Merge bcbe589964 into 6ab899f621

Instead of one big 2048MB JIT Cache that'd crash the emulator when maxed
out, we now have it where we add 256MB JIT Cache regions when needed,
helping reduce allocated memory where games don't use the JIT cache for
it, and helping bigger games that DO need JIT cache bigger than 2048MB!

![image](https://github.com/user-attachments/assets/ff17dc48-6028-4377-8c73-746ab21ab83b)
(SSBU goes past the 2048MB JIT Cache limit that would normally crash
Ryujinx ^)

Also I added a BSD socket that Baba is You's networking for downloading
custom levels uses.

src/ARMeilleure/Memory/ReservedRegion.cs

@@ -7,6 +7,7 @@ namespace ARMeilleure.Memory
         public const int DefaultGranularity = 65536; // Mapping granularity in Windows.
 
         public IJitMemoryBlock Block { get; }
+        public IJitMemoryAllocator Allocator { get; }
 
         public nint Pointer => Block.Pointer;
 
@@ -21,6 +22,7 @@ namespace ARMeilleure.Memory
                 granularity = DefaultGranularity;
             }
 
+            Allocator = allocator;
             Block = allocator.Reserve(maxSize);
             _maxSize = maxSize;
             _sizeGranularity = granularity;

src/ARMeilleure/Translation/Cache/JitCache.cs

@@ -2,6 +2,8 @@ using ARMeilleure.CodeGen;
 using ARMeilleure.CodeGen.Unwinding;
 using ARMeilleure.Memory;
 using ARMeilleure.Native;
+using Humanizer;
+using Ryujinx.Common.Logging;
 using Ryujinx.Memory;
 using System;
 using System.Collections.Generic;
@@ -18,9 +20,8 @@ namespace ARMeilleure.Translation.Cache
         private static readonly int _pageMask = _pageSize - 1;
 
         private const int CodeAlignment = 4; // Bytes.
-        private const int CacheSize = 2047 * 1024 * 1024;
+        private const int CacheSize = 256 * 1024 * 1024;
 
-        private static ReservedRegion _jitRegion;
         private static JitCacheInvalidation _jitCacheInvalidator;
 
         private static CacheMemoryAllocator _cacheAllocator;
@@ -30,6 +31,9 @@ namespace ARMeilleure.Translation.Cache
         private static readonly Lock _lock = new();
         private static bool _initialized;
 
+        private static readonly List<ReservedRegion> _jitRegions = new();
+        private static int _activeRegionIndex = 0;
+
         [SupportedOSPlatform("windows")]
         [LibraryImport("kernel32.dll", SetLastError = true)]
         public static partial nint FlushInstructionCache(nint hProcess, nint lpAddress, nuint dwSize);
@@ -48,7 +52,9 @@ namespace ARMeilleure.Translation.Cache
                     return;
                 }
 
-                _jitRegion = new ReservedRegion(allocator, CacheSize);
+                ReservedRegion firstRegion = new(allocator, CacheSize);
+                _jitRegions.Add(firstRegion);
+                _activeRegionIndex = 0;
 
                 if (!OperatingSystem.IsWindows() && !OperatingSystem.IsMacOS())
                 {
@@ -59,7 +65,9 @@ namespace ARMeilleure.Translation.Cache
 
                 if (OperatingSystem.IsWindows())
                 {
-                    JitUnwindWindows.InstallFunctionTableHandler(_jitRegion.Pointer, CacheSize, _jitRegion.Pointer + Allocate(_pageSize));
+                    JitUnwindWindows.InstallFunctionTableHandler(
+                        firstRegion.Pointer, CacheSize, firstRegion.Pointer + Allocate(_pageSize)
+                    );
                 }
 
