improvements via the new routines pmap_qenter/pmap_qremove and pmap_kenter/
pmap_kremove. These routine allow fast mapping of pages for those
architectures that have "normal" MMUs. Also included is a fix to the
pageout daemon to properly check a queue end condition.
Submitted by: John Dyson
me:
1) TLB flush optimization that effectively eliminates half of all of the
TLB flushes. This works by only flushing the TLB when a page is "present"
in memory (i.e. the valid bit is set in the page table entry). See section
5.3.5 of the Intel 386 Programmer's Reference Manual.
2) The handling of "CMAP" has been improved to catch attempts at multiple
simultaneous use.
John:
1) Added pmap_qenter/pmap_qremove functions for fast mapping of pages into
the kernel. This is for future optimizations and support for the upcoming
merged VM/buffer cache.
Reviewed by: John Dyson
From Bruce Evans:
fu[i]byte() checked the wrong register. This caused interesting behaviour
in the GPL math emulator. The emulator does not check the values returned
by fu*() or su*() (:-() and it interpreted the address of -12(%ebp) as
-1(%ebp). The same probably occurs for all signed 8-bit offsets from
registers.
I cleaned up the new bzero() a bit.
Vastly improved trap.c from me. This rewritten version has a variety of
features, amoung them: higher performance and much higher code quality.
support.s, cpufunc.h:
No longer use gs override to enforce range limits - compare directly
against VM_MAXUSER_ADDRESS instead. The old way caused problems in
preserving the gs selector...and this method is just as fast or faster.
the NTP kernel PLL is disabled, and acquire_timer0() is enabled, thus
opening the door for microtime() (and hence gettimeofday()) to return
bogus timestamps. This option is necessary for the `pca' driver to
work, but is implemented to underscore the fact that accurate timekeeping
and the `pca' driver are incompatible at present. If someone writes a version
of microtime() that works when the `pca' driver is being used, this can get
junked.
1) check va before clearing the page clean flag. Not doing so was
causing the vnode pager error 5 messages when paging from
NFS. (pmap.c)
2) put back interrupt protection in idle_loop. Bruce didn't think
it was necessary, John insists that it is (and I agree). (swtch.s)
3) various improvements to the clustering code (vm_machdep.c). It's
now enabled/used by default.
4) bad disk blocks are now handled properly when doing clustered IOs.
(wd.c, vm_machdep.c)
5) bogus bad block handling fixed in wd.c.
6) algorithm improvements to the pageout/pagescan daemons. It's amazing
how well 4MB machines work now.
1) Removed all instances of disable_intr()/enable_intr() and changed
them back to splimp/splx. The previous method was done to improve
the performance, but Bruces recent changes to inline spl* have
made this unnecessary.
2) Cleaned up vm_machdep.c considerably. Probably fixed a few bugs, too.
3) Added a new mechanism for collecting page statistics - now done by
a new system process "pagescan". Previously this was done by the
pageout daemon, but this proved to be impractical.
4) Improved the page usage statistics gathering mechanism - performance is
much improved in small memory machines.
5) Modified mbuf.h to enable the support for an external free routine when
using mbuf clusters. Added appropriate glue in various places to
allow this to work.
6) Adapted a suggested change to the NFS code from Yuval Yurom to take
advantage of #5.
7) Added fault/swap statistics support.
1) fixed some bugs related to the bounce buffer code
2) vnode pager now supports clustered pageouts
3) experimental code for clustering all I/O via a new "cldisksort"
4) added >16MB check to Bustek driver
5) made some experimental algorithmic changes to the pageout daemon
6) fixed bugs in truncating mapped files (esp when mapped via NFS)
7) reorganized vnode pager I/O code
list of changes, I've made the following additional changes:
1) i386/include/ipl.h renamed to spl.h as the name conflicts with the
file of the same name in i386/isa/ipl.h.
2) changed all use of *mask (i.e. netmask, biomask, ttymask, etc) to
*_imask (net_imask, etc).
3) changed vestige of splnet use in if_is to splimp.
4) got rid of "impmask" completely (Bruce had gotten rid of netmask),
and are now using net_imask instead.
5) dozens of minor cruft to glue in Bruce's changes.
These require changes I made to config(8) as well, and thus it must
be rebuilt.
-DG
from Bruce Evans:
sio:
o No diff is supplied. Remove the define of setsofttty(). I hope
that is enough.
*.s:
o i386/isa/debug.h no longer exists. The event counters became too
much trouble to maintain. All function call entry and exception
entry counters can be recovered by using profiling kernel (the new
profiling supports all entry points; however, it is too slow to
leave enabled all the time; it also). Only BDBTRAP() from debug.h
is now used. That is moved to exception.s. It might be worth
preserving SHOW_BITS() and calling it from _mcount() (if enabled).
o T_ASTFLT is now only set just before calling trap().
o All exception handlers set SWI_AST_MASK in cpl as soon as possible
after entry and arrange for _doreti to restore it atomically with
exiting. It is not possible to set it atomically with entering
the kernel, so it must be checked against the user mode bits in
the trap frame before committing to using it. There is no place
to store the old value of cpl for syscalls or traps, so there are
some complications restoring it.
