Merge pull request #2555 from lioncash/tls

kernel/process: Decouple TLS handling from threads

src/core/hle/kernel/process.cpp

@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
+#include <bitset>
 #include <memory>
 #include <random>
 #include "common/alignment.h"
@@ -48,8 +49,58 @@ void SetupMainThread(Process& owner_process, KernelCore& kernel, u32 priority) {
 }
 } // Anonymous namespace
 
-SharedPtr<Process> Process::Create(Core::System& system, std::string name,
-                                   Process::ProcessType type) {
+// Represents a page used for thread-local storage.
+//
+// Each TLS page contains slots that may be used by processes and threads.
+// Every process and thread is created with a slot in some arbitrary page
+// (whichever page happens to have an available slot).
+class TLSPage {
+public:
+    static constexpr std::size_t num_slot_entries = Memory::PAGE_SIZE / Memory::TLS_ENTRY_SIZE;
+
+    explicit TLSPage(VAddr address) : base_address{address} {}
+
+    bool HasAvailableSlots() const {
+        return !is_slot_used.all();
+    }
+
+    VAddr GetBaseAddress() const {
+        return base_address;
+    }
+
+    std::optional<VAddr> ReserveSlot() {
+        for (std::size_t i = 0; i < is_slot_used.size(); i++) {
+            if (is_slot_used[i]) {
+                continue;
+            }
+
+            is_slot_used[i] = true;
+            return base_address + (i * Memory::TLS_ENTRY_SIZE);
+        }
+
+        return std::nullopt;
+    }
+
+    void ReleaseSlot(VAddr address) {
+        // Ensure that all given addresses are consistent with how TLS pages
+        // are intended to be used when releasing slots.
+        ASSERT(IsWithinPage(address));
+        ASSERT((address % Memory::TLS_ENTRY_SIZE) == 0);
+
+        const std::size_t index = (address - base_address) / Memory::TLS_ENTRY_SIZE;
+        is_slot_used[index] = false;
+    }
+
+private:
+    bool IsWithinPage(VAddr address) const {
+        return base_address <= address && address < base_address + Memory::PAGE_SIZE;
+    }
+
+    VAddr base_address;
+    std::bitset<num_slot_entries> is_slot_used;
+};
+
+SharedPtr<Process> Process::Create(Core::System& system, std::string name, ProcessType type) {
     auto& kernel = system.Kernel();
 
     SharedPtr<Process> process(new Process(system));
@@ -181,61 +232,55 @@ void Process::PrepareForTermination() {
 }
 
 /**
- * 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).
+ * Attempts to find a TLS page that contains a free slot for
+ * use by a thread.
+ *
+ * @returns If a page with an available slot is found, then an iterator
+ *          pointing to the page is returned. Otherwise the end iterator
+ *          is returned instead.
  */
-static std::tuple<std::size_t, std::size_t, bool> FindFreeThreadLocalSlot(
-    const 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 (std::size_t 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 (std::size_t slot = 0; slot < page_tls_slots.size(); ++slot) {
-                if (!page_tls_slots.test(slot)) {
-                    return std::make_tuple(page, slot, false);
-                }
-            }
-        }
+static auto FindTLSPageWithAvailableSlots(std::vector<TLSPage>& tls_pages) {
+    return std::find_if(tls_pages.begin(), tls_pages.end(),
+                        [](const auto& page) { return page.HasAvailableSlots(); });
 }
 
-    return std::make_tuple(0, 0, true);
-}
+VAddr Process::CreateTLSRegion() {
+    auto tls_page_iter = FindTLSPageWithAvailableSlots(tls_pages);
 
-VAddr Process::MarkNextAvailableTLSSlotAsUsed(Thread& thread) {
-    auto [available_page, available_slot, needs_allocation] = FindFreeThreadLocalSlot(tls_slots);
-    const VAddr tls_begin = vm_manager.GetTLSIORegionBaseAddress();
+    if (tls_page_iter == tls_pages.cend()) {
+        const auto region_address =
+            vm_manager.FindFreeRegion(vm_manager.GetTLSIORegionBaseAddress(),
+                                      vm_manager.GetTLSIORegionEndAddress(), Memory::PAGE_SIZE);
+        ASSERT(region_address.Succeeded());
 
-    if (needs_allocation) {
-        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
-
-        // Allocate some memory from the end of the linear heap for this region.
-        auto& tls_memory = thread.GetTLSMemory();
-        tls_memory->insert(tls_memory->end(), Memory::PAGE_SIZE, 0);
-
-        vm_manager.RefreshMemoryBlockMappings(tls_memory.get());
-
-        vm_manager.MapMemoryBlock(tls_begin + available_page * Memory::PAGE_SIZE, tls_memory, 0,
+        const auto map_result = vm_manager.MapMemoryBlock(
+            *region_address, std::make_shared<std::vector<u8>>(Memory::PAGE_SIZE), 0,
             Memory::PAGE_SIZE, MemoryState::ThreadLocal);
+        ASSERT(map_result.Succeeded());
+
+        tls_pages.emplace_back(*region_address);
+
+        const auto reserve_result = tls_pages.back().ReserveSlot();
+        ASSERT(reserve_result.has_value());
+
+        return *reserve_result;
     }
 