                 _initialized = true;
@@ -75,8 +83,8 @@ namespace ARMeilleure.Translation.Cache
                 Debug.Assert(_initialized);
 
                 int funcOffset = Allocate(code.Length);
-
-                nint funcPtr = _jitRegion.Pointer + funcOffset;
+                ReservedRegion targetRegion = _jitRegions[_activeRegionIndex];
+                nint funcPtr = targetRegion.Pointer + funcOffset;
 
                 if (OperatingSystem.IsMacOS() && RuntimeInformation.ProcessArchitecture == Architecture.Arm64)
                 {
@@ -90,9 +98,9 @@ namespace ARMeilleure.Translation.Cache
                 }
                 else
                 {
-                    ReprotectAsWritable(funcOffset, code.Length);
+                    ReprotectAsWritable(targetRegion, funcOffset, code.Length);
                     Marshal.Copy(code, 0, funcPtr, code.Length);
-                    ReprotectAsExecutable(funcOffset, code.Length);
+                    ReprotectAsExecutable(targetRegion, funcOffset, code.Length);
 
                     if (OperatingSystem.IsWindows() && RuntimeInformation.ProcessArchitecture == Architecture.Arm64)
                     {
@@ -116,52 +124,83 @@ namespace ARMeilleure.Translation.Cache
             {
                 Debug.Assert(_initialized);
 
-                int funcOffset = (int)(pointer.ToInt64() - _jitRegion.Pointer.ToInt64());
-
-                if (TryFind(funcOffset, out CacheEntry entry, out int entryIndex) && entry.Offset == funcOffset)
+                foreach (ReservedRegion region in _jitRegions)
                 {
-                    _cacheAllocator.Free(funcOffset, AlignCodeSize(entry.Size));
-                    _cacheEntries.RemoveAt(entryIndex);
+                    if (pointer.ToInt64() < region.Pointer.ToInt64() ||
+                        pointer.ToInt64() >= (region.Pointer + CacheSize).ToInt64())
+                    {
+                        continue;
+                    }
+
+                    int funcOffset = (int)(pointer.ToInt64() - region.Pointer.ToInt64());
+
+                    if (TryFind(funcOffset, out CacheEntry entry, out int entryIndex) && entry.Offset == funcOffset)
+                    {
+                        _cacheAllocator.Free(funcOffset, AlignCodeSize(entry.Size));
+                        _cacheEntries.RemoveAt(entryIndex);
+                    }
+
+                    return;
                 }
             }
         }
 
-        private static void ReprotectAsWritable(int offset, int size)
+        private static void ReprotectAsWritable(ReservedRegion region, int offset, int size)
         {
             int endOffs = offset + size;
-
             int regionStart = offset & ~_pageMask;
             int regionEnd = (endOffs + _pageMask) & ~_pageMask;
 
-            _jitRegion.Block.MapAsRwx((ulong)regionStart, (ulong)(regionEnd - regionStart));
+            region.Block.MapAsRwx((ulong)regionStart, (ulong)(regionEnd - regionStart));
         }
 
-        private static void ReprotectAsExecutable(int offset, int size)
+        private static void ReprotectAsExecutable(ReservedRegion region, int offset, int size)
         {
             int endOffs = offset + size;
-
             int regionStart = offset & ~_pageMask;
             int regionEnd = (endOffs + _pageMask) & ~_pageMask;
 
-            _jitRegion.Block.MapAsRx((ulong)regionStart, (ulong)(regionEnd - regionStart));
+            region.Block.MapAsRx((ulong)regionStart, (ulong)(regionEnd - regionStart));
         }
 
         private static int Allocate(int codeSize)
         {
             codeSize = AlignCodeSize(codeSize);
 
-            int allocOffset = _cacheAllocator.Allocate(codeSize);
-
-            if (allocOffset < 0)
+            for (int i = _activeRegionIndex; i < _jitRegions.Count; i++)
             {
-                throw new OutOfMemoryException("JIT Cache exhausted.");
+                int allocOffset = _cacheAllocator.Allocate(codeSize);
+        
+                if (allocOffset >= 0)
+                {
+                    _jitRegions[i].ExpandIfNeeded((ulong)allocOffset + (ulong)codeSize);
+                    _activeRegionIndex = i;
+                    return allocOffset;
+                }
             }
 
-            _jitRegion.ExpandIfNeeded((ulong)allocOffset + (ulong)codeSize);
+            int exhaustedRegion = _activeRegionIndex;
+            var newRegion = new ReservedRegion(_jitRegions[0].Allocator, CacheSize);
+            _jitRegions.Add(newRegion);
+            _activeRegionIndex = _jitRegions.Count - 1;
+            
+            int newRegionNumber = _activeRegionIndex;
 
