RGBCube/serenity
1
Fork 0
mirror of https://github.com/RGBCube/serenity synced 2025-07-25 18:57:45 +00:00
Code Issues Activity Actions

Kernel: Implement lazy committed page allocation

Browse source 
By designating a committed page pool we can guarantee to have physical
pages available for lazy allocation in mappings. However, when forking
we will overcommit. The assumption is that worst-case it's better for
the fork to die due to insufficient physical memory on COW access than
the parent that created the region. If a fork wants to ensure that all
memory is available (trigger a commit) then it can use madvise.

This also means that fork now can gracefully fail if we don't have
enough physical pages available.
This commit is contained in:
Tom 2020-09-04 21:12:25 -06:00 • committed by Andreas Kling
parent e21cc4cff6
commit b2a52f6208
20 changed files with 329 additions and 67 deletions
Download patch file Download diff file Expand all files Collapse all files

60
Kernel/VM/Region.cpp
View file

@ -87,7 +87,7 @@ void Region::unregister_purgeable_page_ranges()
}
}
NonnullOwnPtr<Region> Region::clone()
OwnPtr<Region> Region::clone()
{
ASSERT(Process::current());
@ -122,13 +122,17 @@ NonnullOwnPtr<Region> Region::clone()
if (vmobject().is_inode())
ASSERT(vmobject().is_private_inode());
auto vmobject_clone = m_vmobject->clone();
if (!vmobject_clone)
return {};
#ifdef MM_DEBUG
dbg() << "Region::clone(): CoWing " << name() << " (" << vaddr() << ")";
#endif
// Set up a COW region. The parent (this) region becomes COW as well!
ensure_cow_map().fill(true);
remap();
auto clone_region = Region::create_user_accessible(get_owner().ptr(), m_range, m_vmobject->clone(), m_offset_in_vmobject, m_name, m_access);
auto clone_region = Region::create_user_accessible(get_owner().ptr(), m_range, vmobject_clone.release_nonnull(), m_offset_in_vmobject, m_name, m_access);
clone_region->set_purgeable_page_ranges(*this);
clone_region->ensure_cow_map();
if (m_stack) {
@ -187,7 +191,7 @@ auto Region::set_volatile(VirtualAddress vaddr, size_t size, bool is_volatile, b
// Attempt to remap the page range. We want to make sure we have
// enough memory, if not we need to inform the caller of that
// fact
if (!remap_page_range(first_page_index, last_page_index - first_page_index, true))
if (!remap_page_range(first_page_index, last_page_index - first_page_index))
return SetVolatileError::OutOfMemory;
}
}
@ -224,10 +228,15 @@ bool Region::commit(size_t page_index)
auto& vmobject_physical_page_entry = physical_page_slot(page_index);
if (!vmobject_physical_page_entry.is_null() && !vmobject_physical_page_entry->is_shared_zero_page())
return true;
auto physical_page = MM.allocate_user_physical_page(MemoryManager::ShouldZeroFill::Yes);
if (!physical_page) {
klog() << "MM: commit was unable to allocate a physical page";
return false;
RefPtr<PhysicalPage> physical_page;
if (vmobject_physical_page_entry->is_lazy_committed_page()) {
physical_page = static_cast<AnonymousVMObject&>(*m_vmobject).allocate_committed_page(page_index);
} else {
physical_page = MM.allocate_user_physical_page(MemoryManager::ShouldZeroFill::Yes);
if (!physical_page) {
klog() << "MM: commit was unable to allocate a physical page";
return false;
}
}
vmobject_physical_page_entry = move(physical_page);
remap_page(page_index, false); // caller is in charge of flushing tlb
@ -292,7 +301,7 @@ NonnullOwnPtr<Region> Region::create_kernel_only(const Range& range, NonnullRefP
bool Region::should_cow(size_t page_index) const
