introduce a modified allocation mechanism for mbufs and mbuf clusters; one
which can scale under SMP and which offers the possibility of resource
reclamation to be implemented in the future. Notable advantages:
o Reduce contention for SMP by offering per-CPU pools and locks.
o Better use of data cache due to per-CPU pools.
o Much less code cache pollution due to excessively large allocation macros.
o Framework for `grouping' objects from same page together so as to be able
to possibly free wired-down pages back to the system if they are no longer
needed by the network stacks.
Additional things changed with this addition:
- Moved some mbuf specific declarations and initializations from
sys/conf/param.c into mbuf-specific code where they belong.
- m_getclr() has been renamed to m_get_clrd() because the old name is really
confusing. m_getclr() HAS been preserved though and is defined to the new
name. No tree sweep has been done "to change the interface," as the old
name will continue to be supported and is not depracated. The change was
merely done because m_getclr() sounds too much like "m_get a cluster."
- TEMPORARILY disabled mbtypes statistics displaying in netstat(1) and
systat(1) (see TODO below).
- Fixed systat(1) to display number of "free mbufs" based on new per-CPU
stat structures.
- Fixed netstat(1) to display new per-CPU stats based on sysctl-exported
per-CPU stat structures. All infos are fetched via sysctl.
TODO (in order of priority):
- Re-enable mbtypes statistics in both netstat(1) and systat(1) after
introducing an SMP friendly way to collect the mbtypes stats under the
already introduced per-CPU locks (i.e. hopefully don't use atomic() - it
seems too costly for a mere stat update, especially when other locks are
already present).
- Optionally have systat(1) display not only "total free mbufs" but also
"total free mbufs per CPU pool."
- Fix minor length-fetching issues in netstat(1) related to recently
re-enabled option to read mbuf stats from a core file.
- Move reference counters at least for mbuf clusters into an unused portion
of the cluster itself, to save space and need to allocate a counter.
- Look into introducing resource freeing possibly from a kproc.
Reviewed by (in parts): jlemon, jake, silby, terry
Tested by: jlemon (Intel & Alpha), mjacob (Intel & Alpha)
Preliminary performance measurements: jlemon (and me, obviously)
URL: http://people.freebsd.org/~bmilekic/mb_alloc/
something: offset into the first mbuf of the target chain before copying
the source data over.
Make drivers using m_devget() with a first argument "data - ETHER_ALIGN"
to use the offset argument to pass ETHER_ALIGN in. The way it was previously
done is potentially dangerous if the source data was at the top of a page
and the offset caused the previous page to be copied (if the
previous page has not yet been appropriately mapped).
The old `offset' argument in m_devget() is not used anywhere (it's always
0) and dates back to ~1995 (and earlier?) when support for ethernet trailers
existed. With that support gone, it was merely collecting dust.
Tested on alpha by: jlemon
Partially submitted by: jlemon
Reviewed by: jlemon
MFC after: 3 weeks
This work was based on kame-20010528-freebsd43-snap.tgz and some
critical problem after the snap was out were fixed.
There are many many changes since last KAME merge.
TODO:
- The definitions of SADB_* in sys/net/pfkeyv2.h are still different
from RFC2407/IANA assignment because of binary compatibility
issue. It should be fixed under 5-CURRENT.
- ip6po_m member of struct ip6_pktopts is no longer used. But, it
is still there because of binary compatibility issue. It should
be removed under 5-CURRENT.
Reviewed by: itojun
Obtained from: KAME
MFC after: 3 weeks
around, use a common function for looking up and extracting the tunables
from the kernel environment. This saves duplicating the same function
over and over again. This way typically has an overhead of 8 bytes + the
path string, versus about 26 bytes + the path string.
other "system" header files.
Also help the deprecation of lockmgr.h by making it a sub-include of
sys/lock.h and removing sys/lockmgr.h form kernel .c files.
Sort sys/*.h includes where possible in affected files.
OK'ed by: bde (with reservations)
we also reserve _adequate_ space for the mb_map submap; i.e. we need
space for nmbclusters, nmbufs, _and_ nmbcnt. Furthermore, we need to
rounddown, and not roundup, so that we are consistent.
Pointed out by: bde
The mbuf and mcluster free lists now each "own" a condition variable,
m_starved.