-            return allocOffset;
+            Logger.Warning?.Print(LogClass.Cpu, $"JIT Cache Region {exhaustedRegion} exhausted, creating new Cache Region {newRegionNumber} ({((newRegionNumber + 1) * CacheSize).Bytes()} Total Allocation).");
+        
+            _cacheAllocator = new CacheMemoryAllocator(CacheSize);
+
+            int allocOffsetNew = _cacheAllocator.Allocate(codeSize);
+            if (allocOffsetNew < 0)
+            {
+                throw new OutOfMemoryException("Failed to allocate in new Cache Region!");
+            }
+
+            newRegion.ExpandIfNeeded((ulong)allocOffsetNew + (ulong)codeSize);
+            return allocOffsetNew;
         }
 
+
         private static int AlignCodeSize(int codeSize)
         {
             return checked(codeSize + (CodeAlignment - 1)) & ~(CodeAlignment - 1);
@@ -185,18 +224,21 @@ namespace ARMeilleure.Translation.Cache
         {
             lock (_lock)
             {
-                int index = _cacheEntries.BinarySearch(new CacheEntry(offset, 0, default));
-
-                if (index < 0)
+                foreach (ReservedRegion _ in _jitRegions)
                 {
-                    index = ~index - 1;
-                }
+                    int index = _cacheEntries.BinarySearch(new CacheEntry(offset, 0, default));
 
-                if (index >= 0)
-                {
-                    entry = _cacheEntries[index];
-                    entryIndex = index;
-                    return true;
+                    if (index < 0)
+                    {
+                        index = ~index - 1;
+                    }
+
+                    if (index >= 0)
+                    {
+                        entry = _cacheEntries[index];
+                        entryIndex = index;
+                        return true;
+                    }
                 }
             }
 

src/ARMeilleure/Translation/PTC/PtcProfiler.cs

@@ -144,17 +144,15 @@ namespace ARMeilleure.Translation.PTC
 
         public List<ulong> GetBlacklistedFunctions()
         {
-            List<ulong> funcs = new List<ulong>();
+            List<ulong> funcs = [];
 
-            foreach (var profiledFunc in ProfiledFuncs)
+            foreach ((ulong ptr, FuncProfile funcProfile) in ProfiledFuncs)
             {
-                if (profiledFunc.Value.Blacklist)
-                {
-                    if (!funcs.Contains(profiledFunc.Key))
-                    {
-                        funcs.Add(profiledFunc.Key);
-                    }
-                }
+                if (!funcProfile.Blacklist)
+                    continue;
+
+                if (!funcs.Contains(ptr)) 
+                    funcs.Add(ptr);
             }
 
             return funcs;

src/Ryujinx.Cpu/LightningJit/Cache/JitCache.cs

@@ -1,4 +1,6 @@
 using ARMeilleure.Memory;
+using Humanizer;
+using Ryujinx.Common.Logging;
 using Ryujinx.Memory;
 using System;
 using System.Collections.Generic;
@@ -15,9 +17,8 @@ namespace Ryujinx.Cpu.LightningJit.Cache
         private static readonly int _pageMask = _pageSize - 1;
 
         private const int CodeAlignment = 4; // Bytes.
-        private const int CacheSize = 2047 * 1024 * 1024;
+        private const int CacheSize = 256 * 1024 * 1024;
 
-        private static ReservedRegion _jitRegion;
         private static JitCacheInvalidation _jitCacheInvalidator;
 
         private static CacheMemoryAllocator _cacheAllocator;
@@ -26,6 +27,8 @@ namespace Ryujinx.Cpu.LightningJit.Cache
 
         private static readonly Lock _lock = new();
         private static bool _initialized;
+        private static readonly List<ReservedRegion> _jitRegions = new();
+        private static int _activeRegionIndex = 0;
 
         [SupportedOSPlatform("windows")]
         [LibraryImport("kernel32.dll", SetLastError = true)]
@@ -45,7 +48,9 @@ namespace Ryujinx.Cpu.LightningJit.Cache
                     return;
                 }
 