-    tls_slots[available_page].set(available_slot);
-
-    return tls_begin + available_page * Memory::PAGE_SIZE + available_slot * Memory::TLS_ENTRY_SIZE;
+    return *tls_page_iter->ReserveSlot();
 }
 
-void Process::FreeTLSSlot(VAddr tls_address) {
-    const VAddr tls_base = tls_address - vm_manager.GetTLSIORegionBaseAddress();
-    const VAddr tls_page = tls_base / Memory::PAGE_SIZE;
-    const VAddr tls_slot = (tls_base % Memory::PAGE_SIZE) / Memory::TLS_ENTRY_SIZE;
+void Process::FreeTLSRegion(VAddr tls_address) {
+    const VAddr aligned_address = Common::AlignDown(tls_address, Memory::PAGE_SIZE);
+    auto iter =
+        std::find_if(tls_pages.begin(), tls_pages.end(), [aligned_address](const auto& page) {
+            return page.GetBaseAddress() == aligned_address;
+        });
 
-    tls_slots[tls_page].reset(tls_slot);
+    // Something has gone very wrong if we're freeing a region
+    // with no actual page available.
+    ASSERT(iter != tls_pages.cend());
+
+    iter->ReleaseSlot(tls_address);
 }
 
 void Process::LoadModule(CodeSet module_, VAddr base_addr) {

src/core/hle/kernel/process.h

@@ -5,7 +5,6 @@
 #pragma once
 
 #include <array>
-#include <bitset>
 #include <cstddef>
 #include <list>
 #include <string>
@@ -32,6 +31,7 @@ namespace Kernel {
 class KernelCore;
 class ResourceLimit;
 class Thread;
+class TLSPage;
 
 struct CodeSet;
 
@@ -260,10 +260,10 @@ public:
     // Thread-local storage management
 
     // Marks the next available region as used and returns the address of the slot.
-    VAddr MarkNextAvailableTLSSlotAsUsed(Thread& thread);
+    [[nodiscard]] VAddr CreateTLSRegion();
 
     // Frees a used TLS slot identified by the given address
-    void FreeTLSSlot(VAddr tls_address);
+    void FreeTLSRegion(VAddr tls_address);
 
 private:
     explicit Process(Core::System& system);
@@ -290,7 +290,7 @@ private:
     u64 code_memory_size = 0;
 
     /// Current status of the process
-    ProcessStatus status;
+    ProcessStatus status{};
 
     /// The ID of this process
     u64 process_id = 0;
@@ -310,7 +310,7 @@ private:
     /// 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;
+    std::vector<TLSPage> tls_pages;
 
     /// Contains the parsed process capability descriptors.
     ProcessCapabilities capabilities;
@@ -339,7 +339,7 @@ private:
     Mutex mutex;
 
     /// Random values for svcGetInfo RandomEntropy
-    std::array<u64, RANDOM_ENTROPY_SIZE> random_entropy;
+    std::array<u64, RANDOM_ENTROPY_SIZE> random_entropy{};
 
     /// List of threads that are running with this process as their owner.
     std::list<const Thread*> thread_list;

src/core/hle/kernel/thread.cpp

@@ -65,7 +65,7 @@ void Thread::Stop() {
     owner_process->UnregisterThread(this);
 
     // Mark the TLS slot in the thread's page as free.
-    owner_process->FreeTLSSlot(tls_address);
+    owner_process->FreeTLSRegion(tls_address);
 }
 
 void Thread::WakeAfterDelay(s64 nanoseconds) {
@@ -205,9 +205,9 @@ ResultVal<SharedPtr<Thread>> Thread::Create(KernelCore& kernel, std::string name
     thread->name = std::move(name);
     thread->callback_handle = kernel.ThreadWakeupCallbackHandleTable().Create(thread).Unwrap();
     thread->owner_process = &owner_process;
+    thread->tls_address = thread->owner_process->CreateTLSRegion();
     thread->scheduler = &system.Scheduler(processor_id);
     thread->scheduler->AddThread(thread);
-    thread->tls_address = thread->owner_process->MarkNextAvailableTLSSlotAsUsed(*thread);
 
     thread->owner_process->RegisterThread(thread.get());
 

src/core/hle/kernel/thread.h

@@ -5,7 +5,6 @@
 #pragma once
 
 #include <functional>
-#include <memory>
 #include <string>
 #include <vector>
 
@@ -78,9 +77,6 @@ enum class ThreadActivity : u32 {
 
 class Thread final : public WaitObject {
 public:
-    using TLSMemory = std::vector<u8>;
-    using TLSMemoryPtr = std::shared_ptr<TLSMemory>;
-
     using MutexWaitingThreads = std::vector<SharedPtr<Thread>>;
 
     using ThreadContext = Core::ARM_Interface::ThreadContext;
@@ -169,14 +165,6 @@ public:
         return thread_id;
     }
 
