
|
|
|
|
sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
|
|
|
|
|
{
|
|
|
|
|
int error, newnmbjumbo9;
|
|
|
|
|
|
|
|
|
|
newnmbjumbo9 = nmbjumbo9;
|
2012-12-07 22:19:41 +00:00
|
|
|
error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
|
2013-10-10 16:11:34 +00:00
|
|
|
if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
|
2013-07-11 12:46:35 +00:00
|
|
|
if (newnmbjumbo9 > nmbjumbo9 &&
|
Base the mbuf related limits on the available physical memory or
kernel memory, whichever is lower. The overall mbuf related memory
limit must be set so that mbufs (and clusters of various sizes)
can't exhaust physical RAM or KVM.
The limit is set to half of the physical RAM or KVM (whichever is
lower) as the baseline. In any normal scenario we want to leave
at least half of the physmem/kvm for other kernel functions and
userspace to prevent it from swapping too easily. Via a tunable
kern.maxmbufmem the limit can be upped to at most 3/4 of physmem/kvm.
At the same time divorce maxfiles from maxusers and set maxfiles to
physpages / 8 with a floor based on maxusers. This way busy servers
can make use of the significantly increased mbuf limits with a much
larger number of open sockets.
Tidy up ordering in init_param2() and check up on some users of
those values calculated here.
Out of the overall mbuf memory limit 2K clusters and 4K (page size)
clusters to get 1/4 each because these are the most heavily used mbuf
sizes. 2K clusters are used for MTU 1500 ethernet inbound packets.
4K clusters are used whenever possible for sends on sockets and thus
outbound packets. The larger cluster sizes of 9K and 16K are limited
to 1/6 of the overall mbuf memory limit. When jumbo MTU's are used
these large clusters will end up only on the inbound path. They are
not used on outbound, there it's still 4K. Yes, that will stay that
way because otherwise we run into lots of complications in the
stack. And it really isn't a problem, so don't make a scene.
Normal mbufs (256B) weren't limited at all previously. This was
problematic as there are certain places in the kernel that on
allocation failure of clusters try to piece together their packet
from smaller mbufs.
The mbuf limit is the number of all other mbuf sizes together plus
some more to allow for standalone mbufs (ACK for example) and to
send off a copy of a cluster. Unfortunately there isn't a way to
set an overall limit for all mbuf memory together as UMA doesn't
support such a limiting.
NB: Every cluster also has an mbuf associated with it.
Two examples on the revised mbuf sizing limits:
1GB KVM:
512MB limit for mbufs
419,430 mbufs
65,536 2K mbuf clusters
32,768 4K mbuf clusters
9,709 9K mbuf clusters
5,461 16K mbuf clusters
16GB RAM:
8GB limit for mbufs
33,554,432 mbufs
1,048,576 2K mbuf clusters
524,288 4K mbuf clusters
155,344 9K mbuf clusters
87,381 16K mbuf clusters
These defaults should be sufficient for even the most demanding
network loads.
MFC after: 1 month
2012-11-27 21:19:58 +00:00
|
|
|
nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
|
2007-12-05 15:29:44 +00:00
|
|
|
nmbjumbo9 = newnmbjumbo9;
|
2013-07-11 12:53:13 +00:00
|
|
|
nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
|
2007-12-05 15:29:44 +00:00
|
|
|
} else
|
|
|
|
|
error = EINVAL;
|
|
|
|
|
}
|
|
|
|
|
return (error);
|
|
|
|
|
}
|
|
|
|
|
SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9, CTLTYPE_INT|CTLFLAG_RW,
|
|
|
|
|
&nmbjumbo9, 0, sysctl_nmbjumbo9, "IU",
|
2012-12-07 22:19:41 +00:00
|
|
|
"Maximum number of mbuf 9k jumbo clusters allowed");
|
2007-12-05 15:29:44 +00:00
|
|
|
|
|
|
|
|
static int
|
|
|
|
|
sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
|
|
|
|
|
{
|
|
|
|
|
int error, newnmbjumbo16;
|
|
|
|
|
|
|
|
|
|
newnmbjumbo16 = nmbjumbo16;
|
2012-12-07 22:19:41 +00:00
|
|
|
error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
|
2013-10-10 16:11:34 +00:00
|
|
|
if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
|
Base the mbuf related limits on the available physical memory or
kernel memory, whichever is lower. The overall mbuf related memory
limit must be set so that mbufs (and clusters of various sizes)
can't exhaust physical RAM or KVM.
The limit is set to half of the physical RAM or KVM (whichever is
lower) as the baseline. In any normal scenario we want to leave
at least half of the physmem/kvm for other kernel functions and
userspace to prevent it from swapping too easily. Via a tunable
kern.maxmbufmem the limit can be upped to at most 3/4 of physmem/kvm.
At the same time divorce maxfiles from maxusers and set maxfiles to
physpages / 8 with a floor based on maxusers. This way busy servers
can make use of the significantly increased mbuf limits with a much
larger number of open sockets.
Tidy up ordering in init_param2() and check up on some users of
those values calculated here.
Out of the overall mbuf memory limit 2K clusters and 4K (page size)
clusters to get 1/4 each because these are the most heavily used mbuf
sizes. 2K clusters are used for MTU 1500 ethernet inbound packets.
4K clusters are used whenever possible for sends on sockets and thus
outbound packets. The larger cluster sizes of 9K and 16K are limited
to 1/6 of the overall mbuf memory limit. When jumbo MTU's are used
these large clusters will end up only on the inbound path. They are
not used on outbound, there it's still 4K. Yes, that will stay that
way because otherwise we run into lots of complications in the
stack. And it really isn't a problem, so don't make a scene.
Normal mbufs (256B) weren't limited at all previously. This was
problematic as there are certain places in the kernel that on
allocation failure of clusters try to piece together their packet
from smaller mbufs.
The mbuf limit is the number of all other mbuf sizes together plus
some more to allow for standalone mbufs (ACK for example) and to
send off a copy of a cluster. Unfortunately there isn't a way to
set an overall limit for all mbuf memory together as UMA doesn't
support such a limiting.
NB: Every cluster also has an mbuf associated with it.
Two examples on the revised mbuf sizing limits:
1GB KVM:
512MB limit for mbufs
419,430 mbufs
65,536 2K mbuf clusters
32,768 4K mbuf clusters
9,709 9K mbuf clusters
5,461 16K mbuf clusters
16GB RAM:
8GB limit for mbufs
33,554,432 mbufs
1,048,576 2K mbuf clusters
524,288 4K mbuf clusters
155,344 9K mbuf clusters
87,381 16K mbuf clusters
These defaults should be sufficient for even the most demanding
network loads.
MFC after: 1 month
2012-11-27 21:19:58 +00:00
|
|
|
if (newnmbjumbo16 > nmbjumbo16 &&
|
|
|
|
|
nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
|
2007-12-05 15:29:44 +00:00
|
|
|
nmbjumbo16 = newnmbjumbo16;
|
2013-07-11 12:53:13 +00:00
|
|
|
nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
|
2007-12-05 15:29:44 +00:00
|
|
|
} else
|
|
|
|
|
error = EINVAL;
|
|
|
|
|
}
|
|
|
|
|
return (error);
|
|
|
|
|
}
|
|
|
|
|
SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16, CTLTYPE_INT|CTLFLAG_RW,
|
|
|
|
|
&nmbjumbo16, 0, sysctl_nmbjumbo16, "IU",
|
2005-11-02 16:20:36 +00:00
|
|
|
"Maximum number of mbuf 16k jumbo clusters allowed");
|
2007-12-05 15:29:44 +00:00
|
|
|
|
Base the mbuf related limits on the available physical memory or
kernel memory, whichever is lower. The overall mbuf related memory
limit must be set so that mbufs (and clusters of various sizes)
can't exhaust physical RAM or KVM.
The limit is set to half of the physical RAM or KVM (whichever is
lower) as the baseline. In any normal scenario we want to leave
at least half of the physmem/kvm for other kernel functions and
userspace to prevent it from swapping too easily. Via a tunable
kern.maxmbufmem the limit can be upped to at most 3/4 of physmem/kvm.
At the same time divorce maxfiles from maxusers and set maxfiles to
physpages / 8 with a floor based on maxusers. This way busy servers
can make use of the significantly increased mbuf limits with a much
larger number of open sockets.
Tidy up ordering in init_param2() and check up on some users of
those values calculated here.
Out of the overall mbuf memory limit 2K clusters and 4K (page size)
clusters to get 1/4 each because these are the most heavily used mbuf
sizes. 2K clusters are used for MTU 1500 ethernet inbound packets.
