* Removed DEBUG_PAGE_CACHE_TRANSITIONS debugging.

* Added VMCache::MovePage() and MoveAllPages() to move pages between caches.
* VMAnonymousCache:
  - _MergeSwapPages(): Avoid doing anything, if neither cache has swapped out
    pages.
  - _MergeSwapPages() does now also remove source cache pages that are
    shadowed by consumer swap pages. This allows us to call _MergeSwapPages()
    before _MergePagesSmallerSource(), save the swap page shadowing check
    there and get rid of the vm_page::merge_swap flag. This is an
    optimization based on the assumption that usually none or only few pages
    are swapped out, so we save a lot of checks.
  - Implemented _MergePagesSmallerConsumer() as an alternative to
    _MergePagesSmallerSource(). The former is used when the source cache has
    more pages than the consumer cache. It iterates over the consumer cache's
    pages, moves them to the source and finally moves all pages back to the
    consumer. The final move is relatively cheap (though unfortunately we
    still have to update all pages' vm_page::cache field), so that overall we
    save iterations of the main loop with the more expensive checks.

The optimizations particularly improve the common fork()+exec*() situations.
fork() uses CoW, which is implemented by putting two new empty caches between
the to be copied area and its cache. exec*() destroys one copy of the area,
its cache and thus causes merging of the other new cache with the old cache.
Since this usually happens in a very short time, the old cache does still
contain many pages and the new cache only few. Previously the many pages were
all checked and moved individually. Now we do that for the few pages instead.

A very extreme example of this situation is the Haiku image build. jam has a
huge heap (> 200 MB) and it fork()s+exec*()s for every action to be executed.
Since during the cache merging the cache is locked, any write access to a
heap page causes jam to block until the cache merging is done. Formerly that
took so long that it killed a lot of parallelism in multi-job builds. That
could be observed particularly well when lots of small actions where executed
(like the Link, XRes, Mimeset, SetType, SetVersion combos when building
executables/libraries/add-ons). Those look dramatically better now.
The overall speed improvement for a -j8 image build on my machine is only
about 15%, though.


git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@34784 a95241bf-73f2-0310-859d-f6bbb57e9c96
This commit is contained in:
Ingo Weinhold 2009-12-27 16:14:13 +00:00
parent 6afe50d424
commit eb8dc1ebfb
6 changed files with 118 additions and 81 deletions

View File

@ -84,10 +84,6 @@
// be in.
#define DEBUG_PAGE_QUEUE 0
// Enables extra debug fields in the vm_page used to track page transitions
// between caches.
#define DEBUG_PAGE_CACHE_TRANSITIONS 0
// Enables a global list of all vm_cache structures.
#define DEBUG_CACHE_LIST KDEBUG_LEVEL_1

View File

@ -96,6 +96,8 @@ public:
vm_page* LookupPage(off_t offset);
void InsertPage(vm_page* page, off_t offset);
void RemovePage(vm_page* page);
void MovePage(vm_page* page);
void MoveAllPages(VMCache* fromCache);
void AddConsumer(VMCache* consumer);

View File

@ -90,18 +90,13 @@ struct vm_page {
void* queue;
#endif
#if DEBUG_PAGE_CACHE_TRANSITIONS
uint32 debug_flags;
struct vm_page* collided_page;
#endif
uint8 type : 2;
uint8 state : 3;
uint8 is_cleared : 1;
// is currently only used in vm_page_allocate_page_run()
uint8 busy_writing : 1;
uint8 merge_swap : 1;
uint8 unused : 1;
// used in VMAnonymousCache::Merge()
int8 usage_count;