Profiling stuff (mostly in *.s):
o Changes to kern/subr_mcount.c, gcc and gprof are not supplied yet.
o All interesting labels `foo' are renamed `_foo' and all
uninteresting labels `_bar' are renamed `bar'. A small change
to gprof allows ignoring labels not starting with underscores.
o MCOUNT_LABEL() is to provide names for counters for times spent
in exception handlers.
o FAKE_MCOUNT() is a version of MCOUNT() suitable for exception
handlers. Its arg is the pc where the exception occurred. The
new mcount() pretends that this was a call from that pc to a
suitable MCOUNT_LABEL().
o MEXITCOUNT is to turn off any timer started by MCOUNT().
/usr/src/sys/i386/i386/exception.s:
o The non-BDB BPTTRAP() macros were doing a sti even when interrupts
were disabled when the trap occurred. The sti (fixed) sti is
actually a no-op unless you have my changes to machdep.c that make
the debugger trap gates interrupt gates, but fixing that would
make the ifdefs messier. ddb seems to be unharmed by both
interrupts always disabled and always enabled (I had the branch in
the fix back to front for some time :-().
o There is no known pushal bug.
o tf_err can be left as garbage for syscalls.
/usr/src/sys/i386/i386/locore.s:
o Fix and update BDE_DEBUGGER support.
o ENTRY(btext) before initialization was dangerous.
o Warm boot shot was longer than intended.
/usr/src/sys/i386/i386/machdep.c:
o DON'T APPLY ALL OF THIS DIFF. It's what I'm using, but may require
other changes.
Use the following:
o Remove aston() and setsoftclock().
Maybe use the following:
o No netisr.h.
o Spelling fix.
o Delay to read the Rebooting message.
o Fix for vm system unmapping a reduced area of memory
after bounds_check_with_label() reduces the size of
a physical i/o for a partition boundary. A similar
fix is required in kern_physio.c.
o Correct use of __CONCAT. It never worked here for non-
ANSI cpp's. Is it time to drop support for non-ANSI?
o gdt_segs init. 0xffffffffUL is bogus because ssd_limit
is not 32 bits. The replacement may have the same
value :-), but is more natural.
o physmem was one page too low. Confusing variable names.
Don't use the following:
o Better numbers of buffers. Each 8K page requires up to
16 buffer headers. On my system, this results in 5576
buffers containing [up to] 2854912 bytes of memory.
The usual allocation of about 384 buffers only holds
192K of disk if you use it on an fs with a block size
of 512.
o gdt changes for bdb.
o *TGT -> *IDT changes for bdb.
o #ifdefed changes for bdb.
/usr/src/sys/i386/i386/microtime.s:
o Use the correct asm macros. I think asm.h was copied from Mach
just for microtime and isn't used now. It certainly doesn't
belong in <sys>. Various macros are also duplicated in
sys/i386/boot.h and libc/i386/*.h.
o Don't switch to and from the IRR; it is guaranteed to be selected
(default after ICU init and explicitly selected in isa.c too, and
never changed until the old microtime clobbered it).
/usr/src/sys/i386/i386/support.s:
o Non-essential changes (none related to spls or profiling).
o Removed slow loads of %gs again. The LDT support may require
not relying on %gs, but loading it is not the way to fix it!
Some places (copyin ...) forgot to load it. Loading it clobbers
the user %gs. trap() still loads it after certain types of
faults so that fuword() etc can rely on it without loading it
explicitly. Exception handlers don't restore it. If we want
to preserve the user %gs, then the fastest method is to not
touch it except for context switches. Comparing with
VM_MAXUSER_ADDRESS and branching takes only 2 or 4 cycles on
a 486, while loading %gs takes 9 cycles and using it takes
another.
o Fixed a signed branch to unsigned.
/usr/src/sys/i386/i386/swtch.s:
o Move spl0() outside of idle loop.
o Remove cli/sti from idle loop. sw1 does a cli, and in the
unlikely event of an interrupt occurring and whichqs becoming
zero, sw1 will just jump back to _idle.
o There's no spl0() function in asm any more, so use splz().
o swtch() doesn't need to be superaligned, at least with the
new mcounting.
o Fixed a signed branch to unsigned.
o Removed astoff().
/usr/src/sys/i386/i386/trap.c:
o The decentralized extern decls were inconsistent, of course.
o Fixed typo MATH_EMULTATE in comments. */
o Removed unused variables.
o Old netmask is now impmask; print it instead. Perhaps we
should print some of the new masks.
o BTW, trap() should not print anything for normal debugger
traps.
