This updates the CFQ io scheduler to the new time sliced design (cfq
v3). It provides full process fairness, while giving excellent
aggregate system throughput even for many competing processes. It
supports io priorities, either inherited from the cpu nice value or set
directly with the ioprio_get/set syscalls. The latter closely mimic
set/getpriority.
This import is based on my latest from -mm.
Signed-off-by: Jens Axboe <axboe@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This is the core of the (much simplified) early reclaim. The goal of this
patch is to reclaim some easily-freed pages from a zone before falling back
onto another zone.
One of the major uses of this is NUMA machines. With the default allocator
behavior the allocator would look for memory in another zone, which might be
off-node, before trying to reclaim from the current zone.
This adds a zone tuneable to enable early zone reclaim. It is selected on a
per-zone basis and is turned on/off via syscall.
Adding some extra throttling on the reclaim was also required (patch
4/4). Without the machine would grind to a crawl when doing a "make -j"
kernel build. Even with this patch the System Time is higher on
average, but it seems tolerable. Here are some numbers for kernbench
runs on a 2-node, 4cpu, 8Gig RAM Altix in the "make -j" run:
wall user sys %cpu ctx sw. sleeps
---- ---- --- ---- ------ ------
No patch 1009 1384 847 258 298170 504402
w/patch, no reclaim 880 1376 667 288 254064 396745
w/patch & reclaim 1079 1385 926 252 291625 548873
These numbers are the average of 2 runs of 3 "make -j" runs done right
after system boot. Run-to-run variability for "make -j" is huge, so
these numbers aren't terribly useful except to seee that with reclaim
the benchmark still finishes in a reasonable amount of time.
I also looked at the NUMA hit/miss stats for the "make -j" runs and the
reclaim doesn't make any difference when the machine is thrashing away.
Doing a "make -j8" on a single node that is filled with page cache pages
takes 700 seconds with reclaim turned on and 735 seconds without reclaim
(due to remote memory accesses).
The simple zone_reclaim syscall program is at
http://www.bork.org/~mort/sgi/zone_reclaim.c
Signed-off-by: Martin Hicks <mort@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Some time ago, GAS was fixed to bring the .spillpsp directive in line
with the Intel assembler manual (there was some disagreement as to
whether or not there is a built-in 16-byte offset). Unfortunately,
there are two places in the kernel where this directive is used in
handwritten assembly files and those of course relied on the "buggy"
behavior. As a result, when using a "fixed" assembler, the kernel
picks up the UNaT bits from the wrong place (off by 16) and randomly
sets NaT bits on the scratch registers. This can be noticed easily by
looking at a coredump and finding various scratch registers with
unexpected NaT values. The patch below fixes this by using the
.spillsp directive instead, which works correctly no matter what
assembler is in use.
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Patch below fixes 3 trivial typos which are caught by the new
assembler (v2.169.90). Please apply.
[Note: fix to memcpy that was also part of this patch was separately
applied from patches by H.J. and Andreas ... so the delta here only
has the other two fixes. -Tony]
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
This patch switches the srlz.i in ia64_leave_kernel() to srlz.d. As
per architecture manual, the former is needed only to ensure that the
clearing of PSR.IC is seen by the VHPT for subsequent instruction
fetches. However, since the remainder of the code (up to and
including the RFI instruction) is mapped by a pinned TLB entry, there
is no chance of an iTLB miss and we don't care whether or not the VHPT
sees PSR.IC cleared. Since srlz.d is substantially cheaper than
srlz.i, this should shave off a few cycles off the interrupt path
(unverified though; I'm not setup to measure this at the moment).
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Reschedule code to read ar.bsp as early as possible. To enable this,
don't bother clearing some of the registers when we're returning to
kernel stacks. Also, instead of trying to support the pNonSys case
(which makes no sense), do a bugcheck instead (with break 0). Finally,
remove a clear of r14 which is a left-over from the previous patch.
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Why is this a good idea? Clearing b7 to 0 is guaranteed to do us no
good and writing it with __kernel_syscall_via_epc() yields a 6 cycle
improvement _if_ the application performs another EPC-based system-
call without overwriting b7, which is not all that uncommon. Well
worth the minimal cost of 1 bundle of code.
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Decreases syscall overhead by approximately 6 cycles.
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
This by itself is good for a 1-2 cycle speed up. Effect is bigger
when combined with the later patches.
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Sadly, I goofed in this syscall-tuning patch:
ChangeSet 1.1966.1.40 2005/01/22 13:31:05 davidm@hpl.hp.com
[IA64] Improve ia64_leave_syscall() for McKinley-type cores.
Optimize ia64_leave_syscall() a bit better for McKinley-type cores.
The patch looks big, but that's mostly due to renaming r16/r17 to r2/r3.
Good for a 13 cycle improvement.
The problem is that the size of the physical stacked registers was
loaded into the wrong register (r3 instead of r17). Since r17 by
coincidence always had the value 1, this had the effect of turning
rse_clear_invalid into a no-op. That poses the risk of leaking kernel
state back to user-land and is hence not acceptable.
The fix below is simple, but unfortunately it costs us about 28 cycles
in syscall overhead. ;-(
Unfortunately, there isn't much we can do about that since those
registers have to be cleared one way or another.
--david
Signed-off-by: Tony Luck <tony.luck@intel.com>
Recently I noticed that clearing ar.ssd/ar.csd right before srlz.d is
causing significant stalling in the syscall path. The patch below
fixes that by moving the register-writes after srlz.d. On a Madison,
this drops break-based getpid() from 241 to 226 cycles (-15 cycles).
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
Jens Axboe