View File

@ -751,11 +751,15 @@ VMAnonymousCache::Merge(VMCache* _source)
if (committed_size > actualSize)
_Commit(actualSize);
// Move all not shadowed pages from the source to the consumer cache.
_MergePagesSmallerSource(source);
// Move all not shadowed swap pages from the source to the consumer cache.
// Also remove all source pages that are shadowed by consumer swap pages.
_MergeSwapPages(source);
// Move all not shadowed pages from the source to the consumer cache.
if (source->page_count < page_count)
_MergePagesSmallerSource(source);
else
_MergePagesSmallerConsumer(source);
}
@ -916,44 +920,59 @@ VMAnonymousCache::_Commit(off_t size)
void
VMAnonymousCache::_MergePagesSmallerSource(VMAnonymousCache* source)
{
// The source cache has less pages than the consumer (this cache), so we
// iterate through the source's pages and move the ones that are not
// shadowed up to the consumer.
for (VMCachePagesTree::Iterator it = source->pages.GetIterator();
vm_page* page = it.Next();) {
// Note: Removing the current node while iterating through a
// IteratableSplayTree is safe.
vm_page* consumerPage = LookupPage(
(off_t)page->cache_offset << PAGE_SHIFT);
swap_addr_t consumerSwapSlot = _SwapBlockGetAddress(page->cache_offset);
if (consumerPage == NULL && consumerSwapSlot == SWAP_SLOT_NONE) {
if (consumerPage == NULL) {
// the page is not yet in the consumer cache - move it upwards
source->RemovePage(page);
InsertPage(page, (off_t)page->cache_offset << PAGE_SHIFT);
// If the moved-up page has a swap page associated, we mark it, so
// that the swap page is moved upwards, too. We would lose if the
// page was modified and written to swap, and is now not marked
// modified.
if (source->_SwapBlockGetAddress(page->cache_offset)
!= SWAP_SLOT_NONE) {
page->merge_swap = true;
}
#if DEBUG_PAGE_CACHE_TRANSITIONS
} else {
page->debug_flags = 0;
if (consumerPage->state == PAGE_STATE_BUSY)
page->debug_flags |= 0x1;
if (consumerPage->type == PAGE_TYPE_DUMMY)
page->debug_flags |= 0x2;
page->collided_page = consumerPage;
consumerPage->collided_page = page;
#endif // DEBUG_PAGE_CACHE_TRANSITIONS
}
}
}
void
VMAnonymousCache::_MergePagesSmallerConsumer(VMAnonymousCache* source)
{
// The consumer (this cache) has less pages than the source, so we move the
// consumer's pages to the source (freeing shadowed ones) and finally just
// all pages of the source back to the consumer.
for (VMCachePagesTree::Iterator it = pages.GetIterator();
vm_page* page = it.Next();) {
// If a source page is in the way, remove and free it.
vm_page* sourcePage = source->LookupPage(
(off_t)page->cache_offset << PAGE_SHIFT);
if (sourcePage != NULL) {
source->RemovePage(sourcePage);
vm_page_free(source, sourcePage);
}
// Note: Removing the current node while iterating through a
// IteratableSplayTree is safe.
source->MovePage(page);
}
MoveAllPages(source);
}
void
VMAnonymousCache::_MergeSwapPages(VMAnonymousCache* source)
{
// If neither source nor consumer have swap pages, we don't have to do
// anything.
if (source->fAllocatedSwapSize == 0 && fAllocatedSwapSize == 0)
return;
for (off_t offset = source->virtual_base
& ~(off_t)(B_PAGE_SIZE * SWAP_BLOCK_PAGES - 1);
offset < source->virtual_end;
@ -965,48 +984,43 @@ VMAnonymousCache::_MergeSwapPages(VMAnonymousCache* source)
swap_hash_key key = { source, swapBlockPageIndex };
swap_block* sourceSwapBlock = sSwapHashTable.Lookup(key);
if (sourceSwapBlock == NULL)
continue;
// remove the source swap block -- we will either take over the swap
// space (and the block) or free it
sSwapHashTable.RemoveUnchecked(sourceSwapBlock);
if (sourceSwapBlock != NULL)
sSwapHashTable.RemoveUnchecked(sourceSwapBlock);
key.cache = this;
swap_block* swapBlock = sSwapHashTable.Lookup(key);
locker.Unlock();
// remove all source pages that are shadowed by consumer swap pages
if (swapBlock != NULL) {
for (uint32 i = 0; i < SWAP_BLOCK_PAGES; i++) {
if (swapBlock->swap_slots[i] != SWAP_SLOT_NONE) {
vm_page* page = source->LookupPage(
(off_t)(swapBlockPageIndex + i) << PAGE_SHIFT);
source->RemovePage(page);
vm_page_free(source, page);
}
}
}
if (sourceSwapBlock == NULL)
continue;
for (uint32 i = 0; i < SWAP_BLOCK_PAGES; i++) {
off_t pageIndex = swapBlockPageIndex + i;
swap_addr_t sourceSlotIndex = sourceSwapBlock->swap_slots[i];
if (sourceSlotIndex == SWAP_SLOT_NONE)
// this page is not swapped out
continue;
vm_page* page = LookupPage((off_t)pageIndex << PAGE_SHIFT);
bool keepSwapPage = true;
if (page != NULL && !page->merge_swap) {
// The consumer already has a page at this index and it wasn't
// one taken over from the source. So we can simply free the
// swap space.
keepSwapPage = false;
} else {
if (page != NULL) {
// The page was taken over from the source cache. Clear the
// indicator flag. We'll take over the swap page too.
page->merge_swap = false;
} else if (swapBlock != NULL
&& swapBlock->swap_slots[i] != SWAP_SLOT_NONE) {
// There's no page in the consumer cache, but a swap page.
// Free the source swap page.
keepSwapPage = false;
}
}
if (!keepSwapPage) {
if ((swapBlock != NULL
&& swapBlock->swap_slots[i] != SWAP_SLOT_NONE)
|| LookupPage((off_t)pageIndex << PAGE_SHIFT) != NULL) {
// The consumer already has a page or a swapped out page
// at this index. So we can free the source swap space.
swap_slot_dealloc(sourceSlotIndex, 1);
sourceSwapBlock->swap_slots[i] = SWAP_SLOT_NONE;
sourceSwapBlock->used--;
@ -1036,7 +1050,7 @@ VMAnonymousCache::_MergeSwapPages(VMAnonymousCache* source)
locker.Unlock();
} else {
// We need to take over some of the source's swap pages and there's
// already swap block in the consumer cache. Copy the respective
// already a swap block in the consumer cache. Copy the respective
// swap addresses and discard the source swap block.
for (uint32 i = 0; i < SWAP_BLOCK_PAGES; i++) {
if (sourceSwapBlock->swap_slots[i] != SWAP_SLOT_NONE)