- Clean up minor indentention issues in sys/mbuf.h caused by previous
commit.
Don't use atomic operations for the stats updating, instead protect
the counts with the mbuf mutex. Most twiddling of the stats was
done right before or after releasing a mutex. By doing this we
reduce the number of locked ops needed as well as allow a sysctl
to gain a consitant view of the entire stats structure.
In the future...
This will allow us to chain common mbuf operations that would
normally need to aquire/release 2 or 3 of the locks to build an
mbuf with a cluster or external data attached into a single op
requiring only one lock.
Simplify the per-cpu locks that are planned.
There's also some if (1) code that should check if the "how"
operation specifies blocking/non-blocking behavior, we _could_ make
it so that we hold onto the mutex through calls into kmem_alloc
when non-blocking requests are made, but for safety reasons we
currently drop and reaquire the mutex around the calls.
Also, note that calling kmem_alloc is rare and only happens during
a shortage so drop/re-getting the mutex will not be a common
occurance.
Remove some #define's that seemed to obfuscate the code to me.
Remove an extranious comment.
Remove an XXX, including mutex.h isn't a crime.
Reviewed by: bmilekic
MCLGET macros in order to avoid incrementing the drop count twice.
Otherwise, in some cases, we may increment m_drops once in m_mballoc()
for example, and increment it again in m_mballoc_wait() if the
wait fails.
- Make sure that m_mballoc() really doesn't allow over nmbufs mbufs to
be allocated from mb_map. In the case where nmbufs-reserved space is not
an exact multiple of PAGE_SIZE (which it should be, but anyway...), we
hold nmbufs as an absolute maximum which need not ever be reached.
- Clean up m_clalloc(); make it more consistent in the sense that the first
argument `ncl' really means "the number of clusters ensured to be allocated"
and not "the number of pages worth of clusters to be allocated," as was
previously the case. This also makes it consistent with m_mballoc() as well
as the comment that preceeds it.
Reviewed by: jlemon
This is useful when doing copies of packet where some leading
space has been preallocated to insert protocol headers.
Note that there are in fact almost no users of m_copypacket.
MFC candidate.
allocation, as required.
If m_getm() receives NULL as a first argument, then it allocates `len'
(second argument) bytes worth of mbufs + clusters and returns the chain
only if it was able to allocate everything.
If the first argument is non-NULL, then it should be an existing mbuf
chain (e.g. pre-allocated mbuf sitting on a ring, on some list, etc.) and
so it will allocate `len' bytes worth of clusters and mbufs, as needed,
and append them to the tail of the passed in chain, only if it was able
to allocate everything requested.
If allocation fails, only what was allocated by the routine will be freed,
and NULL will be returned.
Also, get rid of existing m_getm() in netncp code and replace calls to it
to calls to this new generic code.
Heavily Reviewed by: bp
and function argument declarations. Make sure that functions that are
supposed to return a pointer return NULL in case of failure. Don't cast
NULL. Finally, get rid of annoying `register' uses.
mtx_enter(lock, type) becomes:
mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks)
mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized)
similarily, for releasing a lock, we now have:
mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN.
We change the caller interface for the two different types of locks
because the semantics are entirely different for each case, and this
makes it explicitly clear and, at the same time, it rids us of the
extra `type' argument.
The enter->lock and exit->unlock change has been made with the idea
that we're "locking data" and not "entering locked code" in mind.
Further, remove all additional "flags" previously passed to the
lock acquire/release routines with the exception of two:
MTX_QUIET and MTX_NOSWITCH
The functionality of these flags is preserved and they can be passed
to the lock/unlock routines by calling the corresponding wrappers:
mtx_{lock, unlock}_flags(lock, flag(s)) and
mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN
locks, respectively.
Re-inline some lock acq/rel code; in the sleep lock case, we only
inline the _obtain_lock()s in order to ensure that the inlined code
fits into a cache line. In the spin lock case, we inline recursion and
actually only perform a function call if we need to spin. This change
has been made with the idea that we generally tend to avoid spin locks
and that also the spin locks that we do have and are heavily used
(i.e. sched_lock) do recurse, and therefore in an effort to reduce
function call overhead for some architectures (such as alpha), we
inline recursion for this case.
Create a new malloc type for the witness code and retire from using
the M_DEV type. The new type is called M_WITNESS and is only declared
if WITNESS is enabled.