4K clusters are used whenever possible for sends on sockets and thus
outbound packets. The larger cluster sizes of 9K and 16K are limited
to 1/6 of the overall mbuf memory limit. When jumbo MTU's are used
these large clusters will end up only on the inbound path. They are
not used on outbound, there it's still 4K. Yes, that will stay that
way because otherwise we run into lots of complications in the
stack. And it really isn't a problem, so don't make a scene.
Normal mbufs (256B) weren't limited at all previously. This was
problematic as there are certain places in the kernel that on
allocation failure of clusters try to piece together their packet
from smaller mbufs.
The mbuf limit is the number of all other mbuf sizes together plus
some more to allow for standalone mbufs (ACK for example) and to
send off a copy of a cluster. Unfortunately there isn't a way to
set an overall limit for all mbuf memory together as UMA doesn't
support such a limiting.
NB: Every cluster also has an mbuf associated with it.
Two examples on the revised mbuf sizing limits:
1GB KVM:
512MB limit for mbufs
419,430 mbufs
65,536 2K mbuf clusters
32,768 4K mbuf clusters
9,709 9K mbuf clusters
5,461 16K mbuf clusters
16GB RAM:
8GB limit for mbufs
33,554,432 mbufs
1,048,576 2K mbuf clusters
524,288 4K mbuf clusters
155,344 9K mbuf clusters
87,381 16K mbuf clusters
These defaults should be sufficient for even the most demanding
network loads.
MFC after: 1 month
2012-11-27 21:19:58 +00:00
|
|
|
static int
|
|
|
|
|
sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
|
|
|
|
|
{
|
|
|
|
|
int error, newnmbufs;
|
|
|
|
|
|
|
|
|
|
newnmbufs = nmbufs;
|
2012-12-07 22:19:41 +00:00
|
|
|
error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
|
2013-10-10 16:11:34 +00:00
|
|
|
if (error == 0 && req->newptr && newnmbufs != nmbufs) {
|
Base the mbuf related limits on the available physical memory or
kernel memory, whichever is lower. The overall mbuf related memory
limit must be set so that mbufs (and clusters of various sizes)
can't exhaust physical RAM or KVM.
The limit is set to half of the physical RAM or KVM (whichever is
lower) as the baseline. In any normal scenario we want to leave
at least half of the physmem/kvm for other kernel functions and
userspace to prevent it from swapping too easily. Via a tunable
kern.maxmbufmem the limit can be upped to at most 3/4 of physmem/kvm.
At the same time divorce maxfiles from maxusers and set maxfiles to
physpages / 8 with a floor based on maxusers. This way busy servers
can make use of the significantly increased mbuf limits with a much
larger number of open sockets.
Tidy up ordering in init_param2() and check up on some users of
those values calculated here.
Out of the overall mbuf memory limit 2K clusters and 4K (page size)
clusters to get 1/4 each because these are the most heavily used mbuf
sizes. 2K clusters are used for MTU 1500 ethernet inbound packets.
4K clusters are used whenever possible for sends on sockets and thus
outbound packets. The larger cluster sizes of 9K and 16K are limited
to 1/6 of the overall mbuf memory limit. When jumbo MTU's are used
these large clusters will end up only on the inbound path. They are
not used on outbound, there it's still 4K. Yes, that will stay that
way because otherwise we run into lots of complications in the
stack. And it really isn't a problem, so don't make a scene.
Normal mbufs (256B) weren't limited at all previously. This was
problematic as there are certain places in the kernel that on
allocation failure of clusters try to piece together their packet
from smaller mbufs.
The mbuf limit is the number of all other mbuf sizes together plus
some more to allow for standalone mbufs (ACK for example) and to
send off a copy of a cluster. Unfortunately there isn't a way to
set an overall limit for all mbuf memory together as UMA doesn't
support such a limiting.
NB: Every cluster also has an mbuf associated with it.
Two examples on the revised mbuf sizing limits:
1GB KVM:
512MB limit for mbufs
419,430 mbufs
65,536 2K mbuf clusters
32,768 4K mbuf clusters
9,709 9K mbuf clusters
5,461 16K mbuf clusters
16GB RAM:
8GB limit for mbufs
33,554,432 mbufs
1,048,576 2K mbuf clusters
524,288 4K mbuf clusters
155,344 9K mbuf clusters
87,381 16K mbuf clusters
These defaults should be sufficient for even the most demanding
network loads.
MFC after: 1 month
2012-11-27 21:19:58 +00:00
|
|
|
if (newnmbufs > nmbufs) {
|
|
|
|
|
nmbufs = newnmbufs;
|
2013-07-11 12:53:13 +00:00
|
|
|
nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
|
Base the mbuf related limits on the available physical memory or
kernel memory, whichever is lower. The overall mbuf related memory
limit must be set so that mbufs (and clusters of various sizes)
can't exhaust physical RAM or KVM.
The limit is set to half of the physical RAM or KVM (whichever is
lower) as the baseline. In any normal scenario we want to leave
at least half of the physmem/kvm for other kernel functions and
userspace to prevent it from swapping too easily. Via a tunable
kern.maxmbufmem the limit can be upped to at most 3/4 of physmem/kvm.
At the same time divorce maxfiles from maxusers and set maxfiles to
physpages / 8 with a floor based on maxusers. This way busy servers
can make use of the significantly increased mbuf limits with a much
larger number of open sockets.
Tidy up ordering in init_param2() and check up on some users of
those values calculated here.
Out of the overall mbuf memory limit 2K clusters and 4K (page size)
clusters to get 1/4 each because these are the most heavily used mbuf
sizes. 2K clusters are used for MTU 1500 ethernet inbound packets.
4K clusters are used whenever possible for sends on sockets and thus
outbound packets. The larger cluster sizes of 9K and 16K are limited
to 1/6 of the overall mbuf memory limit. When jumbo MTU's are used
these large clusters will end up only on the inbound path. They are
not used on outbound, there it's still 4K. Yes, that will stay that
way because otherwise we run into lots of complications in the
stack. And it really isn't a problem, so don't make a scene.
Normal mbufs (256B) weren't limited at all previously. This was
problematic as there are certain places in the kernel that on
allocation failure of clusters try to piece together their packet
from smaller mbufs.
The mbuf limit is the number of all other mbuf sizes together plus
some more to allow for standalone mbufs (ACK for example) and to
send off a copy of a cluster. Unfortunately there isn't a way to
set an overall limit for all mbuf memory together as UMA doesn't
support such a limiting.
NB: Every cluster also has an mbuf associated with it.
Two examples on the revised mbuf sizing limits:
1GB KVM:
512MB limit for mbufs
419,430 mbufs
65,536 2K mbuf clusters
32,768 4K mbuf clusters
9,709 9K mbuf clusters
5,461 16K mbuf clusters
16GB RAM:
8GB limit for mbufs
33,554,432 mbufs
1,048,576 2K mbuf clusters
524,288 4K mbuf clusters
155,344 9K mbuf clusters
87,381 16K mbuf clusters
These defaults should be sufficient for even the most demanding
network loads.
MFC after: 1 month
2012-11-27 21:19:58 +00:00
|
|
|
EVENTHANDLER_INVOKE(nmbufs_change);
|
|
|
|
|
} else
|
|
|
|
|
error = EINVAL;
|
|
|
|
|
}
|
|
|
|
|
return (error);
|
|
|
|
|
}
|
2013-07-11 12:46:35 +00:00
|
|
|
SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs, CTLTYPE_INT|CTLFLAG_RW,
|
Base the mbuf related limits on the available physical memory or
kernel memory, whichever is lower. The overall mbuf related memory
limit must be set so that mbufs (and clusters of various sizes)
can't exhaust physical RAM or KVM.
The limit is set to half of the physical RAM or KVM (whichever is
lower) as the baseline. In any normal scenario we want to leave
at least half of the physmem/kvm for other kernel functions and
userspace to prevent it from swapping too easily. Via a tunable
kern.maxmbufmem the limit can be upped to at most 3/4 of physmem/kvm.
At the same time divorce maxfiles from maxusers and set maxfiles to
physpages / 8 with a floor based on maxusers. This way busy servers
can make use of the significantly increased mbuf limits with a much
larger number of open sockets.
Tidy up ordering in init_param2() and check up on some users of
those values calculated here.
Out of the overall mbuf memory limit 2K clusters and 4K (page size)
clusters to get 1/4 each because these are the most heavily used mbuf
sizes. 2K clusters are used for MTU 1500 ethernet inbound packets.
