Kernel/Threads: Dynamically allocate the TLS region for threads in the BASE region of the linear heap.

Each thread gets a 0x200-byte area from the 0x1000-sized page, when all 8 thread slots in a single page are used up, the kernel allocates a new page to hold another 8 entries.

This is consistent with what the real kernel does.
This commit is contained in:
Subv 2016-04-19 17:12:48 -05:00
parent 3e7e8daf59
commit d192fb066d
5 changed files with 74 additions and 28 deletions

View File

@ -109,7 +109,6 @@ struct MemoryArea {
static MemoryArea memory_areas[] = {
{SHARED_MEMORY_VADDR, SHARED_MEMORY_SIZE, "Shared Memory"}, // Shared memory
{VRAM_VADDR, VRAM_SIZE, "VRAM"}, // Video memory (VRAM)
{TLS_AREA_VADDR, TLS_AREA_SIZE, "TLS Area"}, // TLS memory
};
}

View File

@ -140,8 +140,11 @@ public:
MemoryRegionInfo* memory_region = nullptr;
/// Bitmask of the used TLS slots
std::bitset<300> used_tls_slots;
/// The Thread Local Storage area is allocated as processes create threads,
/// each TLS area is 0x200 bytes, so one page (0x1000) is split up in 8 parts, and each part
/// holds the TLS for a specific thread. This vector contains which parts are in use for each page as a bitmask.
/// This vector will grow as more pages are allocated for new threads.
std::vector<std::bitset<8>> tls_slots;
VAddr GetLinearHeapAreaAddress() const;
VAddr GetLinearHeapBase() const;