-                _jitRegion = new ReservedRegion(allocator, CacheSize);
+                ReservedRegion firstRegion = new(allocator, CacheSize);
+                _jitRegions.Add(firstRegion);
+                _activeRegionIndex = 0;
 
                 if (!OperatingSystem.IsWindows() && !OperatingSystem.IsMacOS())
                 {
@@ -65,8 +70,8 @@ namespace Ryujinx.Cpu.LightningJit.Cache
                 Debug.Assert(_initialized);
 
                 int funcOffset = Allocate(code.Length);
-
-                nint funcPtr = _jitRegion.Pointer + funcOffset;
+                ReservedRegion targetRegion = _jitRegions[_activeRegionIndex];
+                nint funcPtr = targetRegion.Pointer + funcOffset;
 
                 if (OperatingSystem.IsMacOS() && RuntimeInformation.ProcessArchitecture == Architecture.Arm64)
                 {
@@ -80,18 +85,11 @@ namespace Ryujinx.Cpu.LightningJit.Cache
                 }
                 else
                 {
-                    ReprotectAsWritable(funcOffset, code.Length);
-                    code.CopyTo(new Span<byte>((void*)funcPtr, code.Length));
-                    ReprotectAsExecutable(funcOffset, code.Length);
+                    ReprotectAsWritable(targetRegion, funcOffset, code.Length);
+                    Marshal.Copy(code.ToArray(), 0, funcPtr, code.Length);
+                    ReprotectAsExecutable(targetRegion, funcOffset, code.Length);
 
-                    if (OperatingSystem.IsWindows() && RuntimeInformation.ProcessArchitecture == Architecture.Arm64)
-                    {
-                        FlushInstructionCache(Process.GetCurrentProcess().Handle, funcPtr, (nuint)code.Length);
-                    }
-                    else
-                    {
-                        _jitCacheInvalidator?.Invalidate(funcPtr, (ulong)code.Length);
-                    }
+                    _jitCacheInvalidator?.Invalidate(funcPtr, (ulong)code.Length);
                 }
 
                 Add(funcOffset, code.Length);
@@ -106,50 +104,80 @@ namespace Ryujinx.Cpu.LightningJit.Cache
             {
                 Debug.Assert(_initialized);
 
-                int funcOffset = (int)(pointer.ToInt64() - _jitRegion.Pointer.ToInt64());
-
-                if (TryFind(funcOffset, out CacheEntry entry, out int entryIndex) && entry.Offset == funcOffset)
+                foreach (ReservedRegion region in _jitRegions)
                 {
-                    _cacheAllocator.Free(funcOffset, AlignCodeSize(entry.Size));
-                    _cacheEntries.RemoveAt(entryIndex);
+                    if (pointer.ToInt64() < region.Pointer.ToInt64() ||
+                        pointer.ToInt64() >= (region.Pointer + CacheSize).ToInt64())
+                    {
+                        continue;
+                    }
+
+                    int funcOffset = (int)(pointer.ToInt64() - region.Pointer.ToInt64());
+
+                    if (TryFind(funcOffset, out CacheEntry entry, out int entryIndex) && entry.Offset == funcOffset)
+                    {
+                        _cacheAllocator.Free(funcOffset, AlignCodeSize(entry.Size));
+                        _cacheEntries.RemoveAt(entryIndex);
+                    }
+
+                    return;
                 }
             }
         }
 
-        private static void ReprotectAsWritable(int offset, int size)
+        private static void ReprotectAsWritable(ReservedRegion region, int offset, int size)
         {
             int endOffs = offset + size;
-
             int regionStart = offset & ~_pageMask;
             int regionEnd = (endOffs + _pageMask) & ~_pageMask;
 
-            _jitRegion.Block.MapAsRwx((ulong)regionStart, (ulong)(regionEnd - regionStart));
+            region.Block.MapAsRwx((ulong)regionStart, (ulong)(regionEnd - regionStart));
         }
 
-        private static void ReprotectAsExecutable(int offset, int size)
+        private static void ReprotectAsExecutable(ReservedRegion region, int offset, int size)
         {
             int endOffs = offset + size;
-
             int regionStart = offset & ~_pageMask;
             int regionEnd = (endOffs + _pageMask) & ~_pageMask;
 