4K clusters are used whenever possible for sends on sockets and thus
outbound packets. The larger cluster sizes of 9K and 16K are limited
to 1/6 of the overall mbuf memory limit. When jumbo MTU's are used
these large clusters will end up only on the inbound path. They are
not used on outbound, there it's still 4K. Yes, that will stay that
way because otherwise we run into lots of complications in the
stack. And it really isn't a problem, so don't make a scene.
Normal mbufs (256B) weren't limited at all previously. This was
problematic as there are certain places in the kernel that on
allocation failure of clusters try to piece together their packet
from smaller mbufs.
The mbuf limit is the number of all other mbuf sizes together plus
some more to allow for standalone mbufs (ACK for example) and to
send off a copy of a cluster. Unfortunately there isn't a way to
set an overall limit for all mbuf memory together as UMA doesn't
support such a limiting.
NB: Every cluster also has an mbuf associated with it.
Two examples on the revised mbuf sizing limits:
1GB KVM:
512MB limit for mbufs
419,430 mbufs
65,536 2K mbuf clusters
32,768 4K mbuf clusters
9,709 9K mbuf clusters
5,461 16K mbuf clusters
16GB RAM:
8GB limit for mbufs
33,554,432 mbufs
1,048,576 2K mbuf clusters
524,288 4K mbuf clusters
155,344 9K mbuf clusters
87,381 16K mbuf clusters
These defaults should be sufficient for even the most demanding
network loads.
MFC after: 1 month
2012-11-27 21:19:58 +00:00
|
|
|
&nmbufs, 0, sysctl_nmbufs, "IU",
|
|
|
|
|
"Maximum number of mbufs allowed");
|
2007-12-05 15:29:44 +00:00
|
|
|
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
/*
|
|
|
|
|
* Zones from which we allocate.
|
|
|
|
|
*/
|
|
|
|
|
uma_zone_t zone_mbuf;
|
|
|
|
|
uma_zone_t zone_clust;
|
|
|
|
|
uma_zone_t zone_pack;
|
2006-02-17 14:14:15 +00:00
|
|
|
uma_zone_t zone_jumbop;
|
2005-11-02 16:20:36 +00:00
|
|
|
uma_zone_t zone_jumbo9;
|
|
|
|
|
uma_zone_t zone_jumbo16;
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Local prototypes.
|
|
|
|
|
*/
|
2004-08-02 00:18:36 +00:00
|
|
|
static int mb_ctor_mbuf(void *, int, void *, int);
|
|
|
|
|
static int mb_ctor_clust(void *, int, void *, int);
|
|
|
|
|
static int mb_ctor_pack(void *, int, void *, int);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
static void mb_dtor_mbuf(void *, int, void *);
|
2005-11-02 16:20:36 +00:00
|
|
|
static void mb_dtor_pack(void *, int, void *);
|
|
|
|
|
static int mb_zinit_pack(void *, int, int);
|
|
|
|
|
static void mb_zfini_pack(void *, int);
|
2016-02-03 23:30:17 +00:00
|
|
|
static void mb_reclaim(uma_zone_t, int);
|
2018-01-12 23:25:05 +00:00
|
|
|
static void *mbuf_jumbo_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
|
2013-01-17 21:28:31 +00:00
|
|
|
/* Ensure that MSIZE is a power of 2. */
|
2004-09-20 08:52:04 +00:00
|
|
|
CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
|
|
|
|
|
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
/*
|
|
|
|
|
* Initialize FreeBSD Network buffer allocation.
|
|
|
|
|
*/
|
|
|
|
|
static void
|
|
|
|
|
mbuf_init(void *dummy)
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Configure UMA zones for Mbufs, Clusters, and Packets.
|
|
|
|
|
*/
|
2005-11-02 16:20:36 +00:00
|
|
|
zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
|
|
|
|
|
mb_ctor_mbuf, mb_dtor_mbuf,
|
2005-06-23 04:33:39 +00:00
|
|
|
#ifdef INVARIANTS
|
2005-11-02 16:20:36 +00:00
|
|
|
trash_init, trash_fini,
|
2005-06-23 04:33:39 +00:00
|
|
|
#else
|
2005-11-02 16:20:36 +00:00
|
|
|
NULL, NULL,
|
2005-06-23 04:33:39 +00:00
|
|
|
#endif
|
2005-11-02 16:20:36 +00:00
|
|
|
MSIZE - 1, UMA_ZONE_MAXBUCKET);
|
2012-12-07 22:23:53 +00:00
|
|
|
if (nmbufs > 0)
|
|
|
|
|
nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
|
2012-12-07 22:30:30 +00:00
|
|
|
uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
|
2016-02-03 23:30:17 +00:00
|
|
|
uma_zone_set_maxaction(zone_mbuf, mb_reclaim);
|
2005-11-02 16:20:36 +00:00
|
|
|
|
2005-07-17 14:04:03 +00:00
|
|
|
zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
|
2016-03-01 00:17:14 +00:00
|
|
|
mb_ctor_clust,
|
2005-06-23 04:33:39 +00:00
|
|
|
#ifdef INVARIANTS
|
2016-03-01 00:17:14 +00:00
|
|
|
trash_dtor, trash_init, trash_fini,
|
2005-06-23 04:33:39 +00:00
|
|
|
#else
|
2016-03-01 00:17:14 +00:00
|
|
|
NULL, NULL, NULL,
|
2005-06-23 04:33:39 +00:00
|
|
|
#endif
|
2016-03-01 00:17:14 +00:00
|
|
|
UMA_ALIGN_PTR, 0);
|
2012-12-07 22:23:53 +00:00
|
|
|
if (nmbclusters > 0)
|
|
|
|
|
nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
|
2012-12-07 22:30:30 +00:00
|
|
|
uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
|
2016-02-03 23:30:17 +00:00
|
|
|
uma_zone_set_maxaction(zone_clust, mb_reclaim);
|
2005-11-02 16:20:36 +00:00
|
|
|
|
2005-07-17 14:04:03 +00:00
|
|
|
zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
|
2005-11-02 16:20:36 +00:00
|
|
|
mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
|
|
|
|
|
|
2006-02-26 11:44:28 +00:00
|
|
|
/* Make jumbo frame zone too. Page size, 9k and 16k. */
|
2006-02-17 14:14:15 +00:00
|
|
|
zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
|
2016-03-01 00:17:14 +00:00
|
|
|
mb_ctor_clust,
|
2005-12-08 13:13:06 +00:00
|
|
|
#ifdef INVARIANTS
|
2016-03-01 00:17:14 +00:00
|
|
|
trash_dtor, trash_init, trash_fini,
|
2005-12-08 13:13:06 +00:00
|
|
|
#else
|
2016-03-01 00:17:14 +00:00
|
|
|
NULL, NULL, NULL,
|
2005-12-08 13:13:06 +00:00
|
|
|
#endif
|
2016-03-01 00:17:14 +00:00
|
|
|
UMA_ALIGN_PTR, 0);
|
2012-12-07 22:23:53 +00:00
|
|
|
if (nmbjumbop > 0)
|
|
|
|
|
nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
|
2012-12-07 22:30:30 +00:00
|
|
|
uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
|
2016-02-03 23:30:17 +00:00
|
|
|
uma_zone_set_maxaction(zone_jumbop, mb_reclaim);
|
2005-12-08 13:13:06 +00:00
|
|
|
|
2005-11-02 16:20:36 +00:00
|
|
|
zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
|
2016-03-01 00:17:14 +00:00
|
|
|
mb_ctor_clust,
|
2005-11-02 16:20:36 +00:00
|
|
|
#ifdef INVARIANTS
|
2016-03-01 00:17:14 +00:00
|
|
|
trash_dtor, trash_init, trash_fini,
|
2005-11-02 16:20:36 +00:00
|
|
|
#else
|
2016-03-01 00:17:14 +00:00
|
|
|
NULL, NULL, NULL,
|
2005-11-02 16:20:36 +00:00
|
|
|
#endif
|
2016-03-01 00:17:14 +00:00
|
|
|
UMA_ALIGN_PTR, 0);
|
2007-12-04 07:06:08 +00:00
|
|
|
uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc);
|
2012-12-07 22:23:53 +00:00
|
|
|
if (nmbjumbo9 > 0)
|
|
|
|
|
nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
|
2012-12-07 22:30:30 +00:00
|
|
|
uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
|
2016-02-03 23:30:17 +00:00
|
|
|
uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
|
2005-11-02 16:20:36 +00:00
|
|
|
|
|
|
|
|
zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
|
2016-03-01 00:17:14 +00:00
|
|
|
mb_ctor_clust,
|
2005-11-02 16:20:36 +00:00
|
|
|
#ifdef INVARIANTS
|
2016-03-01 00:17:14 +00:00
|
|
|
trash_dtor, trash_init, trash_fini,
|
2005-11-02 16:20:36 +00:00
|
|
|
#else
|
2016-03-01 00:17:14 +00:00
|
|
|
NULL, NULL, NULL,
|
2005-11-02 16:20:36 +00:00
|
|
|
#endif
|
2016-03-01 00:17:14 +00:00
|
|
|
UMA_ALIGN_PTR, 0);
|
2007-12-04 07:06:08 +00:00
|
|
|
uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc);
|
2012-12-07 22:23:53 +00:00
|
|
|
if (nmbjumbo16 > 0)
|
|
|
|
|
nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
|
2012-12-07 22:30:30 +00:00
|
|
|
uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
|
2016-02-03 23:30:17 +00:00
|
|
|
uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Hook event handler for low-memory situation, used to
|
|
|
|
|
* drain protocols and push data back to the caches (UMA
|
|
|
|
|
* later pushes it back to VM).
