instead of ephemeral mappings using pmap_qenter() by the writer. The
writer is still, however, responsible for wiring the pages, just not
mapping them. Consequently, the allocation of KVA for the direct case is
unnecessary. Remove it and the sysctls limiting it, i.e.,
kern.ipc.maxpipekvawired and kern.ipc.amountpipekvawired. The number
of temporarily wired pages is still, however, limited by
kern.ipc.maxpipekva.
Note: On platforms lacking a direct virtual-to-physical mapping,
uiomove_fromphys() uses sf_bufs to cache ephemeral mappings. Thus,
the number of available sf_bufs can influence the performance of pipes
on platforms such i386. Surprisingly, I saw the greatest gain from this
change on such a machine: lmbench's pipe bandwidth result increased from
~1050MB/s to ~1850MB/s on my 2.4GHz, 400MHz FSB P4 Xeon.
From alc:
Move pageable pipe memory to a seperate kernel submap to avoid awkward
vm map interlocking issues. (Bad explanation provided by me.)
From me:
Rework pipespace accounting code to handle this new layout, and adjust
our default values to account for the fact that we now have a solid
limit on allocations.
Also, remove the "maxpipes" limit, as it no longer has a purpose.
(The limit on kva usage solves the problem of having two many pipes.)
immediately after the kernel map has been sized, and is
the optimal place for the autosizing of memory allocations
which occur within the kernel map to occur.
Suggested by: bde
than the shortcircuited version I had been using, which only worked
properly on i386 & amd64.
Also, change an autoscale constant to account for the more correct
kmem_map size.
Problem noticed by: mux
- Limit the total number of pipes so that we do not
exhaust all vm objects in the kernel map. When
this limit is reached, a ratelimited message will
be printed to the console.
- Put a soft limit on the amount of memory consumable
by pipes. Once the limit has been reached, all new
pipes will be limited to 4K in size, rather than the
default of 16K.
- Put a limit on the number of pages that may be used
for high speed page flipping in order to reduce the
amount of wired memory. Pipe writes that occur
while this limit is exceeded will fall back to
non-page flipping mode.
The above values are auto-tuned in subr_param.c and
are scaled to take into account both the size of
physical memory and the size of the kernel map.
These limits help to reduce the "kernel resources exhausted"
panics that could be caused by opening a large
number of pipes. (Pipes alone are no longer able
to exhaust all resources, but other kernel memory hogs
in league with pipes may still be able to do so.)
PR: 53627
Ideas / comments from: hsu, tjr, dillon@apollo.backplane.com
MFC after: 1 week
in the original hardwired sysctl implementation.
The buf size calculator still overflows an integer on machines with large
KVA (eg: ia64) where the number of pages does not fit into an int. Use
'long' there.
Change Maxmem and physmem and related variables to 'long', mostly for
completeness. Machines are not likely to overflow 'int' pages in the
near term, but then again, 640K ought to be enough for anybody. This
comes for free on 32 bit machines, so why not?
Apply the change as a continuous slew rather than as a series of
discrete steps and make it possible to adjust arbitraryly huge
amounts of time in either direction.
In practice this is done by hooking into the same once-per-second
loop as the NTP PLL and setting a suitable frequency offset deducting
the amount slewed from the remainder. If the remaining delta is
larger than 1 second we slew at 5000PPM (5msec/sec), for a delta
less than a second we slew at 500PPM (500usec/sec) and for the last
one second period we will slew at whatever rate (less than 500PPM)
it takes to eliminate the delta entirely.
The old implementation stepped the clock a number of microseconds
every HZ to acheive the same effect, using the same rates of change.
Eliminate the global variables tickadj, tickdelta and timedelta and
their various use and initializations.
This removes the most significant obstacle to running timecounter and
NTP housekeeping from a timeout rather than hardclock.
to exhaust all kmaps. The only reward for setting maxproc
to a value which will cause kmap exhaustion is a panic
during a forkbomb attack.
MFC after: 3 days
from 1 megabyte of ram per user to 2 megabytes of ram per user, and
reduce the cap from 512 to 384. 512 leaves around 240 MB of KVM available
while 384 leaves 270 MB of KVM available. Available KVM is important
in order to deal with zalloc and kernel malloc area growth.
