To avoid a MAXMACH limit of 32 and make
txtflush into an array for the bitmap.
Provide portable macros for testing and clearing
the bits: needtxtflush(), donetxtflush().
On pc/pc64, define inittxtflush()/settxtflush()
as no-op macros, avoiding the storage overhead of
the txtflush array alltogether.
We can take advantage of the fact that xinit() allocates
kernel memory from conf.mem[] banks always at the beginning
of a bank, so the separate palloc.mem[] array can be eleminated
as we can calculate the amount of non-kernel memory like:
upages = cm->npage - (PGROUND(cm->klimit - cm->kbase)/BY2PG)
for the number of reserved kernel pages,
we provide the new function: ulong nkpages(Confmem*)
This eleminates the error case of running out of slots in
the array and avoids wasting memory in ports that have simple
memory configurations (compared to pc/pc64).
Previously, mmurelease() was always called with
palloc spinlock held.
This is unneccesary for some mmurelease()
implementations as they wont release pages
to the palloc pool.
This change removes pagechainhead() and
pagechaindone() and replaces them with just
freepages() call, which aquires the palloc
lock internally as needed.
freepages() avoids holding the palloc lock
while walking the linked list of pages,
avoding some lock contention.
do not touch s->map on SG_PHYSICAL type segments as they do
not have a pte map (s->mapsize == 0 && s->map == nil).
also remove the SG_PHYSICAL switch in freepte(), this is never
reached.
when reclaiming pages from an image, always reclaim all
the hash chains equally. that way, we avoid being biased
towards the chains at the start of the Image.pghash[] array.
images can be in two states: active or inactive. inactive
images are the ones which are not used by program while
active ones aare.
when reclaiming pages, we should try to reclaim pages
from inactive images first and only if that set becomes
exhausted attempt to release text pages and attempt to
reclaim pages from active images.
when we run out of Image structures, it makes only sense
to reclaim pages from inactive images, as reclaiming pages
from active ones will never free any Image structures.
change putimage() to require a image already locked and
make it unlock the image. this avoids many pointless
unlock()/lock() sequences as all callers of putimage()
already had the image locked.
imagereclaim(), pagereclaim():
- move imagereclaim() and pagereclaim() declarations to portfns.h
- consistently use ulong type for page counts
- name number of pages to free "pages" instead of "min"
- check for pages == 0 on entry
freepages():
- move pagechaindone() call to wakeup newpage() consumers inside
palloc critical section.
putimage():
- use long type for refcount
addresses va's of 0 and -BY2PG cause trouble with some memmove()/memset()
implementations and possibly other code because of the nil pointer
and end pointers wrapping to zero.
give kernel processes and local disk file servers (procs
having noswap flag set) a clear advantage for page allocation
under starved condition by giving them ther own wait queue so
they get readied as soon as pages become available.
the mount cache uses Page.va to store cached range offset and
limit, but mips kernel uses cache index bits from Page.va to
maintain page coloring. Page.va was not initialized by auxpage().
this change removes auxpage() which was primarily used only
by the mount cache and use newpage() with cache file offset
page as va so we will get a page of the right color.
mount cache keeps the index bits intact by only using the top
and buttom PGSHIFT bits of Page.va for the range offset/limit.
mcountseg(), mfreeseg():
use Pte.first/last pointers when possible and avoid constructs
like s->map[i]->pages[j].
freepte():
do not zero entries in freepte(), the segment is going away and
here is no point in zeroing page pointers. hoist common code at
the top avoiding duplication.
segpage(), fixfault():
avoid load after store for Pte** pointer.
fixfault():
return -1 in default case to avoid the "used but not set" warning
for mmuphys and get rid of the useless initialization.
syssegflush():
due to len being unsigned, the pe = PGROUND(pe) can make "chunk"
bigger than len causing a overflow. rewrite the function and deal
with page alignment and errors at the beginning.
syssegflush(), segpage(), fixfault(), putseg(), relocateseg(),
mcountseg(), mfreeseg():
keep naming consistent.
ignore physical segments in mcountseg() and mfreeseg(). physical
segments are not backed by user pages, and doing putpage() on
physical segment pages in mfreeseg() is an error.
do now allow physical segemnts to be resized. the segment size
is only checked in segattach() to be within the physical segment!
ignore physical segments in portcountpagerefs() as pagenumber()
does not work on the malloced page structures of a physical segment.
get rid of Physseg.pgalloc() and Physseg.pgfree() indirection as
this was never used and if theres a need to do more efficient
allocation, it should be done in a portable way.
there are no kernels currently that do page coloring,
so the only use of cachectl[] is flushing the icache
(on arm and ppc).
on pc64, cachectl consumes 32 bytes in each page resulting
in over 200 megabytes of overhead for 32gb of ram with 4K
pages.
this change removes cachectl[] and adds txtflush ulong
that is set to ~0 by pio() to instruct putmmu() to flush
the icache.
make the Page stucture less than half its original size by getting rid of
the Lock and the lru.