|
|
|
|
|
*/
|
|
|
|
|
EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
|
|
|
|
|
EVENTHANDLER_PRI_FIRST);
|
|
|
|
|
}
|
2013-01-17 21:28:31 +00:00
|
|
|
SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
|
2018-05-06 00:19:48 +00:00
|
|
|
#ifdef NETDUMP
|
|
|
|
|
/*
|
|
|
|
|
* netdump makes use of a pre-allocated pool of mbufs and clusters. When
|
|
|
|
|
* netdump is configured, we initialize a set of UMA cache zones which return
|
|
|
|
|
* items from this pool. At panic-time, the regular UMA zone pointers are
|
|
|
|
|
* overwritten with those of the cache zones so that drivers may allocate and
|
|
|
|
|
* free mbufs and clusters without attempting to allocate physical memory.
|
|
|
|
|
*
|
|
|
|
|
* We keep mbufs and clusters in a pair of mbuf queues. In particular, for
|
|
|
|
|
* the purpose of caching clusters, we treat them as mbufs.
|
|
|
|
|
*/
|
|
|
|
|
static struct mbufq nd_mbufq =
|
|
|
|
|
{ STAILQ_HEAD_INITIALIZER(nd_mbufq.mq_head), 0, INT_MAX };
|
|
|
|
|
static struct mbufq nd_clustq =
|
|
|
|
|
{ STAILQ_HEAD_INITIALIZER(nd_clustq.mq_head), 0, INT_MAX };
|
|
|
|
|
|
|
|
|
|
static int nd_clsize;
|
|
|
|
|
static uma_zone_t nd_zone_mbuf;
|
|
|
|
|
static uma_zone_t nd_zone_clust;
|
|
|
|
|
static uma_zone_t nd_zone_pack;
|
|
|
|
|
|
|
|
|
|
static int
|
|
|
|
|
nd_buf_import(void *arg, void **store, int count, int domain __unused,
|
|
|
|
|
int flags)
|
|
|
|
|
{
|
|
|
|
|
struct mbufq *q;
|
|
|
|
|
struct mbuf *m;
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
q = arg;
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < count; i++) {
|
|
|
|
|
m = mbufq_dequeue(q);
|
|
|
|
|
if (m == NULL)
|
|
|
|
|
break;
|
|
|
|
|
trash_init(m, q == &nd_mbufq ? MSIZE : nd_clsize, flags);
|
|
|
|
|
store[i] = m;
|
|
|
|
|
}
|
2018-11-12 05:24:20 +00:00
|
|
|
KASSERT((flags & M_WAITOK) == 0 || i == count,
|
|
|
|
|
("%s: ran out of pre-allocated mbufs", __func__));
|
2018-05-06 00:19:48 +00:00
|
|
|
return (i);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void
|
|
|
|
|
nd_buf_release(void *arg, void **store, int count)
|
|
|
|
|
{
|
|
|
|
|
struct mbufq *q;
|
|
|
|
|
struct mbuf *m;
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
q = arg;
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < count; i++) {
|
|
|
|
|
m = store[i];
|
|
|
|
|
(void)mbufq_enqueue(q, m);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static int
|
|
|
|
|
nd_pack_import(void *arg __unused, void **store, int count, int domain __unused,
|
|
|
|
|
int flags __unused)
|
|
|
|
|
{
|
|
|
|
|
struct mbuf *m;
|
|
|
|
|
void *clust;
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < count; i++) {
|
|
|
|
|
m = m_get(MT_DATA, M_NOWAIT);
|
|
|
|
|
if (m == NULL)
|
|
|
|
|
break;
|
|
|
|
|
clust = uma_zalloc(nd_zone_clust, M_NOWAIT);
|
|
|
|
|
if (clust == NULL) {
|
|
|
|
|
m_free(m);
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
mb_ctor_clust(clust, nd_clsize, m, 0);
|
|
|
|
|
store[i] = m;
|
|
|
|
|
}
|
2018-11-12 05:24:20 +00:00
|
|
|
KASSERT((flags & M_WAITOK) == 0 || i == count,
|
|
|
|
|
("%s: ran out of pre-allocated mbufs", __func__));
|
2018-05-06 00:19:48 +00:00
|
|
|
return (i);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void
|
|
|
|
|
nd_pack_release(void *arg __unused, void **store, int count)
|
|
|
|
|
{
|
|
|
|
|
struct mbuf *m;
|
|
|
|
|
void *clust;
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < count; i++) {
|
|
|
|
|
m = store[i];
|
|
|
|
|
clust = m->m_ext.ext_buf;
|
|
|
|
|
uma_zfree(nd_zone_clust, clust);
|
|
|
|
|
uma_zfree(nd_zone_mbuf, m);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Free the pre-allocated mbufs and clusters reserved for netdump, and destroy
|
|
|
|
|
* the corresponding UMA cache zones.
|
|
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
netdump_mbuf_drain(void)
|
|
|
|
|
{
|
|
|
|
|
struct mbuf *m;
|
|
|
|
|
void *item;
|
|
|
|
|
|
|
|
|
|
if (nd_zone_mbuf != NULL) {
|
|
|
|
|
uma_zdestroy(nd_zone_mbuf);
|
|
|
|
|
nd_zone_mbuf = NULL;
|
|
|
|
|
}
|
|
|
|
|
if (nd_zone_clust != NULL) {
|
|
|
|
|
uma_zdestroy(nd_zone_clust);
|
|
|
|
|
nd_zone_clust = NULL;
|
|
|
|
|
}
|
|
|
|
|
if (nd_zone_pack != NULL) {
|
|
|
|
|
uma_zdestroy(nd_zone_pack);
|
|
|
|
|
nd_zone_pack = NULL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
while ((m = mbufq_dequeue(&nd_mbufq)) != NULL)
|
|
|
|
|
m_free(m);
|
|
|
|
|
while ((item = mbufq_dequeue(&nd_clustq)) != NULL)
|
|
|
|
|
uma_zfree(m_getzone(nd_clsize), item);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Callback invoked immediately prior to starting a netdump.
|
|
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
netdump_mbuf_dump(void)
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* All cluster zones return buffers of the size requested by the
|
|
|
|
|
* drivers. It's up to the driver to reinitialize the zones if the
|
|
|
|
|
* MTU of a netdump-enabled interface changes.
|
|
|
|
|
*/
|
|
|
|
|
printf("netdump: overwriting mbuf zone pointers\n");
|
|
|
|
|
zone_mbuf = nd_zone_mbuf;
|
|
|
|
|
zone_clust = nd_zone_clust;
|
|
|
|
|
zone_pack = nd_zone_pack;
|
|
|
|
|
zone_jumbop = nd_zone_clust;
|
|
|
|
|
zone_jumbo9 = nd_zone_clust;
|
|
|
|
|
zone_jumbo16 = nd_zone_clust;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Reinitialize the netdump mbuf+cluster pool and cache zones.