View File

@ -13,6 +13,8 @@
#include <stddef.h>
#include <stdlib.h>
#include <algorithm>
#include <arch/cpu.h>
#include <condition_variable.h>
#include <debug.h>
@ -789,6 +791,53 @@ VMCache::RemovePage(vm_page* page)
}
/*! Moves the given page from its current cache inserts it into this cache.
Both caches must be locked.
*/
void
VMCache::MovePage(vm_page* page)
{
VMCache* oldCache = page->cache;
AssertLocked();
oldCache->AssertLocked();
// remove from old cache
oldCache->pages.Remove(page);
oldCache->page_count--;
T2(RemovePage(oldCache, page));
// insert here
pages.Insert(page);
page_count++;
page->cache = this;
T2(InsertPage(this, page, page->cache_offset << PAGE_SHIFT));
}
/*! Moves all pages from the given cache to this one.
Both caches must be locked. This cache must be empty.
*/
void
VMCache::MoveAllPages(VMCache* fromCache)
{
AssertLocked();
fromCache->AssertLocked();
ASSERT(page_count == 0);
std::swap(fromCache->pages, pages);
page_count = fromCache->page_count;
fromCache->page_count = 0;
for (VMCachePagesTree::Iterator it = pages.GetIterator();
vm_page* page = it.Next();) {
page->cache = this;
T2(RemovePage(fromCache, page));
T2(InsertPage(this, page, page->cache_offset << PAGE_SHIFT));
}
}
/*! Waits until one or more events happened for a given page which belongs to
this cache.
The cache must be locked. It will be unlocked by the method. \a relock
@ -1165,16 +1214,6 @@ VMCache::Merge(VMCache* source)
// the page is not yet in the consumer cache - move it upwards
source->RemovePage(page);
InsertPage(page, (off_t)page->cache_offset << PAGE_SHIFT);
#if DEBUG_PAGE_CACHE_TRANSITIONS
} else {
page->debug_flags = 0;
if (consumerPage->state == PAGE_STATE_BUSY)
page->debug_flags |= 0x1;
if (consumerPage->type == PAGE_TYPE_DUMMY)
page->debug_flags |= 0x2;
page->collided_page = consumerPage;
consumerPage->collided_page = page;
#endif // DEBUG_PAGE_CACHE_TRANSITIONS
}
}
}

View File

@ -564,10 +564,6 @@ dump_page(int argc, char **argv)
#if DEBUG_PAGE_QUEUE
kprintf("queue: %p\n", page->queue);
#endif
#if DEBUG_PAGE_CACHE_TRANSITIONS
kprintf("debug_flags: 0x%lx\n", page->debug_flags);
kprintf("collided page: %p\n", page->collided_page);
#endif // DEBUG_PAGE_CACHE_TRANSITIONS
kprintf("area mappings:\n");
vm_page_mappings::Iterator iterator = page->mappings.GetIterator();
@ -1906,15 +1902,10 @@ vm_page_init(kernel_args *args)
sPages[i].wired_count = 0;
sPages[i].usage_count = 0;
sPages[i].busy_writing = false;
sPages[i].merge_swap = false;
sPages[i].cache = NULL;
#if DEBUG_PAGE_QUEUE
sPages[i].queue = NULL;
#endif
#if DEBUG_PAGE_CACHE_TRANSITIONS
sPages[i].debug_flags = 0;
sPages[i].collided_page = NULL;
#endif // DEBUG_PAGE_CACHE_TRANSITIONS
enqueue_page(&sFreePageQueue, &sPages[i]);
}