{
auto* page = physical_page(page_index);
if (page && page->is_shared_zero_page())
if (page && (page->is_shared_zero_page() || page->is_lazy_committed_page()))
return true;
if (m_shared)
return false;
@ -344,7 +353,7 @@ bool Region::map_individual_page_impl(size_t page_index)
return true;
}
bool Region::remap_page_range(size_t page_index, size_t page_count, bool do_commit)
bool Region::remap_page_range(size_t page_index, size_t page_count)
{
bool success = true;
ScopedSpinLock lock(s_mm_lock);
@ -352,10 +361,6 @@ bool Region::remap_page_range(size_t page_index, size_t page_count, bool do_comm
ScopedSpinLock page_lock(m_page_directory->get_lock());
size_t index = page_index;
while (index < page_index + page_count) {
if (do_commit && !commit(index)) {
success = false;
break;
}
if (!map_individual_page_impl(index)) {
success = false;
break;
@ -455,9 +460,16 @@ PageFaultResponse Region::handle_fault(const PageFault& fault)
#endif
return handle_inode_fault(page_index_in_region);
}
auto& page_slot = physical_page_slot(page_index_in_region);
if (page_slot->is_lazy_committed_page()) {
page_slot = static_cast<AnonymousVMObject&>(*m_vmobject).allocate_committed_page(page_index_in_region);
remap_page(page_index_in_region);
return PageFaultResponse::Continue;
}
#ifdef MAP_SHARED_ZERO_PAGE_LAZILY
if (fault.is_read()) {
physical_page_slot(page_index_in_region) = MM.shared_zero_page();
page_slot = MM.shared_zero_page();
remap_page(page_index_in_region);
return PageFaultResponse::Continue;
}
@ -472,7 +484,8 @@ PageFaultResponse Region::handle_fault(const PageFault& fault)
#ifdef PAGE_FAULT_DEBUG
dbg() << "PV(cow) fault in Region{" << this << "}[" << page_index_in_region << "]";
#endif
if (physical_page(page_index_in_region)->is_shared_zero_page()) {
auto* phys_page = physical_page(page_index_in_region);
if (phys_page->is_shared_zero_page() || phys_page->is_lazy_committed_page()) {
#ifdef PAGE_FAULT_DEBUG
dbg() << "NP(zero) fault in Region{" << this << "}[" << page_index_in_region << "]";
#endif
@ -493,7 +506,7 @@ PageFaultResponse Region::handle_zero_fault(size_t page_index_in_region)
auto& page_slot = physical_page_slot(page_index_in_region);
if (!page_slot.is_null() && !page_slot->is_shared_zero_page()) {
if (!page_slot.is_null() && !page_slot->is_shared_zero_page() && !page_slot->is_lazy_committed_page()) {
#ifdef PAGE_FAULT_DEBUG
dbg() << "MM: zero_page() but page already present. Fine with me!";
#endif
@ -506,16 +519,19 @@ PageFaultResponse Region::handle_zero_fault(size_t page_index_in_region)
if (current_thread != nullptr)
current_thread->did_zero_fault();
auto page = MM.allocate_user_physical_page(MemoryManager::ShouldZeroFill::Yes);
if (page.is_null()) {
klog() << "MM: handle_zero_fault was unable to allocate a physical page";
return PageFaultResponse::OutOfMemory;
if (page_slot->is_lazy_committed_page()) {
page_slot = static_cast<AnonymousVMObject&>(*m_vmobject).allocate_committed_page(page_index_in_region);
} else {
page_slot = MM.allocate_user_physical_page(MemoryManager::ShouldZeroFill::Yes);
if (page_slot.is_null()) {
klog() << "MM: handle_zero_fault was unable to allocate a physical page";
return PageFaultResponse::OutOfMemory;
}
}
#ifdef PAGE_FAULT_DEBUG
dbg() << " >> ZERO " << page->paddr();
dbg() << " >> ZERO " << page_slot->paddr();
#endif
page_slot = move(page);
if (!remap_page(page_index_in_region)) {
klog() << "MM: handle_zero_fault was unable to allocate a page table to map " << page_slot;
return PageFaultResponse::OutOfMemory;