|
|
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
netdump_mbuf_reinit(int nmbuf, int nclust, int clsize)
|
|
|
|
|
{
|
|
|
|
|
struct mbuf *m;
|
|
|
|
|
void *item;
|
|
|
|
|
|
|
|
|
|
netdump_mbuf_drain();
|
|
|
|
|
|
|
|
|
|
nd_clsize = clsize;
|
|
|
|
|
|
|
|
|
|
nd_zone_mbuf = uma_zcache_create("netdump_" MBUF_MEM_NAME,
|
|
|
|
|
MSIZE, mb_ctor_mbuf, mb_dtor_mbuf,
|
|
|
|
|
#ifdef INVARIANTS
|
|
|
|
|
trash_init, trash_fini,
|
|
|
|
|
#else
|
|
|
|
|
NULL, NULL,
|
|
|
|
|
#endif
|
|
|
|
|
nd_buf_import, nd_buf_release,
|
|
|
|
|
&nd_mbufq, UMA_ZONE_NOBUCKET);
|
|
|
|
|
|
|
|
|
|
nd_zone_clust = uma_zcache_create("netdump_" MBUF_CLUSTER_MEM_NAME,
|
|
|
|
|
clsize, mb_ctor_clust,
|
|
|
|
|
#ifdef INVARIANTS
|
|
|
|
|
trash_dtor, trash_init, trash_fini,
|
|
|
|
|
#else
|
|
|
|
|
NULL, NULL, NULL,
|
|
|
|
|
#endif
|
|
|
|
|
nd_buf_import, nd_buf_release,
|
|
|
|
|
&nd_clustq, UMA_ZONE_NOBUCKET);
|
|
|
|
|
|
|
|
|
|
nd_zone_pack = uma_zcache_create("netdump_" MBUF_PACKET_MEM_NAME,
|
|
|
|
|
MCLBYTES, mb_ctor_pack, mb_dtor_pack, NULL, NULL,
|
|
|
|
|
nd_pack_import, nd_pack_release,
|
|
|
|
|
NULL, UMA_ZONE_NOBUCKET);
|
|
|
|
|
|
|
|
|
|
while (nmbuf-- > 0) {
|
|
|
|
|
m = m_get(MT_DATA, M_WAITOK);
|
|
|
|
|
uma_zfree(nd_zone_mbuf, m);
|
|
|
|
|
}
|
|
|
|
|
while (nclust-- > 0) {
|
|
|
|
|
item = uma_zalloc(m_getzone(nd_clsize), M_WAITOK);
|
|
|
|
|
uma_zfree(nd_zone_clust, item);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
#endif /* NETDUMP */
|
|
|
|
|
|
2007-12-04 07:06:08 +00:00
|
|
|
/*
|
|
|
|
|
* UMA backend page allocator for the jumbo frame zones.
|
|
|
|
|
*
|
|
|
|
|
* Allocates kernel virtual memory that is backed by contiguous physical
|
|
|
|
|
* pages.
|
|
|
|
|
*/
|
|
|
|
|
static void *
|
2018-01-12 23:25:05 +00:00
|
|
|
mbuf_jumbo_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
|
|
|
|
|
int wait)
|
2007-12-04 07:06:08 +00:00
|
|
|
{
|
|
|
|
|
|
2008-04-04 18:41:12 +00:00
|
|
|
/* Inform UMA that this allocator uses kernel_map/object. */
|
|
|
|
|
*flags = UMA_SLAB_KERNEL;
|
2018-10-30 18:26:34 +00:00
|
|
|
return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
|
|
|
|
|
bytes, wait, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
|
|
|
|
|
VM_MEMATTR_DEFAULT));
|
2007-12-04 07:06:08 +00:00
|
|
|
}
|
|
|
|
|
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
/*
|
|
|
|
|
* Constructor for Mbuf master zone.
|
|
|
|
|
*
|
|
|
|
|
* The 'arg' pointer points to a mb_args structure which
|
|
|
|
|
* contains call-specific information required to support the
|
2005-11-02 16:20:36 +00:00
|
|
|
* mbuf allocation API. See mbuf.h.
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
*/
|
2004-08-02 00:18:36 +00:00
|
|
|
static int
|
|
|
|
|
mb_ctor_mbuf(void *mem, int size, void *arg, int how)
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
{
|
|
|
|
|
struct mbuf *m;
|
|
|
|
|
struct mb_args *args;
|
2004-08-02 00:18:36 +00:00
|
|
|
int error;
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
int flags;
|
|
|
|
|
short type;
|
|
|
|
|
|
2005-06-23 04:33:39 +00:00
|
|
|
#ifdef INVARIANTS
|
|
|
|
|
trash_ctor(mem, size, arg, how);
|
|
|
|
|
#endif
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
args = (struct mb_args *)arg;
|
|
|
|
|
type = args->type;
|
|
|
|
|
|
2005-11-02 16:20:36 +00:00
|
|
|
/*
|
|
|
|
|
* The mbuf is initialized later. The caller has the
|
2006-02-27 07:22:32 +00:00
|
|
|
* responsibility to set up any MAC labels too.
|
2005-11-02 16:20:36 +00:00
|
|
|
*/
|
|
|
|
|
if (type == MT_NOINIT)
|
|
|
|
|
return (0);
|
|
|
|
|
|
2013-08-24 12:24:58 +00:00
|
|
|
m = (struct mbuf *)mem;
|
|
|
|
|
flags = args->flags;
|
2016-03-26 23:39:53 +00:00
|
|
|
MPASS((flags & M_NOFREE) == 0);
|
2013-08-24 12:24:58 +00:00
|
|
|
|
2016-02-10 18:54:18 +00:00
|
|
|
error = m_init(m, how, type, flags);
|
2013-08-24 12:24:58 +00:00
|
|
|
|
|
|
|
|
return (error);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2005-11-02 16:20:36 +00:00
|
|
|
* The Mbuf master zone destructor.
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
*/
|
|
|
|
|
static void
|
|
|
|
|
mb_dtor_mbuf(void *mem, int size, void *arg)
|
|
|
|
|
{
|
|
|
|
|
struct mbuf *m;
|
2012-12-07 22:19:41 +00:00
|
|
|
unsigned long flags;
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
|
|
|
|
|
m = (struct mbuf *)mem;
|
2007-10-06 21:13:55 +00:00
|
|
|
flags = (unsigned long)arg;
|
2007-10-26 16:33:47 +00:00
|
|
|
|
2014-09-30 23:16:26 +00:00
|
|
|
KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
|
2016-05-18 04:35:58 +00:00
|
|
|
if (!(flags & MB_DTOR_SKIP) && (m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
m_tag_delete_chain(m, NULL);
|
2005-06-23 04:33:39 +00:00
|
|
|
#ifdef INVARIANTS
|
|
|
|
|
trash_dtor(mem, size, arg);
|
|
|
|
|
#endif
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
}
|
|
|
|
|
|
2005-11-02 16:20:36 +00:00
|
|
|
/*
|
|
|
|
|
* The Mbuf Packet zone destructor.
|
|
|
|
|
*/
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
static void
|
|
|
|
|
mb_dtor_pack(void *mem, int size, void *arg)
|
|
|
|
|
{
|
|
|
|
|
struct mbuf *m;
|
|
|
|
|
|
|
|
|
|
m = (struct mbuf *)mem;
|
|
|
|
|
if ((m->m_flags & M_PKTHDR) != 0)
|
|
|
|
|
m_tag_delete_chain(m, NULL);
|
2005-11-02 16:20:36 +00:00
|
|
|
|
|
|
|
|
/* Make sure we've got a clean cluster back. */
|
|
|
|
|
KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
|
|
|
|
|
KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
|
|
|
|
|
KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
|
2008-02-01 19:36:27 +00:00
|
|
|
KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
|
|
|
|
|
KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
|
2005-11-02 16:20:36 +00:00
|
|
|
KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
|
2005-11-06 16:47:59 +00:00
|
|
|
KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
|
2005-06-23 04:33:39 +00:00
|
|
|
#ifdef INVARIANTS
|
|
|
|
|
trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
|
|
|
|
|
#endif
|
2007-01-25 01:05:23 +00:00
|
|
|
/*
|
2007-10-26 16:33:47 +00:00
|
|
|
* If there are processes blocked on zone_clust, waiting for pages
|
|
|
|
|
* to be freed up, * cause them to be woken up by draining the
|
|
|
|
|
* packet zone. We are exposed to a race here * (in the check for
|
|
|
|
|
* the UMA_ZFLAG_FULL) where we might miss the flag set, but that
|
|
|
|
|
* is deliberate. We don't want to acquire the zone lock for every
|
|
|
|
|
* mbuf free.
|
2007-01-25 01:05:23 +00:00
|
|
|
*/
|
2007-10-26 16:33:47 +00:00
|
|
|
if (uma_zone_exhausted_nolock(zone_clust))
|
|
|
|
|
zone_drain(zone_pack);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2006-02-17 14:14:15 +00:00
|
|
|
* The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
*
|
|
|
|
|
* Here the 'arg' pointer points to the Mbuf which we
|
2005-11-02 16:20:36 +00:00
|
|
|
* are configuring cluster storage for. If 'arg' is
|
|
|
|
|
* empty we allocate just the cluster without setting
|
|
|
|
|
* the mbuf to it. See mbuf.h.