-            _jitRegion.Block.MapAsRx((ulong)regionStart, (ulong)(regionEnd - regionStart));
+            region.Block.MapAsRx((ulong)regionStart, (ulong)(regionEnd - regionStart));
         }
 
         private static int Allocate(int codeSize)
         {
             codeSize = AlignCodeSize(codeSize);
 
-            int allocOffset = _cacheAllocator.Allocate(codeSize);
-
-            if (allocOffset < 0)
+            for (int i = _activeRegionIndex; i < _jitRegions.Count; i++)
             {
-                throw new OutOfMemoryException("JIT Cache exhausted.");
+                int allocOffset = _cacheAllocator.Allocate(codeSize);
+        
+                if (allocOffset >= 0)
+                {
+                    _jitRegions[i].ExpandIfNeeded((ulong)allocOffset + (ulong)codeSize);
+                    _activeRegionIndex = i;
+                    return allocOffset;
+                }
             }
 
-            _jitRegion.ExpandIfNeeded((ulong)allocOffset + (ulong)codeSize);
+            int exhaustedRegion = _activeRegionIndex;
+            ReservedRegion newRegion = new(_jitRegions[0].Allocator, CacheSize);
+            _jitRegions.Add(newRegion);
+            _activeRegionIndex = _jitRegions.Count - 1;
+            
+            int newRegionNumber = _activeRegionIndex;
 
-            return allocOffset;
+            Logger.Warning?.Print(LogClass.Cpu, $"JIT Cache Region {exhaustedRegion} exhausted, creating new Cache Region {newRegionNumber} ({((newRegionNumber + 1) * CacheSize).Bytes()} Total Allocation).");
+        
+            _cacheAllocator = new CacheMemoryAllocator(CacheSize);
+
+            int allocOffsetNew = _cacheAllocator.Allocate(codeSize);
+            if (allocOffsetNew < 0)
+            {
+                throw new OutOfMemoryException("Failed to allocate in new Cache Region!");
+            }
+
+            newRegion.ExpandIfNeeded((ulong)allocOffsetNew + (ulong)codeSize);
+            return allocOffsetNew;
         }
 
         private static int AlignCodeSize(int codeSize)

src/Ryujinx.Cpu/LightningJit/Cache/NoWxCache.cs

@@ -12,7 +12,7 @@ namespace Ryujinx.Cpu.LightningJit.Cache
     {
         private const int CodeAlignment = 4; // Bytes.
         private const int SharedCacheSize = 2047 * 1024 * 1024;
-        private const int LocalCacheSize = 128 * 1024 * 1024;
+        private const int LocalCacheSize = 256 * 1024 * 1024;
 
         // How many calls to the same function we allow until we pad the shared cache to force the function to become available there
         // and allow the guest to take the fast path.

src/Ryujinx.HLE/HOS/Services/Ldn/UserServiceCreator/LdnRyu/Proxy/P2pProxyServer.cs

@@ -110,24 +110,10 @@ namespace Ryujinx.HLE.HOS.Services.Ldn.UserServiceCreator.LdnRyu.Proxy
             _broadcastAddress = config.ProxyIp | (~config.ProxySubnetMask);
         }
 
-        public async Task<ushort> NatPunch()
+        private async Task<ushort> NatPunchForDevice(NatDevice device)
         {
-            NatDiscoverer discoverer = new();
-            CancellationTokenSource cts = new(5000);
-
-            NatDevice device;
-
-            try
-            {
-                device = await discoverer.DiscoverDeviceAsync(PortMapper.Upnp, cts);
-            }
-            catch (NatDeviceNotFoundException)
-            {
-                return 0;
-            }
-
+            Logger.Info?.PrintMsg(LogClass.ServiceLdn, $"Attempting to map port using {device.ToString()}");
             _publicPort = PublicPortBase;
-
             for (int i = 0; i < PublicPortRange; i++)
             {
                 try
@@ -138,12 +124,14 @@ namespace Ryujinx.HLE.HOS.Services.Ldn.UserServiceCreator.LdnRyu.Proxy
 
                     break;
                 }
-                catch (MappingException)
+                catch (MappingException ex)
                 {
+                    Logger.Info?.PrintMsg(LogClass.ServiceLdn, $"Failed to map port {_publicPort}: {ex.Message}");
                     _publicPort++;
                 }
-                catch (Exception)
+                catch (Exception ex)
                 {
+                    Logger.Info?.PrintMsg(LogClass.ServiceLdn, $"Failed to map port {_publicPort}: {ex.GetType().Name}: {ex.Message}");
                     return 0;
                 }
 