-    TLSMemoryPtr& GetTLSMemory() {
-        return tls_memory;
-    }
-
-    const TLSMemoryPtr& GetTLSMemory() const {
-        return tls_memory;
-    }
-
     /// Resumes a thread from waiting
     void ResumeFromWait();
 
@@ -463,11 +451,9 @@ private:
     u32 ideal_core{0xFFFFFFFF};
     u64 affinity_mask{0x1};
 
-    TLSMemoryPtr tls_memory = std::make_shared<TLSMemory>();
+    ThreadActivity activity = ThreadActivity::Normal;
 
     std::string name;
-
-    ThreadActivity activity = ThreadActivity::Normal;
 };
 
 /**

src/core/hle/kernel/vm_manager.cpp

@@ -152,22 +152,33 @@ ResultVal<VMManager::VMAHandle> VMManager::MapBackingMemory(VAddr target, u8* me
 }
 
 ResultVal<VAddr> VMManager::FindFreeRegion(u64 size) const {
-    // Find the first Free VMA.
-    const VAddr base = GetASLRRegionBaseAddress();
-    const VMAHandle vma_handle = std::find_if(vma_map.begin(), vma_map.end(), [&](const auto& vma) {
-        if (vma.second.type != VMAType::Free)
-            return false;
+    return FindFreeRegion(GetASLRRegionBaseAddress(), GetASLRRegionEndAddress(), size);
+}
 
-        const VAddr vma_end = vma.second.base + vma.second.size;
-        return vma_end > base && vma_end >= base + size;
+ResultVal<VAddr> VMManager::FindFreeRegion(VAddr begin, VAddr end, u64 size) const {
+    ASSERT(begin < end);
+    ASSERT(size <= end - begin);
+
+    const VMAHandle vma_handle =
+        std::find_if(vma_map.begin(), vma_map.end(), [begin, end, size](const auto& vma) {
+            if (vma.second.type != VMAType::Free) {
+                return false;
+            }
+            const VAddr vma_base = vma.second.base;
+            const VAddr vma_end = vma_base + vma.second.size;
+            const VAddr assumed_base = (begin < vma_base) ? vma_base : begin;
+            const VAddr used_range = assumed_base + size;
+
+            return vma_base <= assumed_base && assumed_base < used_range && used_range < end &&
+                   used_range <= vma_end;
         });
 
-    if (vma_handle == vma_map.end()) {
+    if (vma_handle == vma_map.cend()) {
         // TODO(Subv): Find the correct error code here.
         return ResultCode(-1);
     }
 
-    const VAddr target = std::max(base, vma_handle->second.base);
+    const VAddr target = std::max(begin, vma_handle->second.base);
     return MakeResult<VAddr>(target);
 }
 

src/core/hle/kernel/vm_manager.h

@@ -362,13 +362,38 @@ public:
     ResultVal<VMAHandle> MapBackingMemory(VAddr target, u8* memory, u64 size, MemoryState state);
 
     /**
-     * Finds the first free address that can hold a region of the desired size.
+     * Finds the first free memory region of the given size within
+     * the user-addressable ASLR memory region.
      *
-     * @param size Size of the desired region.
-     * @return The found free address.
+     * @param size The size of the desired region in bytes.
+     *
+     * @returns If successful, the base address of the free region with
+     *          the given size.
      */
     ResultVal<VAddr> FindFreeRegion(u64 size) const;
 
+    /**
+     * Finds the first free address range that can hold a region of the desired size
+     *
+     * @param begin The starting address of the range.
+     *              This is treated as an inclusive beginning address.
+     *
+     * @param end   The ending address of the range.
+     *              This is treated as an exclusive ending address.
+     *
+     * @param size  The size of the free region to attempt to locate,
+     *              in bytes.
+     *
+     * @returns If successful, the base address of the free region with
+     *          the given size.
+     *
+     * @returns If unsuccessful, a result containing an error code.
+     *
+     * @pre The starting address must be less than the ending address.
+     * @pre The size must not exceed the address range itself.
+     */
+    ResultVal<VAddr> FindFreeRegion(VAddr begin, VAddr end, u64 size) const;
+
     /**
      * Maps a memory-mapped IO region at a given address.
      *