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
*/
|
2004-08-02 00:18:36 +00:00
|
|
|
static int
|
|
|
|
|
mb_ctor_clust(void *mem, int size, void *arg, int how)
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
{
|
|
|
|
|
struct mbuf *m;
|
2007-10-26 16:33:47 +00:00
|
|
|
|
2005-06-23 04:33:39 +00:00
|
|
|
#ifdef INVARIANTS
|
|
|
|
|
trash_ctor(mem, size, arg, how);
|
|
|
|
|
#endif
|
2007-04-04 21:27:01 +00:00
|
|
|
m = (struct mbuf *)arg;
|
|
|
|
|
if (m != NULL) {
|
2017-10-09 20:35:31 +00:00
|
|
|
m->m_ext.ext_buf = (char *)mem;
|
2005-11-02 16:20:36 +00:00
|
|
|
m->m_data = m->m_ext.ext_buf;
|
|
|
|
|
m->m_flags |= M_EXT;
|
|
|
|
|
m->m_ext.ext_free = NULL;
|
2008-02-01 19:36:27 +00:00
|
|
|
m->m_ext.ext_arg1 = NULL;
|
|
|
|
|
m->m_ext.ext_arg2 = NULL;
|
2005-11-02 16:20:36 +00:00
|
|
|
m->m_ext.ext_size = size;
|
2016-03-01 00:17:14 +00:00
|
|
|
m->m_ext.ext_type = m_gettype(size);
|
|
|
|
|
m->m_ext.ext_flags = EXT_FLAG_EMBREF;
|
|
|
|
|
m->m_ext.ext_count = 1;
|
2005-11-02 16:20:36 +00:00
|
|
|
}
|
2007-04-04 21:27:01 +00:00
|
|
|
|
2004-08-02 00:18:36 +00:00
|
|
|
return (0);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The Packet secondary zone's init routine, executed on the
|
2005-11-02 16:20:36 +00:00
|
|
|
* object's transition from mbuf keg slab to zone cache.
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
*/
|
2004-08-02 00:18:36 +00:00
|
|
|
static int
|
2005-11-02 16:20:36 +00:00
|
|
|
mb_zinit_pack(void *mem, int size, int how)
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
{
|
|
|
|
|
struct mbuf *m;
|
|
|
|
|
|
2005-11-02 16:20:36 +00:00
|
|
|
m = (struct mbuf *)mem; /* m is virgin. */
|
2006-03-08 14:05:38 +00:00
|
|
|
if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
|
|
|
|
|
m->m_ext.ext_buf == NULL)
|
2004-08-02 00:18:36 +00:00
|
|
|
return (ENOMEM);
|
2005-11-05 19:43:55 +00:00
|
|
|
m->m_ext.ext_type = EXT_PACKET; /* Override. */
|
2005-06-23 04:33:39 +00:00
|
|
|
#ifdef INVARIANTS
|
|
|
|
|
trash_init(m->m_ext.ext_buf, MCLBYTES, how);
|
|
|
|
|
#endif
|
2004-08-02 00:18:36 +00:00
|
|
|
return (0);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The Packet secondary zone's fini routine, executed on the
|
|
|
|
|
* object's transition from zone cache to keg slab.
|
|
|
|
|
*/
|
|
|
|
|
static void
|
2005-11-02 16:20:36 +00:00
|
|
|
mb_zfini_pack(void *mem, int size)
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
{
|
|
|
|
|
struct mbuf *m;
|
|
|
|
|
|
|
|
|
|
m = (struct mbuf *)mem;
|
2005-06-23 04:33:39 +00:00
|
|
|
#ifdef INVARIANTS
|
|
|
|
|
trash_fini(m->m_ext.ext_buf, MCLBYTES);
|
|
|
|
|
#endif
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
|
2005-06-29 08:18:26 +00:00
|
|
|
#ifdef INVARIANTS
|
|
|
|
|
trash_dtor(mem, size, NULL);
|
|
|
|
|
#endif
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The "packet" keg constructor.
|
|
|
|
|
*/
|
2004-08-02 00:18:36 +00:00
|
|
|
static int
|
|
|
|
|
mb_ctor_pack(void *mem, int size, void *arg, int how)
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
{
|
|
|
|
|
struct mbuf *m;
|
|
|
|
|
struct mb_args *args;
|
2013-08-24 21:25:53 +00:00
|
|
|
int error, flags;
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
short type;
|
|
|
|
|
|
|
|
|
|
m = (struct mbuf *)mem;
|
|
|
|
|
args = (struct mb_args *)arg;
|
|
|
|
|
flags = args->flags;
|
|
|
|
|
type = args->type;
|
2016-03-26 23:39:53 +00:00
|
|
|
MPASS((flags & M_NOFREE) == 0);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
|
2005-06-23 04:33:39 +00:00
|
|
|
#ifdef INVARIANTS
|
|
|
|
|
trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
|
|
|
|
|
#endif
|
2013-08-24 12:24:58 +00:00
|
|
|
|
2016-02-10 18:54:18 +00:00
|
|
|
error = m_init(m, how, type, flags);
|
2013-08-24 12:24:58 +00:00
|
|
|
|
2005-11-02 16:20:36 +00:00
|
|
|
/* m_ext is already initialized. */
|
2013-08-24 12:24:58 +00:00
|
|
|
m->m_data = m->m_ext.ext_buf;
|
|
|
|
|
m->m_flags = (flags | M_EXT);
|
2005-11-02 16:20:36 +00:00
|
|
|
|
2013-08-24 12:24:58 +00:00
|
|
|
return (error);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2016-02-03 23:30:17 +00:00
|
|
|
* This is the protocol drain routine. Called by UMA whenever any of the
|
|
|
|
|
* mbuf zones is closed to its limit.
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
*
|
|
|
|
|
* No locks should be held when this is called. The drain routines have to
|
|
|
|
|
* presently acquire some locks which raises the possibility of lock order
|
|
|
|
|
* reversal.
|
|
|
|
|
*/
|
|
|
|
|
static void
|
2016-02-03 23:30:17 +00:00
|
|
|
mb_reclaim(uma_zone_t zone __unused, int pending __unused)
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
{
|
|
|
|
|
struct domain *dp;
|
|
|
|
|
struct protosw *pr;
|
|
|
|
|
|
2016-02-03 23:30:17 +00:00
|
|
|
WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, __func__);
|
Bring in mbuma to replace mballoc.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
2004-05-31 21:46:06 +00:00
|
|
|
|
|
|
|
|
for (dp = domains; dp != NULL; dp = dp->dom_next)
|
|
|
|
|
for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
|
|
|
|
|
if (pr->pr_drain != NULL)
|
|
|
|
|
(*pr->pr_drain)();
|
|
|
|
|
}
|
2016-02-11 21:32:23 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Clean up after mbufs with M_EXT storage attached to them if the
|
|
|
|
|
* reference count hits 1.
|
|
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
mb_free_ext(struct mbuf *m)
|
|
|
|
|
{
|
2016-03-01 00:17:14 +00:00
|
|
|
volatile u_int *refcnt;
|
|
|
|
|
struct mbuf *mref;
|
2016-02-11 21:32:23 +00:00
|
|
|
int freembuf;
|
|
|
|
|
|
|
|
|
|
KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
|
|
|
|
|
|
2016-03-01 00:17:14 +00:00
|
|
|
/* See if this is the mbuf that holds the embedded refcount. */
|
|
|
|
|
if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
|
|
|
|
|
refcnt = &m->m_ext.ext_count;
|
|
|
|
|
mref = m;
|
|
|
|
|
} else {
|
|
|
|
|
KASSERT(m->m_ext.ext_cnt != NULL,
|
|
|
|
|
("%s: no refcounting pointer on %p", __func__, m));
|
|
|
|
|
refcnt = m->m_ext.ext_cnt;
|
|
|
|
|
mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
|
|
|
|
|
}
|
|
|
|
|
|
2016-02-11 21:32:23 +00:00
|
|
|
/*
|
2016-03-01 00:17:14 +00:00
|
|
|
* Check if the header is embedded in the cluster. It is
|
|
|
|
|
* important that we can't touch any of the mbuf fields
|
|
|
|
|
* after we have freed the external storage, since mbuf
|
2016-05-17 23:15:16 +00:00
|
|
|
* could have been embedded in it. For now, the mbufs
|
|
|
|
|
* embedded into the cluster are always of type EXT_EXTREF,
|
|
|
|
|
* and for this type we won't free the mref.