@@ -155,17 +143,52 @@ namespace Ryujinx.HLE.HOS.Services.Ldn.UserServiceCreator.LdnRyu.Proxy
 
             if (_publicPort != 0)
             {
+                _natDevice = device;
                 _ = Executor.ExecuteAfterDelayAsync(
-                    PortLeaseRenew.Seconds(), 
+                    PortLeaseRenew.Seconds(),
                     _disposedCancellation.Token,
                     RefreshLease);
             }
 
-            _natDevice = device;
-
             return _publicPort;
         }
 
+        public async Task<ushort> NatPunch()
+        {
+            NatDiscoverer discoverer = new();
+            CancellationTokenSource cts = new(500);
+
+            NatDevice[] devices;
+
+            try
+            {
+                devices = (await discoverer.DiscoverDevicesAsync(PortMapper.Upnp, cts)).ToArray();
+            }
+            catch (Exception ex)
+            {
+                Logger.Error?.PrintMsg(LogClass.ServiceLdn, $"Failed to discover UPnP devices: {ex.Message}");
+                return 0;
+            }
+
+            if (devices.Length == 0)
+            {
+                Logger.Error?.PrintMsg(LogClass.ServiceLdn, "No UPnP devices found.");
+                return 0;
+            }
+
+            foreach (var device in devices)
+            {
+                ushort port = await NatPunchForDevice(device);
+                if (port != 0)
+                {
+                    return port;
+                }
+            }
+            Logger.Info?.PrintMsg(LogClass.ServiceLdn, $"Failed to map port using any device");
+            _publicPort = 0;
+            return 0;
+        }
+
         // Proxy handlers
 
         private void RouteMessage(P2pProxySession sender, ref ProxyInfo info, Action<P2pProxySession> action)

src/Ryujinx.HLE/HOS/Services/Sockets/Bsd/Impl/WinSockHelper.cs

@@ -150,6 +150,7 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd.Impl
             { BsdSocketOption.SoLinger,      SocketOptionName.Linger },
             { BsdSocketOption.SoOobInline,   SocketOptionName.OutOfBandInline },
             { BsdSocketOption.SoReusePort,   SocketOptionName.ReuseAddress },
+            { BsdSocketOption.SoNoSigpipe,   SocketOptionName.DontLinger },
             { BsdSocketOption.SoSndBuf,      SocketOptionName.SendBuffer },
             { BsdSocketOption.SoRcvBuf,      SocketOptionName.ReceiveBuffer },
             { BsdSocketOption.SoSndLoWat,    SocketOptionName.SendLowWater },

src/Ryujinx/Utilities/PlayReport/Specs.cs

@@ -152,7 +152,7 @@ namespace Ryujinx.Ava.Utilities.PlayReport
         public override bool GetData(Horizon.Prepo.Types.PlayReport playReport, out object result)
         {
             List<MessagePackObject> packedObjects = [];
-            foreach (var reportKey in ReportKeys)
+            foreach (string reportKey in ReportKeys)
             {
                 if (!playReport.ReportData.AsDictionary().TryGetValue(reportKey, out MessagePackObject valuePackObject))
                 {
@@ -176,7 +176,7 @@ namespace Ryujinx.Ava.Utilities.PlayReport
         public override bool GetData(Horizon.Prepo.Types.PlayReport playReport, out object result)
         {
             Dictionary<string, MessagePackObject> packedObjects = [];
-            foreach (var reportKey in ReportKeys)
+            foreach (string reportKey in ReportKeys)
             {
                 if (!playReport.ReportData.AsDictionary().TryGetValue(reportKey, out MessagePackObject valuePackObject))
                     continue;