The Lock was required to coordinate the unchaining of pages that where
both cached and on the lru freelist.
now pages have a single next pointer that is used for palloc.head
freelist xor for page cache hash chains in Image.pghash[].
cached pages are not on the freelist anymore, but will be reclaimed
from images by the pager when the freelist runs out of pages.
each Image has its own 512 hash chains for cached page lookup. That is
2MB worth of pages and there should be no collisions for most text images.
page reclaiming can be done without holding palloc.lock as the Image is
the owner of the page hash chains protected by the Image's lock.
reclaiming Image structures can be done quickly by only reclaiming pages from
inactive images, that is images which are not currently in use by segments.
the Ref structure has no Lock anymore. Only a single long that is atomically
incremented or decremnted using cmpswap().
there are various other changes as a consequence code. and lots of pikeshedding,
sorry.
the palloc.pages array takes arround 5% of the upages which
gives us:
16GB = ~0.8GB
32GB = ~1.6GB
64GB = ~3.2GB
we only have 2GB of address space above KZERO so this will not
work for long.
instead, pageinit() was altered to accept a preallocated memory
in palloc.pages. and preallocpages() in pc64/main.c allocates the
in upages memory, mapping it in the VMAP area (which has 512GB).
the drawback is that we cannot poke at Page structures now from
/proc/n/mem as the VMAP area is not accessible from it.
simplifying paging code by getting rid of duppage(). instead,
fixfault() now always makes a copy of the shared/cached page
and leaves the cache alone. newpage() uncaches pages as
neccesary.
thanks charles forsyth for the suggestion.
from http://9fans.net/archive/2014/03/26:
> It isn't needed at all. When a cached page is written, it's trying hard to
> replace the page in the cache by a new copy,
> to return the previously cached page. Instead, I copy the cached page and
> return the copy, which is what it already
> does in another instance. ...
imagereclaim() sabotaged itself by breaking the invariant
that cached pages are kept at the end of the page list.
once we made a hole of uncached pages, we would stop
reclaiming cached pages before it as the loop breaks
once it hits a uncached page. (we iterate backwards from
the tail to the head of the pagelist until pages have been
reclaimed or we hit a uncached page).
the solution is to move pages to the head of the pagelist
after removing them from the image cache.
this change is in preparation for amd64. the systab calling
convention was also changed to return uintptr (as segattach
returns a pointer) and the arguments are now passed as
va_list which handles amd64 arguments properly (all arguments
are passed in 64bit quantities on the stack, tho the upper
part will not be initialized when the element is smaller
than 8 bytes).
this is partial. xalloc needs to be converted in the future.
the image cache should not hold onto the text file channel
when not neccesary. now, the image keeps track of the number
of page cache references in Image.pgref. if the number of
page cache references and Image.ref are equal, this means
all the references to this image are from the page cache.
so no segments are using this image. in that case, we can
close the channel, but keep the Image in the hash table.
when attachimage() finds our image, it will check if Image.c
is nil and reattach the channel to the image before it is
used.
the Image.nocache flag isnt needed anymore.
the image cache has the property of keeping a channel
for the executable binary arround which prevents the
mountpoint from going away.
this can easily be reproduced by running:
@{rfork n; ramfs; cp /bin/echo /tmp; /tmp/echo}
observe how ramfs stays arround until the image is
reclaimed. the echo binary is also cached but is
unreachable from any namespace.
we now restrict the caching to mounts that use the client
cache (-C flag) only. this should always be the case
for /bin. places where this isnt the case might observe
a performance regression.
the lock order of page.Lock -> palloc.hashlock was
violated in cachedel() which is called from the
pager. change the code to do it in the right oder
to prevent deadlock.
change lookpage to retry on false hit. i assume that
a false hit means:
a) we'r low on memory -> cached page got uncached/reused
b) duppage() got called on the page, meaning theres another
cached copy in the image now.
paging in is expensive compared to the hashtable lookup, so
i think retrying is better.
cleanup fixfault, adding comments.