Begin cleaning up some machdep/mutex.h code - specifically updated the
"optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN
and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently
need those.
Finally, caught up to the interface changes in all sys code.
Contributors: jake, jhb, jasone (in no particular order)
kmem_free() for now. Kmem_malloc() and kmem_free() now have appropriate
assertions in place, and these checks aren't feasible until more of the
networking code is locked down. Also, the extra assertions here should
already be caught by the WITNESS code as lock order violations should
mutex operations on Giant be reintroduced here later.
The counters are incremented when a thread goes to sleep and decremented
either when a thread is woken up by another thread or when the sleep
times out. There existed a race where the sleep count could be decremented
twice resulting in an eventual underflow.
Move the decrementing of the "counters" to the thread initiating the sleep
and thus remedy the problem.
allocation routines are being called safely. Since we drop our relevant
mbuf mutex and acquire Giant before we call kmem_malloc(), we have
to make sure that this does not pave the way for a fatal lock order
reversal. Check that either Giant is already held (in which case it's safe
to grab it again and recurse on it) or, if Giant is not held, that no
other locks are held before we try to acquire Giant.
Similarily, add a KASSERT valid in the WITNESS case in m_reclaim() to
nail callers who end up in m_reclaim() and hold a lock.
Pointed out by: jhb
m_reclaim() and re-acquire it when m_reclaim() returns. This means that
we now call the drain routines without holding the mutex lock and
recursing into it. This was done for mainly two reasons:
(i) Avoid the long recursion; long recursions are typically bad and this
is the case here because we block all other code from freeing mbufs
if they need to. Doing that is kind of counter-productive, since we're
really hoping that someone will free.
(ii) More importantly, avoid a potential lock order reversal. Right now,
not all the locks have been added to our networking code; but
without this change, we're introducing the possibility for deadlock.
Consider for example ip_drain(). We will likely eventually introduce
a lock for ipq there, and so ip_freef() will be called with ipq lock
held. But, ip_freef() calls m_freem() which in turn acquires the
mmbfree lock. Since we were previously calling ip_drain() with mmbfree
held, our lock order would be: mmbfree->ipq->mmbfree. Some other code
may very well lock ipq first and then call ip_freef(). This would
result in the regular lock order, ipq->mmbfree. Clearly, we have
deadlock if one thread acquires the ipq lock and sits waiting for
mmbfree while another thread calling m_reclaim() acquires mmbfree
and sits waiting for the ipq lock.
Also, make sure to add a comment above m_reclaim()'s definition briefly
explaining this. Also document this above the call to m_reclaim() in
m_mballoc_wait().
Suggested and reviewed by: alfred
This is because calls with M_WAIT (now M_TRYWAIT) may not wait
forever when nothing is available for allocation, and may end up
returning NULL. Hopefully we now communicate more of the right thing
to developers and make it very clear that it's necessary to check whether
calls with M_(TRY)WAIT also resulted in a failed allocation.
M_TRYWAIT basically means "try harder, block if necessary, but don't
necessarily wait forever." The time spent blocking is tunable with
the kern.ipc.mbuf_wait sysctl.
M_WAIT is now deprecated but still defined for the next little while.
* Fix a typo in a comment in mbuf.h
* Fix some code that was actually passing the mbuf subsystem's M_WAIT to
malloc(). Made it pass M_WAITOK instead. If we were ever to redefine the
value of the M_WAIT flag, this could have became a big problem.
number of ext_buf counters that are possibly allocatable.
Do this because:
(i) It will make it easier to influence EXT_COUNTERS for if_sk,
if_ti (or similar) users where the driver allocates its own
ext_bufs and where it is important for the mbuf system to take
it into account when reserving necessary space for counters.
(ii) Facilitate some percentile calculation for netstat(1)
to accomodate the changes.
Here's a list of things that have changed (I may have left out a few); for a
relatively complete list, see http://people.freebsd.org/~bmilekic/mtx_journal
* Remove old (once useful) mcluster code for MCLBYTES > PAGE_SIZE which
nobody uses anymore. It was great while it lasted, but now we're moving
onto bigger and better things (Approved by: wollman).
* Practically re-wrote the allocation macros in sys/sys/mbuf.h to accomodate
new allocations which grab the necessary lock.
