machine will result in approximately a 4.2% loss of performance (buildworld)
and approximately a 5% reduction in power consumption (when idle). Add XXX
note on how to really make hlt work (send an IPI to wakeup HLTed cpus on
a thread-schedule event? Generate an interrupt somehow?).
running with interrupts disabled, other cpus locked down, and only
making a temporary local mapping that we immediately back out again.
Tested by: gallatin
-finstrument-functions instead of -mprofiler-epilogue. The former
works essentially the same as the latter but has a higher overhead
(about 22 more bytes per function for passing unused args to the
profiling functions).
Removed all traces of the IDENT Makefile variable, which had been
reduced to just a place for holding profiling's contribution to CFLAGS
(the IDENT that gives the kernel identity was renamed to KERN_IDENT).
page-zeroing code as well as from the general page-zeroing code and use a
lazy tlb page invalidation scheme based on a callback made at the end
of mi_switch.
A number of people came up with this idea at the same time so credit
belongs to Peter, John, and Jake as well.
Two-way SMP buildworld -j 5 tests (second run, after stabilization)
2282.76 real 2515.17 user 704.22 sys before peter's IPI commit
2266.69 real 2467.50 user 633.77 sys after peter's commit
2232.80 real 2468.99 user 615.89 sys after this commit
Reviewed by: peter, jhb
Approved by: peter
choosethread() in MI C code instead of doing it in in assembly in all the
various cpu_switch() functions. This fixes problems on ia64 and sparc64.
Reviewed by: julian, peter, benno
Tested on: i386, alpha, sparc64
- It actually works this time, honest!
- Fine grained TLB shootdowns for SMP on i386. IPI's are very expensive,
so try and optimize things where possible.
- Introduce ranged shootdowns that can be done as a single IPI.
- PG_G support for i386
- Specific-cpu targeted shootdowns. For example, there is no sense in
globally purging the TLB cache for where we are stealing a page from
the local unshared process on the local cpu. Use pm_active to track
this.
- Add some instrumentation for the tlb shootdown code.
- Rip out SMP code from <machine/cpufunc.h>
- Try and fix some very bogus PG_G and PG_PS interactions that were bad
enough to cause vm86 bios calls to break. vm86 depended on our existing
bugs and this was the cause of the VESA panics last time.
- Fix the silly one-line error that caused the 'panic: bad pte' last time.
- Fix a couple of other silly one-line errors that should have caused more
pain than they did.
Some more work is needed:
- pmap_{zero,copy}_page[_idle]. These can be done without IPI's if we
have a hook in cpu_switch.
- The IPI handlers need some cleanup. I have a bogus %ds load that can
be avoided.
- APTD handling is rather bogus and appears to be a large source of
global TLB IPI shootdowns for no really good reason.
I see speedups of between 1.5% and ~4% on buildworlds in a while 1 loop.
I expect to see a bigger difference when there is significant pageout
activity or the system otherwise has memory shortages.
I have backed out a few optimizations that I had been using over the last
few days in order to be a little more conservative. I'll revisit these
again over the next few days as the dust settles.
New option: DISABLE_PG_G - In case I missed something.
the default) is now the only method for i386.
Remove the paraphanalia that supported critmode. Remove td_critnest, clean
up the assembly, and clean up (mostly remove) the old junk from
cpu_critical_enter() and cpu_critical_exit().
ipl.s except doreti which really belongs in with the exceptions as it's
just the other side of the same coin. Will remove ipl.s in a separate commit.
Agreed by: several including bde@freebsd.org
threaded VM pagezero kthread outside of Giant. For some platforms, this
is really easy since it can just use the direct mapped region. For others,
IPI sending is involved or there are other issues, so grab Giant when
needed.
We still have preemption issues to deal with, but Alan Cox has an
interesting suggestion on how to minimize the problem on x86.
Use Luigi's hack for preserving the (lack of) priority.
Turn the idle zeroing back on since it can now actually do something useful
outside of Giant in many cases.
mappings from the page tables, which were mapped with PG_G! We could
reuse the page table entry for another mapping (pmap_mapdev) but it
would never have cleared any remaining PG_G TLB entries.
pmap_swapin_proc/pmap_swapout_proc functions from the MD pmap code
and use a single equivalent MI version. There are other cleanups
needed still.
While here, use the UMA zone hooks to keep a cache of preinitialized
proc structures handy, just like the thread system does. This eliminates
one dependency on 'struct proc' being persistent even after being freed.
There are some comments about things that can be factored out into
ctor/dtor functions if it is worth it. For now they are mostly just
doing statistics to get a feel of how it is working.
These functions are always called on new memory so they can
not already be set up, so don't bother testing for that.
(This was left over from before we used UMA (which is cool))
and function) with existing configuration choices. Arguably if
ALT_BREAK_TO_DEBUGGER was present, so should have been
BREAK_TO_DEBUGGER. Regardless, it broke the option sort order in
these kernel configuration files.
Requested by: bde