View File

@ -117,9 +117,10 @@ void Thread::Stop() {
}
wait_objects.clear();
Kernel::g_current_process->used_tls_slots[tls_index] = false;
g_current_process->misc_memory_used -= Memory::TLS_ENTRY_SIZE;
g_current_process->memory_region->used -= Memory::TLS_ENTRY_SIZE;
// Mark the TLS slot in the thread's page as free.
u32 tls_page = (tls_address - Memory::TLS_AREA_VADDR) / Memory::PAGE_SIZE;
u32 tls_slot = ((tls_address - Memory::TLS_AREA_VADDR) % Memory::PAGE_SIZE) / Memory::TLS_ENTRY_SIZE;
Kernel::g_current_process->tls_slots[tls_page].reset(tls_slot);
HLE::Reschedule(__func__);
}
@ -366,6 +367,31 @@ static void DebugThreadQueue() {
}
}
/**
* Finds a free location for the TLS section of a thread.
* @param tls_slots The TLS page array of the thread's owner process.
* Returns a tuple of (page, slot, alloc_needed) where:
* page: The index of the first allocated TLS page that has free slots.
* slot: The index of the first free slot in the indicated page.
* alloc_needed: Whether there's a need to allocate a new TLS page (All pages are full).
*/
std::tuple<u32, u32, bool> GetFreeThreadLocalSlot(std::vector<std::bitset<8>>& tls_slots) {
// Iterate over all the allocated pages, and try to find one where not all slots are used.
for (unsigned page = 0; page < tls_slots.size(); ++page) {
const auto& page_tls_slots = tls_slots[page];
if (!page_tls_slots.all()) {
// We found a page with at least one free slot, find which slot it is
for (unsigned slot = 0; slot < page_tls_slots.size(); ++slot) {
if (!page_tls_slots.test(slot)) {
return std::make_tuple(page, slot, false);
}
}
}
}
return std::make_tuple(0, 0, true);
}
ResultVal<SharedPtr<Thread>> Thread::Create(std::string name, VAddr entry_point, s32 priority,
u32 arg, s32 processor_id, VAddr stack_top) {
if (priority < THREADPRIO_HIGHEST || priority > THREADPRIO_LOWEST) {
@ -403,22 +429,50 @@ ResultVal<SharedPtr<Thread>> Thread::Create(std::string name, VAddr entry_point,
thread->name = std::move(name);
thread->callback_handle = wakeup_callback_handle_table.Create(thread).MoveFrom();
thread->owner_process = g_current_process;
thread->tls_index = -1;
thread->waitsynch_waited = false;
// Find the next available TLS index, and mark it as used
auto& used_tls_slots = Kernel::g_current_process->used_tls_slots;
for (unsigned int i = 0; i < used_tls_slots.size(); ++i) {
if (used_tls_slots[i] == false) {
thread->tls_index = i;
used_tls_slots[i] = true;
break;
auto& tls_slots = Kernel::g_current_process->tls_slots;
bool needs_allocation = true;
u32 available_page; // Which allocated page has free space
u32 available_slot; // Which slot within the page is free
std::tie(available_page, available_slot, needs_allocation) = GetFreeThreadLocalSlot(tls_slots);
if (needs_allocation) {
// There are no already-allocated pages with free slots, lets allocate a new one.
// TLS pages are allocated from the BASE region in the linear heap.
MemoryRegionInfo* memory_region = GetMemoryRegion(MemoryRegion::BASE);
auto& linheap_memory = memory_region->linear_heap_memory;
if (linheap_memory->size() + Memory::PAGE_SIZE > memory_region->size) {
LOG_ERROR(Kernel_SVC, "Not enough space in region to allocate a new TLS page for thread");
return ResultCode(ErrorDescription::OutOfMemory, ErrorModule::Kernel, ErrorSummary::OutOfResource, ErrorLevel::Permanent);
}
u32 offset = linheap_memory->size();
// Allocate some memory from the end of the linear heap for this region.
linheap_memory->insert(linheap_memory->end(), Memory::PAGE_SIZE, 0);
memory_region->used += Memory::PAGE_SIZE;
Kernel::g_current_process->linear_heap_used += Memory::PAGE_SIZE;
tls_slots.emplace_back(0); // The page is completely available at the start
available_page = tls_slots.size() - 1;
available_slot = 0; // Use the first slot in the new page
auto& vm_manager = Kernel::g_current_process->vm_manager;
vm_manager.RefreshMemoryBlockMappings(linheap_memory.get());
// Map the page to the current process' address space.
// TODO(Subv): Find the correct MemoryState for this region.
vm_manager.MapMemoryBlock(Memory::TLS_AREA_VADDR + available_page * Memory::PAGE_SIZE,
linheap_memory, offset, Memory::PAGE_SIZE, MemoryState::Private);
}
ASSERT_MSG(thread->tls_index != -1, "Out of TLS space");
g_current_process->misc_memory_used += Memory::TLS_ENTRY_SIZE;
g_current_process->memory_region->used += Memory::TLS_ENTRY_SIZE;
// Mark the slot as used
tls_slots[available_page].set(available_slot);
thread->tls_address = Memory::TLS_AREA_VADDR + available_page * Memory::PAGE_SIZE + available_slot * Memory::TLS_ENTRY_SIZE;
// TODO(peachum): move to ScheduleThread() when scheduler is added so selected core is used
// to initialize the context
@ -508,10 +562,6 @@ void Thread::SetWaitSynchronizationOutput(s32 output) {
context.cpu_registers[1] = output;
}
VAddr Thread::GetTLSAddress() const {
return Memory::TLS_AREA_VADDR + tls_index * Memory::TLS_ENTRY_SIZE;
}
////////////////////////////////////////////////////////////////////////////////////////////////////
void ThreadingInit() {

View File

@ -127,7 +127,7 @@ public:
* Returns the Thread Local Storage address of the current thread
* @returns VAddr of the thread's TLS
*/
VAddr GetTLSAddress() const;
VAddr GetTLSAddress() const { return tls_address; }
Core::ThreadContext context;
@ -144,7 +144,7 @@ public:
s32 processor_id;
s32 tls_index; ///< Index of the Thread Local Storage of the thread
VAddr tls_address; ///< Virtual address of the Thread Local Storage of the thread
bool waitsynch_waited; ///< Set to true if the last svcWaitSynch call caused the thread to wait

View File

@ -100,15 +100,9 @@ enum : VAddr {
SHARED_PAGE_SIZE = 0x00001000,
SHARED_PAGE_VADDR_END = SHARED_PAGE_VADDR + SHARED_PAGE_SIZE,
// TODO(yuriks): The size of this area is dynamic, the kernel grows
// it as more and more threads are created. For now we'll just use a
// hardcoded value.
/// Area where TLS (Thread-Local Storage) buffers are allocated.
TLS_AREA_VADDR = 0x1FF82000,
TLS_ENTRY_SIZE = 0x200,
TLS_AREA_SIZE = 300 * TLS_ENTRY_SIZE + 0x800, // Space for up to 300 threads + round to page size
TLS_AREA_VADDR_END = TLS_AREA_VADDR + TLS_AREA_SIZE,
/// Equivalent to LINEAR_HEAP_VADDR, but expanded to cover the extra memory in the New 3DS.
NEW_LINEAR_HEAP_VADDR = 0x30000000,