* Make sure that necessary mbstat variables are manipulated with
corresponding atomic() routines.
* Changed the "wait" routines, cleaned it up, made one routine that does
the job.
* Generalized MWAKEUP() macro. Got rid of m_retry and m_retryhdr, as they
are now included in the generalized "wait" routines.
* Sleep routines now use msleep().
* Free lists have locks.
* etc... probably other stuff I'm missing...
Things to look out for and work on later:
* find a better way to (dynamically) adjust EXT_COUNTERS
* move necessity to recurse on a lock from drain routines by providing
lock-free lower-level version of MFREE() (and possibly m_free()?).
* checkout include of mutex.h in sys/sys/mbuf.h - probably violating
general philosophy here.
The code has been reviewed quite a bit, but problems may arise... please,
don't panic! Send me Emails: bmilekic@freebsd.org
Reviewed by: jlemon, cp, alfred, others?
that should be better.
The old code counted references to mbuf clusters by using the offset
of the cluster from the start of memory allocated for mbufs and
clusters as an index into an array of chars, which did the reference
counting. If the external storage was not a cluster then reference
counting had to be done by the code using that external storage.
NetBSD's system of linked lists of mbufs was cosidered, but Alfred
felt it would have locking issues when the kernel was made more
SMP friendly.
The system implimented uses a pool of unions to track external
storage. The union contains an int for counting the references and
a pointer for forming a free list. The reference counts are
incremented and decremented atomically and so should be SMP friendly.
This system can track reference counts for any sort of external
storage.
Access to the reference counting stuff is now through macros defined
in mbuf.h, so it should be easier to make changes to the system in
the future.
The possibility of storing the reference count in one of the
referencing mbufs was considered, but was rejected 'cos it would
often leave extra mbufs allocated. Storing the reference count in
the cluster was also considered, but because the external storage
may not be a cluster this isn't an option.
The size of the pool of reference counters is available in the
stats provided by "netstat -m".
PR: 19866
Submitted by: Bosko Milekic <bmilekic@dsuper.net>
Reviewed by: alfred (glanced at by others on -net)
The variables "m_mclalloc_wid" and "m_mballoc_wid" were not in the
proper place. They should have been in uipc_mbuf.c and have been global,
not in mbuf.h and local per each file that uses mbuf.h.
Sorta bug fix:
In mbuf.h, the definitions of various things for KERNEL and not
KERNEL cases were very screwy. This fixes all of that which I could
find.
means that running out of mbuf space isn't a panic anymore, and code
which runs out of network memory will sleep to wait for it.
Submitted by: Bosko Milekic <bmilekic@dsuper.net>
Reviewed by: green, wollman
because in the case of mbuf clusters they only increment the reference
count rather than actually copying the data.
Add comments to this effect, and add a new routine called m_dup() that
returns a real, writable copy of an mbuf chain.
This is preliminary work required for implementing 'ipfw tee'.
Reviewed by: julian
into uipc_mbuf.c. This reduces three sets of identical tunable code to
one set, and puts the initialisation with the mbuf code proper.
Make NMBUFs tunable as well.
Move the nmbclusters sysctl here as well.
Move the initialisation of maxsockets from param.c to uipc_socket2.c,
next to its corresponding sysctl.
Use the new tunable macros for the kern.vm.kmem.size tunable (this should have
been in a separate commit, whoops).
SYSINIT_KT() etc (which is a static, compile-time procedure), use a
NetBSD-style kthread_create() interface. kproc_start is still available
as a SYSINIT() hook. This allowed simplification of chunks of the
sysinit code in the process. This kthread_create() is our old kproc_start
internals, with the SYSINIT_KT fork hooks grafted in and tweaked to work
the same as the NetBSD one.
One thing I'd like to do shortly is get rid of nfsiod as a user initiated
process. It makes sense for the nfs client code to create them on the
fly as needed up to a user settable limit. This means that nfsiod
doesn't need to be in /sbin and is always "available". This is a fair bit
easier to do outside of the SYSINIT_KT() framework.
This makes it possible to change the sysctl tree at runtime.
* Change KLD to find and register any sysctl nodes contained in the loaded
file and to unregister them when the file is unloaded.
Reviewed by: Archie Cobbs <archie@whistle.com>,
Peter Wemm <peter@netplex.com.au> (well they looked at it anyway)