|
2016-02-11 21:32:23 +00:00
|
|
|
*/
|
2016-05-17 23:15:16 +00:00
|
|
|
if (m->m_flags & M_NOFREE) {
|
|
|
|
|
freembuf = 0;
|
2019-01-31 22:37:28 +00:00
|
|
|
KASSERT(m->m_ext.ext_type == EXT_EXTREF ||
|
|
|
|
|
m->m_ext.ext_type == EXT_RXRING,
|
2016-05-17 23:15:16 +00:00
|
|
|
("%s: no-free mbuf %p has wrong type", __func__, m));
|
|
|
|
|
} else
|
|
|
|
|
freembuf = 1;
|
2016-02-11 21:32:23 +00:00
|
|
|
|
2016-03-01 00:17:14 +00:00
|
|
|
/* Free attached storage if this mbuf is the only reference to it. */
|
|
|
|
|
if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
|
2016-02-11 21:32:23 +00:00
|
|
|
switch (m->m_ext.ext_type) {
|
2016-03-01 00:17:14 +00:00
|
|
|
case EXT_PACKET:
|
|
|
|
|
/* The packet zone is special. */
|
|
|
|
|
if (*refcnt == 0)
|
|
|
|
|
*refcnt = 1;
|
|
|
|
|
uma_zfree(zone_pack, mref);
|
|
|
|
|
break;
|
2016-02-11 21:32:23 +00:00
|
|
|
case EXT_CLUSTER:
|
|
|
|
|
uma_zfree(zone_clust, m->m_ext.ext_buf);
|
2016-03-01 00:17:14 +00:00
|
|
|
uma_zfree(zone_mbuf, mref);
|
2016-02-11 21:32:23 +00:00
|
|
|
break;
|
|
|
|
|
case EXT_JUMBOP:
|
|
|
|
|
uma_zfree(zone_jumbop, m->m_ext.ext_buf);
|
2016-03-01 00:17:14 +00:00
|
|
|
uma_zfree(zone_mbuf, mref);
|
2016-02-11 21:32:23 +00:00
|
|
|
break;
|
|
|
|
|
case EXT_JUMBO9:
|
|
|
|
|
uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
|
2016-03-01 00:17:14 +00:00
|
|
|
uma_zfree(zone_mbuf, mref);
|
2016-02-11 21:32:23 +00:00
|
|
|
break;
|
|
|
|
|
case EXT_JUMBO16:
|
|
|
|
|
uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
|
2016-03-01 00:17:14 +00:00
|
|
|
uma_zfree(zone_mbuf, mref);
|
|
|
|
|
break;
|
|
|
|
|
case EXT_SFBUF:
|
2016-02-11 21:32:23 +00:00
|
|
|
case EXT_NET_DRV:
|
|
|
|
|
case EXT_MOD_TYPE:
|
|
|
|
|
case EXT_DISPOSABLE:
|
2017-10-09 20:51:58 +00:00
|
|
|
KASSERT(mref->m_ext.ext_free != NULL,
|
2018-05-06 00:11:30 +00:00
|
|
|
("%s: ext_free not set", __func__));
|
2017-10-09 20:51:58 +00:00
|
|
|
mref->m_ext.ext_free(mref);
|
2016-03-01 00:17:14 +00:00
|
|
|
uma_zfree(zone_mbuf, mref);
|
2016-03-02 02:12:01 +00:00
|
|
|
break;
|
2016-02-11 21:32:23 +00:00
|
|
|
case EXT_EXTREF:
|
|
|
|
|
KASSERT(m->m_ext.ext_free != NULL,
|
2018-05-06 00:11:30 +00:00
|
|
|
("%s: ext_free not set", __func__));
|
2017-10-09 20:35:31 +00:00
|
|
|
m->m_ext.ext_free(m);
|
2016-02-11 21:32:23 +00:00
|
|
|
break;
|
2019-01-31 22:37:28 +00:00
|
|
|
case EXT_RXRING:
|
|
|
|
|
KASSERT(m->m_ext.ext_free == NULL,
|
|
|
|
|
("%s: ext_free is set", __func__));
|
|
|
|
|
break;
|
2016-02-11 21:32:23 +00:00
|
|
|
default:
|
|
|
|
|
KASSERT(m->m_ext.ext_type == 0,
|
2018-05-06 00:11:30 +00:00
|
|
|
("%s: unknown ext_type", __func__));
|
2016-02-11 21:32:23 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2016-03-01 00:17:14 +00:00
|
|
|
if (freembuf && m != mref)
|
2016-02-11 21:32:23 +00:00
|
|
|
uma_zfree(zone_mbuf, m);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Official mbuf(9) allocation KPI for stack and drivers:
|
|
|
|
|
*
|
|
|
|
|
* m_get() - a single mbuf without any attachments, sys/mbuf.h.
|
|
|
|
|
* m_gethdr() - a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
|
|
|
|
|
* m_getcl() - an mbuf + 2k cluster, sys/mbuf.h.
|
|
|
|
|
* m_clget() - attach cluster to already allocated mbuf.
|
|
|
|
|
* m_cljget() - attach jumbo cluster to already allocated mbuf.
|
|
|
|
|
* m_get2() - allocate minimum mbuf that would fit size argument.
|
|
|
|
|
* m_getm2() - allocate a chain of mbufs/clusters.
|
|
|
|
|
* m_extadd() - attach external cluster to mbuf.
|
|
|
|
|
*
|
|
|
|
|
* m_free() - free single mbuf with its tags and ext, sys/mbuf.h.
|
|
|
|
|
* m_freem() - free chain of mbufs.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
int
|
|
|
|
|
m_clget(struct mbuf *m, int how)
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
|
|
|
|
|
__func__, m));
|
|
|
|
|
m->m_ext.ext_buf = (char *)NULL;
|
|
|
|
|
uma_zalloc_arg(zone_clust, m, how);
|
|
|
|
|
/*
|
|
|
|
|
* On a cluster allocation failure, drain the packet zone and retry,
|
|
|
|
|
* we might be able to loosen a few clusters up on the drain.
|
|
|
|
|
*/
|
|
|
|
|
if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
|
|
|
|
|
zone_drain(zone_pack);
|
|
|
|
|
uma_zalloc_arg(zone_clust, m, how);
|
|
|
|
|
}
|
2016-03-22 13:16:52 +00:00
|
|
|
MBUF_PROBE2(m__clget, m, how);
|
2016-02-11 21:32:23 +00:00
|
|
|
return (m->m_flags & M_EXT);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* m_cljget() is different from m_clget() as it can allocate clusters without
|
|
|
|
|
* attaching them to an mbuf. In that case the return value is the pointer
|
|
|
|
|
* to the cluster of the requested size. If an mbuf was specified, it gets
|
|
|
|
|
* the cluster attached to it and the return value can be safely ignored.
|
|
|
|
|
* For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
|
|
|
|
|
*/
|
|
|
|
|
void *
|
|
|
|
|
m_cljget(struct mbuf *m, int how, int size)
|
|
|
|
|
{
|
|
|
|
|
uma_zone_t zone;
|
2016-03-22 13:16:52 +00:00
|
|
|
void *retval;
|
2016-03-24 20:06:52 +00:00
|
|
|
|
2016-02-11 21:32:23 +00:00
|
|
|
if (m != NULL) {
|
|
|
|
|
KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
|
|
|
|
|
__func__, m));
|
|
|
|
|
m->m_ext.ext_buf = NULL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
zone = m_getzone(size);
|
2016-03-22 13:16:52 +00:00
|
|
|
retval = uma_zalloc_arg(zone, m, how);
|
|
|
|
|
|
|
|
|
|
MBUF_PROBE4(m__cljget, m, how, size, retval);
|
|
|
|
|
|
|
|
|
|
return (retval);
|
2016-02-11 21:32:23 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* m_get2() allocates minimum mbuf that would fit "size" argument.
