core: hle: kernel: k_process: Implement thread local storage accurately.

3 files changed, 99 insertions, 111 deletions

diff --git a/src/core/hle/kernel/k_process.cpp b/src/core/hle/kernel/k_process.cpp
index 9233261cd..60dc9a048 100644
--- a/src/core/hle/kernel/k_process.cpp
+++ b/src/core/hle/kernel/k_process.cpp
@@ -70,58 +70,6 @@ void SetupMainThread(Core::System& system, KProcess& owner_process, u32 priority
 }
 } // Anonymous namespace
 
-// 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 =
-        Core::Memory::PAGE_SIZE / Core::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 * Core::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 % Core::Memory::TLS_ENTRY_SIZE) == 0);
-
-        const std::size_t index = (address - base_address) / Core::Memory::TLS_ENTRY_SIZE;
-        is_slot_used[index] = false;
-    }
-
-private:
-    bool IsWithinPage(VAddr address) const {
-        return base_address <= address && address < base_address + Core::Memory::PAGE_SIZE;
-    }
-
-    VAddr base_address;
-    std::bitset<num_slot_entries> is_slot_used;
-};
-
 ResultCode KProcess::Initialize(KProcess* process, Core::System& system, std::string process_name,
                                 ProcessType type, KResourceLimit* res_limit) {
     auto& kernel = system.Kernel();
@@ -404,7 +352,7 @@ ResultCode KProcess::LoadFromMetadata(const FileSys::ProgramMetadata& metadata,
     }
 
     // Create TLS region
-    tls_region_address = CreateTLSRegion();
+    R_TRY(this->CreateThreadLocalRegion(std::addressof(tls_region_address)));
     memory_reservation.Commit();
 
     return handle_table.Initialize(capabilities.GetHandleTableSize());
@@ -444,7 +392,7 @@ void KProcess::PrepareForTermination() {
 
     stop_threads(kernel.System().GlobalSchedulerContext().GetThreadList());
 
-    FreeTLSRegion(tls_region_address);
+    this->DeleteThreadLocalRegion(tls_region_address);
     tls_region_address = 0;
 
     if (resource_limit) {
@@ -487,63 +435,103 @@ void KProcess::Finalize() {
     KAutoObjectWithSlabHeapAndContainer<KProcess, KWorkerTask>::Finalize();
 }
 
-/**
- * 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 auto FindTLSPageWithAvailableSlots(std::vector<TLSPage>& tls_pages) {
-    return std::find_if(tls_pages.begin(), tls_pages.end(),
-                        [](const auto& page) { return page.HasAvailableSlots(); });
-}
+ResultCode KProcess::CreateThreadLocalRegion(VAddr* out) {
+    KThreadLocalPage* tlp = nullptr;
+    VAddr tlr = 0;
 
-VAddr KProcess::CreateTLSRegion() {
-    KScopedSchedulerLock lock(kernel);
-    if (auto tls_page_iter{FindTLSPageWithAvailableSlots(tls_pages)};
-        tls_page_iter != tls_pages.cend()) {
-        return *tls_page_iter->ReserveSlot();
-    }
+    // See if we can get a region from a partially used TLP.
+    {
+        KScopedSchedulerLock sl{kernel};
+
+        if (auto it = partially_used_tlp_tree.begin(); it != partially_used_tlp_tree.end()) {
+            tlr = it->Reserve();
+            ASSERT(tlr != 0);
+
+            if (it->IsAllUsed()) {
+                tlp = std::addressof(*it);
+                partially_used_tlp_tree.erase(it);
+                fully_used_tlp_tree.insert(*tlp);
+            }
 
-    Page* const tls_page_ptr{kernel.GetUserSlabHeapPages().Allocate()};
-    ASSERT(tls_page_ptr);
+            *out = tlr;
+            return ResultSuccess;
+        }
+    }
 
-    const VAddr start{page_table->GetKernelMapRegionStart()};
-    const VAddr size{page_table->GetKernelMapRegionEnd() - start};
-    const PAddr tls_map_addr{kernel.System().DeviceMemory().GetPhysicalAddr(tls_page_ptr)};
-    const VAddr tls_page_addr{page_table
-                                  ->AllocateAndMapMemory(1, PageSize, true, start, size / PageSize,
-                                                         KMemoryState::ThreadLocal,
-                                                         KMemoryPermission::UserReadWrite,
-                                                         tls_map_addr)
-                                  .ValueOr(0)};
+    // Allocate a new page.
+    tlp = KThreadLocalPage::Allocate(kernel);
+    R_UNLESS(tlp != nullptr, ResultOutOfMemory);
+    auto tlp_guard = SCOPE_GUARD({ KThreadLocalPage::Free(kernel, tlp); });
 
-    ASSERT(tls_page_addr);
+    // Initialize the new page.
+    R_TRY(tlp->Initialize(kernel, this));
 
-    std::memset(tls_page_ptr, 0, PageSize);
-    tls_pages.emplace_back(tls_page_addr);
+    // Reserve a TLR.
+    tlr = tlp->Reserve();
+    ASSERT(tlr != 0);
 
-    const auto reserve_result{tls_pages.back().ReserveSlot()};
-    ASSERT(reserve_result.has_value());
+    // Insert into our tree.
+    {
+        KScopedSchedulerLock sl{kernel};
+        if (tlp->IsAllUsed()) {
+            fully_used_tlp_tree.insert(*tlp);
+        } else {
+            partially_used_tlp_tree.insert(*tlp);
+        }
+    }
 