Reviewed by: mckusick
MFC: either before 4.5 if re's agree, or after 4.5
information. The default limits only effect machines with > 1GB of ram
and can be overriden with two new kernel conf variables VM_SWZONE_SIZE_MAX
and VM_BCACHE_SIZE_MAX, or with loader variables kern.maxswzone and
kern.maxbcache. This has the effect of leaving more KVM available for
sizing NMBCLUSTERS and 'maxusers' and should avoid tripups where a sysad
adds memory to a machine and then sees the kernel panic on boot due to
running out of KVM.
Also change the default swap-meta auto-sizing calculation to allocate half
of what it was previously allocating. The prior defaults were way too high.
Note that we cannot afford to run out of swap-meta structures so we still
stay somewhat conservative here.
Tunables are now derived at boot time from maxusers. ie: change maxusers
via a tunable and all the derivative settings change. You can change
the other tunables individually as well. Even hz etc is tunable.
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/
and initialized during boot. This avoids bloating sizeof(struct lock).
As a side effect, it is no longer necessary to enforce the assumtion that
lockinit()/lockdestroy() calls are paired, so the LK_VALID flag has been
removed.
Idea taken from: BSD/OS.
Remove evil allocation macros from machdep.c (why was that there???) and
use malloc() instead.
Move paramters out of param.h and into the code itself.
Move a bunch of internal definitions from public sys/*.h headers (without
#ifdef _KERNEL even) into the code itself.
I had hoped to make some of this more dynamic, but the cost of doing
wakeups on all sleeping processes on old arrays was too frightening.
The other possibility is to initialize on the first use, and allow
dynamic sysctl changes to parameters right until that point. That would
allow /etc/rc.sysctl to change SEM* and MSG* defaults as we presently
do with SHM*, but without the nightmare of changing a running system.
via sysctl. It's done pretty simply but it should be quite adequate.
Also move SHMMAXPGS from $machine/include/vmparam.h as the comments that
went with it were wrong... we don't allocate KVM space for the pages so
that comment is bogus.. The only practical limit is how much physical
ram you want to lock up as this stuff isn't paged out or swap backed.
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
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).
file to a stream socket. sendfile(2) is similar to implementations in
HP-UX, Linux, and other systems, but the API is more extensive and
addresses many of the complaints that the Apache Group and others have
had with those other implementations. Thanks to Marc Slemko of the
Apache Group for helping me work out the best API for this.
Anyway, this has the "net" result of speeding up sends of files over
TCP/IP sockets by about 10X (that is to say, uses 1/10th of the CPU
cycles) when compared to a traditional read/write loop.
which makes adjtime(2) useless and confuses xntpd(8) into refusing
to start even when it would use the kernel PLL instead of adjtime().
The result is the same as recommended by tickadj(8), at least when
HZ divides 10^6. Of course, you wouldn't want to actually use
adjtime() when HZ is large. In the silly boundary case of HZ == 10^6,
tickadj == tick == 1 so the clock stops while adjtime() is active.
Define a parameter which indicates the maximum number of sockets in a
system, and use this to size the zone allocators used for sockets and
for certain PCBs.
Convert PF_LOCAL PCB structures to be type-stable and add a version number.
Define an external format for infomation about socket structures and use
it in several places.
Define a mechanism to get all PF_LOCAL and PF_INET PCB lists through
sysctl(3) without blocking network interrupts for an unreasonable
length of time. This probably still has some bugs and/or race
conditions, but it seems to work well enough on my machines.
It is now possible for `netstat' to get almost all of its information
via the sysctl(3) interface rather than reading kmem (changes to follow).
variable `kern.maxvnodes' which gives much better control over vnode
allocation than EXTRAVNODES (except in -current between 1995/10/28 and
1996/11/12, kern.maxvnodes was read-only and thus useless).
when allocating memory for network buffers at interrupt time. This is due
to inadequate checking for the new mcl_map. Fixed by merging mb_map and
mcl_map into a single mb_map.
Reviewed by: wollman
This will make a number of things easier in the future, as well as (finally!)
avoiding the Id-smashing problem which has plagued developers for so long.
Boy, I'm glad we're not using sup anymore. This update would have been
insane otherwise.
to be allocated at boot time. This is an expensive option, as they
consume physical ram and are not pageable etc. In certain situations,
this kind of option is quite useful, especially for news servers that
access a large number of directories at random and torture the name cache.
Defining 5000 or 10000 extra vnodes should cut down the amount of vnode
recycling somewhat, which should allow better name and directory caching
etc.
This is a "your mileage may vary" option, with no real indication of
what works best for your machine except trial and error. Too many will
cost you ram that you could otherwise use for disk buffers etc.
This is based on something John Dyson mentioned to me a while ago.