|
|
|
|
|
*/
|
|
|
|
|
struct mbuf *
|
|
|
|
|
m_get2(int size, int how, short type, int flags)
|
|
|
|
|
{
|
|
|
|
|
struct mb_args args;
|
|
|
|
|
struct mbuf *m, *n;
|
|
|
|
|
|
|
|
|
|
args.flags = flags;
|
|
|
|
|
args.type = type;
|
|
|
|
|
|
|
|
|
|
if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
|
|
|
|
|
return (uma_zalloc_arg(zone_mbuf, &args, how));
|
|
|
|
|
if (size <= MCLBYTES)
|
|
|
|
|
return (uma_zalloc_arg(zone_pack, &args, how));
|
|
|
|
|
|
|
|
|
|
if (size > MJUMPAGESIZE)
|
|
|
|
|
return (NULL);
|
|
|
|
|
|
|
|
|
|
m = uma_zalloc_arg(zone_mbuf, &args, how);
|
|
|
|
|
if (m == NULL)
|
|
|
|
|
return (NULL);
|
|
|
|
|
|
|
|
|
|
n = uma_zalloc_arg(zone_jumbop, m, how);
|
|
|
|
|
if (n == NULL) {
|
|
|
|
|
uma_zfree(zone_mbuf, m);
|
|
|
|
|
return (NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return (m);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* m_getjcl() returns an mbuf with a cluster of the specified size attached.
|
|
|
|
|
* For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
|
|
|
|
|
*/
|
|
|
|
|
struct mbuf *
|
|
|
|
|
m_getjcl(int how, short type, int flags, int size)
|
|
|
|
|
{
|
|
|
|
|
struct mb_args args;
|
|
|
|
|
struct mbuf *m, *n;
|
|
|
|
|
uma_zone_t zone;
|
|
|
|
|
|
|
|
|
|
if (size == MCLBYTES)
|
|
|
|
|
return m_getcl(how, type, flags);
|
|
|
|
|
|
|
|
|
|
args.flags = flags;
|
|
|
|
|
args.type = type;
|
|
|
|
|
|
|
|
|
|
m = uma_zalloc_arg(zone_mbuf, &args, how);
|
|
|
|
|
if (m == NULL)
|
|
|
|
|
return (NULL);
|
|
|
|
|
|
|
|
|
|
zone = m_getzone(size);
|
|
|
|
|
n = uma_zalloc_arg(zone, m, how);
|
|
|
|
|
if (n == NULL) {
|
|
|
|
|
uma_zfree(zone_mbuf, m);
|
|
|
|
|
return (NULL);
|
|
|
|
|
}
|
|
|
|
|
return (m);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Allocate a given length worth of mbufs and/or clusters (whatever fits
|
|
|
|
|
* best) and return a pointer to the top of the allocated chain. If an
|
|
|
|
|
* existing mbuf chain is provided, then we will append the new chain
|
2019-01-27 16:44:27 +00:00
|
|
|
* to the existing one and return a pointer to the provided mbuf.
|
2016-02-11 21:32:23 +00:00
|
|
|
*/
|
|
|
|
|
struct mbuf *
|
|
|
|
|
m_getm2(struct mbuf *m, int len, int how, short type, int flags)
|
|
|
|
|
{
|
|
|
|
|
struct mbuf *mb, *nm = NULL, *mtail = NULL;
|
|
|
|
|
|
|
|
|
|
KASSERT(len >= 0, ("%s: len is < 0", __func__));
|
|
|
|
|
|
|
|
|
|
/* Validate flags. */
|
|
|
|
|
flags &= (M_PKTHDR | M_EOR);
|
|
|
|
|
|
|
|
|
|
/* Packet header mbuf must be first in chain. */
|
|
|
|
|
if ((flags & M_PKTHDR) && m != NULL)
|
|
|
|
|
flags &= ~M_PKTHDR;
|
|
|
|
|
|
|
|
|
|
/* Loop and append maximum sized mbufs to the chain tail. */
|
|
|
|
|
while (len > 0) {
|
|
|
|
|
if (len > MCLBYTES)
|
|
|
|
|
mb = m_getjcl(how, type, (flags & M_PKTHDR),
|
|
|
|
|
MJUMPAGESIZE);
|
|
|
|
|
else if (len >= MINCLSIZE)
|
|
|
|
|
mb = m_getcl(how, type, (flags & M_PKTHDR));
|
|
|
|
|
else if (flags & M_PKTHDR)
|
|
|
|
|
mb = m_gethdr(how, type);
|
|
|
|
|
else
|
|
|
|
|
mb = m_get(how, type);
|
|
|
|
|
|
|
|
|
|
/* Fail the whole operation if one mbuf can't be allocated. */
|
|
|
|
|
if (mb == NULL) {
|
|
|
|
|
if (nm != NULL)
|
|
|
|
|
m_freem(nm);
|
|
|
|
|
return (NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Book keeping. */
|
|
|
|
|
len -= M_SIZE(mb);
|
|
|
|
|
if (mtail != NULL)
|
|
|
|
|
mtail->m_next = mb;
|
|
|
|
|
else
|
|
|
|
|
nm = mb;
|
|
|
|
|
mtail = mb;
|
|
|
|
|
flags &= ~M_PKTHDR; /* Only valid on the first mbuf. */
|
|
|
|
|
}
|
|
|
|
|
if (flags & M_EOR)
|
|
|
|
|
mtail->m_flags |= M_EOR; /* Only valid on the last mbuf. */
|
|
|
|
|
|
|
|
|
|
/* If mbuf was supplied, append new chain to the end of it. */
|
|
|
|
|
if (m != NULL) {
|
|
|
|
|
for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
|
|
|
|
|
;
|
|
|
|
|
mtail->m_next = nm;
|
|
|
|
|
mtail->m_flags &= ~M_EOR;
|
|
|
|
|
} else
|
|
|
|
|
m = nm;
|
|
|
|
|
|
|
|
|
|
return (m);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*-
|
|
|
|
|
* Configure a provided mbuf to refer to the provided external storage
|
2016-03-01 00:17:14 +00:00
|
|
|
* buffer and setup a reference count for said buffer.
|
2016-02-11 21:32:23 +00:00
|
|
|
*
|
|
|
|
|
* Arguments:
|
|
|
|
|
* mb The existing mbuf to which to attach the provided buffer.
|
|
|
|
|
* buf The address of the provided external storage buffer.
|
|
|
|
|
* size The size of the provided buffer.
|
|
|
|
|
* freef A pointer to a routine that is responsible for freeing the
|
|
|
|
|
* provided external storage buffer.
|
|
|
|
|
* args A pointer to an argument structure (of any type) to be passed
|
|
|
|
|
* to the provided freef routine (may be NULL).
|
|
|
|
|
* flags Any other flags to be passed to the provided mbuf.
|
|
|
|
|
* type The type that the external storage buffer should be
|
|
|
|
|
* labeled with.
|
|
|
|
|
*
|
|
|
|
|
* Returns:
|
|
|
|
|
* Nothing.
|
|
|
|
|
*/
|
2016-03-01 00:17:14 +00:00
|
|
|
void
|
2017-10-09 20:35:31 +00:00
|
|
|
m_extadd(struct mbuf *mb, char *buf, u_int size, m_ext_free_t freef,
|
|
|
|
|
void *arg1, void *arg2, int flags, int type)
|
2016-02-11 21:32:23 +00:00
|
|
|
{
|
|
|
|
|
|
2016-03-01 00:17:14 +00:00
|
|
|
KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
|
2016-02-11 21:32:23 +00:00
|
|
|
|
|
|
|
|
mb->m_flags |= (M_EXT | flags);
|
|
|
|
|
mb->m_ext.ext_buf = buf;
|
|
|
|
|
mb->m_data = mb->m_ext.ext_buf;
|
|
|
|
|
mb->m_ext.ext_size = size;
|
|
|
|
|
mb->m_ext.ext_free = freef;
|
|
|
|
|
mb->m_ext.ext_arg1 = arg1;
|
|
|
|
|
mb->m_ext.ext_arg2 = arg2;
|
|
|
|
|
mb->m_ext.ext_type = type;
|
|
|
|
|
|
2016-03-01 00:17:14 +00:00
|
|
|
if (type != EXT_EXTREF) {
|
|
|
|
|
mb->m_ext.ext_count = 1;
|
|
|
|
|
mb->m_ext.ext_flags = EXT_FLAG_EMBREF;
|
|
|
|
|
} else
|
|
|
|
|
mb->m_ext.ext_flags = 0;
|
2016-02-11 21:32:23 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Free an entire chain of mbufs and associated external buffers, if
|
|
|
|
|
* applicable.
|
|
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
m_freem(struct mbuf *mb)
|
|
|
|
|
{
|
|
|
|
|
|
2016-03-24 20:06:52 +00:00
|
|
|
MBUF_PROBE1(m__freem, mb);
|
2016-02-11 21:32:23 +00:00
|
|
|
while (mb != NULL)
|
|
|
|
|
mb = m_free(mb);
|
|
|
|
|
}
|