-    return *reserve_result;
+    // We succeeded!
+    tlp_guard.Cancel();
+    *out = tlr;
+    return ResultSuccess;
 }
 
-void KProcess::FreeTLSRegion(VAddr tls_address) {
-    KScopedSchedulerLock lock(kernel);
-    const VAddr aligned_address = Common::AlignDown(tls_address, Core::Memory::PAGE_SIZE);
-    auto iter =
-        std::find_if(tls_pages.begin(), tls_pages.end(), [aligned_address](const auto& page) {
-            return page.GetBaseAddress() == aligned_address;
-        });
+ResultCode KProcess::DeleteThreadLocalRegion(VAddr addr) {
+    KThreadLocalPage* page_to_free = nullptr;
+
+    // Release the region.
+    {
+        KScopedSchedulerLock sl{kernel};
+
+        // Try to find the page in the partially used list.
+        auto it = partially_used_tlp_tree.find_key(Common::AlignDown(addr, PageSize));
+        if (it == partially_used_tlp_tree.end()) {
+            // If we don't find it, it has to be in the fully used list.
+            it = fully_used_tlp_tree.find_key(Common::AlignDown(addr, PageSize));
+            R_UNLESS(it != fully_used_tlp_tree.end(), ResultInvalidAddress);
+
+            // Release the region.
+            it->Release(addr);
+
+            // Move the page out of the fully used list.
+            KThreadLocalPage* tlp = std::addressof(*it);
+            fully_used_tlp_tree.erase(it);
+            if (tlp->IsAllFree()) {
+                page_to_free = tlp;
+            } else {
+                partially_used_tlp_tree.insert(*tlp);
+            }
+        } else {
+            // Release the region.
+            it->Release(addr);
+
+            // Handle the all-free case.
+            KThreadLocalPage* tlp = std::addressof(*it);
+            if (tlp->IsAllFree()) {
+                partially_used_tlp_tree.erase(it);
+                page_to_free = tlp;
+            }
+        }
+    }
+
+    // If we should free the page it was in, do so.
+    if (page_to_free != nullptr) {
+        page_to_free->Finalize();
 
-    // Something has gone very wrong if we're freeing a region
-    // with no actual page available.
-    ASSERT(iter != tls_pages.cend());
+        KThreadLocalPage::Free(kernel, page_to_free);
+    }
 
-    iter->ReleaseSlot(tls_address);
+    return ResultSuccess;
 }
 
 void KProcess::LoadModule(CodeSet code_set, VAddr base_addr) {
diff --git a/src/core/hle/kernel/k_process.h b/src/core/hle/kernel/k_process.h
index cf1b67428..5ed0f2d83 100644
--- a/src/core/hle/kernel/k_process.h
+++ b/src/core/hle/kernel/k_process.h
@@ -15,6 +15,7 @@
 #include "core/hle/kernel/k_condition_variable.h"
 #include "core/hle/kernel/k_handle_table.h"
 #include "core/hle/kernel/k_synchronization_object.h"
+#include "core/hle/kernel/k_thread_local_page.h"
 #include "core/hle/kernel/k_worker_task.h"
 #include "core/hle/kernel/process_capability.h"
 #include "core/hle/kernel/slab_helpers.h"
@@ -362,10 +363,10 @@ public:
     // Thread-local storage management
 
     // Marks the next available region as used and returns the address of the slot.
-    [[nodiscard]] VAddr CreateTLSRegion();
+    [[nodiscard]] ResultCode CreateThreadLocalRegion(VAddr* out);
 
     // Frees a used TLS slot identified by the given address
-    void FreeTLSRegion(VAddr tls_address);
+    ResultCode DeleteThreadLocalRegion(VAddr addr);
 
 private:
     void PinThread(s32 core_id, KThread* thread) {
@@ -413,13 +414,6 @@ private:
     /// The ideal CPU core for this process, threads are scheduled on this core by default.
     u8 ideal_core = 0;
 
-    /// 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<TLSPage> tls_pages;
-
     /// Contains the parsed process capability descriptors.
     ProcessCapabilities capabilities;
 
@@ -482,6 +476,12 @@ private:
     KThread* exception_thread{};
 
     KLightLock state_lock;
+
+    using TLPTree =
+        Common::IntrusiveRedBlackTreeBaseTraits<KThreadLocalPage>::TreeType<KThreadLocalPage>;
+    using TLPIterator = TLPTree::iterator;
+    TLPTree fully_used_tlp_tree;
+    TLPTree partially_used_tlp_tree;
 };
 
 } // namespace Kernel
diff --git a/src/core/hle/kernel/k_thread.cpp b/src/core/hle/kernel/k_thread.cpp
index 736f39e91..ba7f72c6b 100644
--- a/src/core/hle/kernel/k_thread.cpp
+++ b/src/core/hle/kernel/k_thread.cpp
@@ -210,7 +210,7 @@ ResultCode KThread::Initialize(KThreadFunction func, uintptr_t arg, VAddr user_s
     if (owner != nullptr) {
         // Setup the TLS, if needed.
         if (type == ThreadType::User) {
-            tls_address = owner->CreateTLSRegion();
+            R_TRY(owner->CreateThreadLocalRegion(std::addressof(tls_address)));
         }
 
         parent = owner;
@@ -305,7 +305,7 @@ void KThread::Finalize() {
 
     // If the thread has a local region, delete it.
     if (tls_address != 0) {
-        parent->FreeTLSRegion(tls_address);
+        ASSERT(parent->DeleteThreadLocalRegion(tls_address).IsSuccess());
     }
 
     // Release any waiters.