/usr/src/sys/i386/include/asmacros.h:
o DON'T APPLY ALL OF THIS DIFF. Just use some of the null macros
as necessary.
/usr/src/sys/i386/include/cpu.h:
o CLKF_BASEPRI() changes since cpl == SWI_AST_MASK is now normal
while the kernel is running.
o Don't use var++ to set boolean variables. It fails after a mere
4G times :-) and is slower than storing a constant on [3-4]86s.
/usr/src/sys/i386/include/cpufunc.h:
o DON'T APPLY ALL OF THIS DIFF. You need mainly the include of
<machine/ipl.h>. Unfortunately, <machine/ipl.h> is needed by
almost everything for the inlines.
/usr/src/sys/i386/include/ipl.h:
o New file. Defines spl inlines and SWI macros and declares most
variables related to hard and soft interrupt masks.
/usr/src/sys/i386/isa/icu.h:
o Moved definitions to <machine/ipl.h>
/usr/src/sys/i386/isa/icu.s:
o Software interrupts (SWIs) and delayed hardware interrupts (HWIs)
are now handled uniformally, and dispatching them from splx() is
more like dispatching them from _doreti. The dispatcher is
essentially *(handler[ffs(ipending & ~cpl)]().
o More care (not quite enough) is taken to avoid unbounded nesting
of interrupts.
o The interface to softclock() is changed so that a trap frame is
not required.
o Fast interrupt handlers are now handled more uniformally.
Configuration is still too early (new handlers would require
bits in <machine/ipl.h> and functions to vector.s).
o splnnn() and splx() are no longer here; they are inline functions
(could be macros for other compilers). splz() is the nontrivial
part of the old splx().
/usr/src/sys/i386/isa/ipl.h
o New file. Supposed to have only bus-dependent stuff. Perhaps
the h/w masks should be declared here.
/usr/src/sys/i386/isa/isa.c:
o DON'T APPLY ALL OF THIS DIFF. You need only things involving
*mask and *MASK and comments about them. netmask is now a pure
software mask. It works like the softclock mask.
/usr/src/sys/i386/isa/vector.s:
o Reorganize AUTO_EOI* macros.
o Option FAST_INTR_HANDLER_USERS_ES for people who don't trust
fastintr handlers.
o fastintr handlers need to metamorphose into ordinary interrupt
handlers if their SWI bit has become set. Previously, sio had
unintended latency for handling output completions and input
of SLIP framing characters because this was not done.
/usr/src/sys/net/netisr.h:
o The machine-dependent stuff is now imported from <machine/ipl.h>.
/usr/src/sys/sys/systm.h
o DON'T APPLY ALL OF THIS DIFF. You need mainly the different
splx() prototype. The spl*() prototypes are duplicated as
inlines in <machine/ipl.h> but they need to be duplicated here
in case there are no inlines. I sent systm.h and cpufunc.h
to Garrett. We agree that spl0 should be replaced by splnone
and not the other way around like I've done.
/usr/src/sys/kern/kern_clock.c
o splsoftclock() now lowers cpl so the direct call to softclock()
works as intended.
o softclock() interface changed to avoid passing the whole frame
(some machines may need another change for profile_tick()).
o profiling renamed _profiling to avoid ANSI namespace pollution.
(I had to improve the mcount() interface and may as well fix it.)
The GUPROF variant doesn't actually reference profiling here,
but the 'U' in GUPROF should mean to select the microtimer
mcount() and not change the interface.
1) A new mechanism has been added to prevent pages from being paged
out called "vm_page_hold". Similar to vm_page_wire, but
much lower overhead.
2) Scheduling algorithm has been changed to improve interactive
performance.
3) Paging algorithm improved.
4) Some vnode and swap pager bugs fixed.
Eliminates vm_fault overhead on process startup and
mmap referenced data for in-memory pages.
(process startup time using in-memory segments *much* faster)
2) Even more efficient pmap code. Code partially cleaned up.
More comments yet to follow.
(generally more efficient pte management)
3) Pageout clustering ( in addition to the FreeBSD V1.1 pagein
clustering.)
(much faster paging performance on non-write behind disk
subsystems, slightly faster performance on other systems.)
4) Slightly changed vm_pageout code for more efficiency and
better statistics. Also, resist swapout a little more.
(less likely to pageout a recently used page)
5) Slight improvement to the page table page trap efficiency.
(generally faster system VM fault performance)
6) Defer creation of unnamed anonymous regions pager until needed.
(speeds up shared memory bss creation)
7) Remove possible deadlock from swap_pager initialization.
8) Enhanced procfs to provide "vminfo" about vm objects and user
pmaps.
9) Increased MCLSHIFT/MCLBYTES from 2K to 4K to improve net &
socket performance and to prepare for things to come.
John Dyson
dyson@implode.root.com
David Greenman
davidg@root.com
