i2p/haiku-beta6
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

kernel/vm: Numerous fixes to area cutting and commitment logic.

Browse Source 
* Split the "is only cache user" logic from cut_area into
   a helper routine.

 * If we can't modify a cache in cut_area, then we can't modify
   its commitment either, so add that to the check.

 * If we didn't split the areas, then modifying the commitments
   doesn't make sense and will cause problems, so move that logic
   into the "modify cache" branch.

 * Use the new helper routine in set_memory_protection rather than
   checking only cache->temporary.

Combined with the previous commit, seems to fix #19624.
This commit is contained in:
Augustin Cavalier
2025-07-07 17:20:34 -04:00
parent 92f73a4e2c
commit ceec330bbb
1 changed files with 35 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/system/kernel/vm/vm.cpp
+35 -25
View File
@@ -626,6 +626,16 @@ compute_area_page_commitment(VMArea* area)
}
static bool
is_area_only_cache_user(VMArea* area)
{
return area->cache->type == CACHE_TYPE_RAM
&& area->cache->areas.First() == area
&& area->cache->areas.GetNext(area) == NULL
&& area->cache->consumers.IsEmpty();
}
/*! The caller must have reserved enough pages the translation map
implementation might need to map this page.
The page's cache must be locked.
@@ -773,22 +783,21 @@ cut_area(VMAddressSpace* addressSpace, VMArea* area, addr_t address,
allocationFlags = 0;
}
int resizePriority = priority;
const bool overcommitting = (area->protection & B_OVERCOMMITTING_AREA) != 0,
writable = (area->protection & (B_WRITE_AREA | B_KERNEL_WRITE_AREA)) != 0;
if ((area->page_protections != NULL || !writable) && !overcommitting) {
// We'll adjust commitments directly, rather than letting VMCache do it.
resizePriority = -1;
}
VMCache* cache = vm_area_get_locked_cache(area);
VMCacheChainLocker cacheChainLocker(cache);
cacheChainLocker.LockAllSourceCaches();
// If no one else uses the area's cache and it's an anonymous cache, we can
// resize or split it, too.
bool onlyCacheUser = cache->areas.First() == area && cache->areas.GetNext(area) == NULL
&& cache->consumers.IsEmpty() && cache->type == CACHE_TYPE_RAM;
const bool onlyCacheUser = is_area_only_cache_user(area);
int resizePriority = priority;
const bool overcommitting = (area->protection & B_OVERCOMMITTING_AREA) != 0,
writable = (area->protection & (B_WRITE_AREA | B_KERNEL_WRITE_AREA)) != 0;
if (onlyCacheUser && !overcommitting && (area->page_protections != NULL || !writable)) {
// We'll adjust commitments directly, rather than letting VMCache do it.
resizePriority = -1;
}
const addr_t oldSize = area->Size();
@@ -1007,6 +1016,19 @@ cut_area(VMAddressSpace* addressSpace, VMArea* area, addr_t address,
error = cache->Resize(cache->virtual_base + firstNewSize, resizePriority);
ASSERT_ALWAYS(error == B_OK);
if (resizePriority == -1) {
// Adjust commitments.
const off_t areaCommit = compute_area_page_commitment(area) * B_PAGE_SIZE;
if (areaCommit < area->cache->committed_size) {
secondArea->cache->committed_size += area->cache->committed_size - areaCommit;
area->cache->committed_size = areaCommit;
}
area->cache->Commit(areaCommit, priority);
const off_t secondCommit = compute_area_page_commitment(secondArea) * B_PAGE_SIZE;
secondArea->cache->Commit(secondCommit, priority);
}
} else {
// Reuse the existing cache.
error = map_backing_store(addressSpace, cache, secondCacheOffset,
@@ -1047,19 +1069,6 @@ cut_area(VMAddressSpace* addressSpace, VMArea* area, addr_t address,
free_etc(areaOldProtections, allocationFlags);
}
if (resizePriority == -1) {
// Adjust commitments.
const off_t areaCommit = compute_area_page_commitment(area) * B_PAGE_SIZE;
if (areaCommit < area->cache->committed_size) {
secondArea->cache->committed_size += area->cache->committed_size - areaCommit;
area->cache->committed_size = areaCommit;
}
area->cache->Commit(areaCommit, priority);
const off_t secondCommit = compute_area_page_commitment(secondArea) * B_PAGE_SIZE;
secondArea->cache->Commit(secondCommit, priority);
}
if (_secondArea != NULL)
*_secondArea = secondArea;
@@ -6420,7 +6429,7 @@ _user_set_memory_protection(void* _address, size_t size, uint32 protection)
cacheChainLocker.LockAllSourceCaches();
// Adjust the committed size, if necessary.
if (topCache->temporary && !topCache->CanOvercommit()) {
if (is_area_only_cache_user(area) && !topCache->CanOvercommit()) {
const bool becomesWritable = (protection & B_WRITE_AREA) != 0;
ssize_t commitmentChange = 0;
const off_t areaCacheBase = area->Base() - area->cache_offset;
@@ -6442,7 +6451,8 @@ _user_set_memory_protection(void* _address, size_t size, uint32 protection)
if (commitmentChange != 0) {
off_t newCommitment = topCache->committed_size + commitmentChange;
if (newCommitment > PAGE_ALIGN(topCache->virtual_end - topCache->virtual_base)) {
const ssize_t topCacheSize = topCache->virtual_end - topCache->virtual_base;
if (newCommitment > topCacheSize) {
// This should only happen in the case where this process fork()ed,
// duplicating the commitment, and then the child exited, resulting
// in the commitments being merged along with the caches.