using Ryujinx.Common.Memory;
using Ryujinx.Memory;
using Ryujinx.Memory.Range;
using System;
using System.Buffers;
using System.Collections.Generic;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
namespace Ryujinx.Graphics.Gpu.Memory
{
///
/// GPU memory manager.
///
public class MemoryManager : IWritableBlock
{
private const int PtLvl0Bits = 14;
private const int PtLvl1Bits = 14;
public const int PtPageBits = 12;
private const ulong PtLvl0Size = 1UL << PtLvl0Bits;
private const ulong PtLvl1Size = 1UL << PtLvl1Bits;
public const ulong PageSize = 1UL << PtPageBits;
private const ulong PtLvl0Mask = PtLvl0Size - 1;
private const ulong PtLvl1Mask = PtLvl1Size - 1;
public const ulong PageMask = PageSize - 1;
private const int PtLvl0Bit = PtPageBits + PtLvl1Bits;
private const int PtLvl1Bit = PtPageBits;
private const int AddressSpaceBits = PtPageBits + PtLvl1Bits + PtLvl0Bits;
public const ulong PteUnmapped = ulong.MaxValue;
private readonly ulong[][] _pageTable;
public event EventHandler MemoryUnmapped;
///
/// Physical memory where the virtual memory is mapped into.
///
internal PhysicalMemory Physical { get; }
///
/// Virtual range cache.
///
internal VirtualRangeCache VirtualRangeCache { get; }
///
/// Cache of GPU counters.
///
internal CounterCache CounterCache { get; }
///
/// Creates a new instance of the GPU memory manager.
///
/// Physical memory that this memory manager will map into
internal MemoryManager(PhysicalMemory physicalMemory)
{
Physical = physicalMemory;
VirtualRangeCache = new VirtualRangeCache(this);
CounterCache = new CounterCache();
_pageTable = new ulong[PtLvl0Size][];
MemoryUnmapped += Physical.TextureCache.MemoryUnmappedHandler;
MemoryUnmapped += Physical.BufferCache.MemoryUnmappedHandler;
MemoryUnmapped += VirtualRangeCache.MemoryUnmappedHandler;
MemoryUnmapped += CounterCache.MemoryUnmappedHandler;
}
///
/// Reads data from GPU mapped memory.
///
/// Type of the data
/// GPU virtual address where the data is located
/// True if read tracking is triggered on the memory region
/// The data at the specified memory location
public T Read(ulong va, bool tracked = false) where T : unmanaged
{
int size = Unsafe.SizeOf();
if (IsContiguous(va, size))
{
ulong address = Translate(va);
if (tracked)
{
return Physical.ReadTracked(address);
}
else
{
return Physical.Read(address);
}
}
else
{
Span data = new byte[size];
ReadImpl(va, data, tracked);
return MemoryMarshal.Cast(data)[0];
}
}
///
/// Gets a read-only span of data from GPU mapped memory.
///
/// GPU virtual address where the data is located
/// Size of the data
/// True if read tracking is triggered on the span
/// The span of the data at the specified memory location
public ReadOnlySpan GetSpan(ulong va, int size, bool tracked = false)
{
if (IsContiguous(va, size))
{
return Physical.GetSpan(Translate(va), size, tracked);
}
else
{
Span data = new byte[size];
ReadImpl(va, data, tracked);
return data;
}
}
///
/// Gets a read-only span of data from GPU mapped memory, up to the entire range specified,
/// or the last mapped page if the range is not fully mapped.
///
/// GPU virtual address where the data is located
/// Size of the data
/// True if read tracking is triggered on the span
/// The span of the data at the specified memory location
public ReadOnlySpan GetSpanMapped(ulong va, int size, bool tracked = false)
{
bool isContiguous = true;
int mappedSize;
if (ValidateAddress(va) && GetPte(va) != PteUnmapped && Physical.IsMapped(Translate(va)))
{
ulong endVa = va + (ulong)size;
ulong endVaAligned = (endVa + PageMask) & ~PageMask;
ulong currentVa = va & ~PageMask;
int pages = (int)((endVaAligned - currentVa) / PageSize);
for (int page = 0; page < pages - 1; page++)
{
ulong nextVa = currentVa + PageSize;
ulong nextPa = Translate(nextVa);
if (!ValidateAddress(nextVa) || GetPte(nextVa) == PteUnmapped || !Physical.IsMapped(nextPa))
{
break;
}
if (Translate(currentVa) + PageSize != nextPa)
{
isContiguous = false;
}
currentVa += PageSize;
}
currentVa += PageSize;
if (currentVa > endVa)
{
currentVa = endVa;
}
mappedSize = (int)(currentVa - va);
}
else
{
return ReadOnlySpan.Empty;
}
if (isContiguous)
{
return Physical.GetSpan(Translate(va), mappedSize, tracked);
}
else
{
Span data = new byte[mappedSize];
ReadImpl(va, data, tracked);
return data;
}
}
///
/// Reads data from a possibly non-contiguous region of GPU mapped memory.
///
/// GPU virtual address of the data
/// Span to write the read data into
/// True to enable write tracking on read, false otherwise
private void ReadImpl(ulong va, Span data, bool tracked)
{
if (data.Length == 0)
{
return;
}
int offset = 0, size;
if ((va & PageMask) != 0)
{
ulong pa = Translate(va);
size = Math.Min(data.Length, (int)PageSize - (int)(va & PageMask));
Physical.GetSpan(pa, size, tracked).CopyTo(data[..size]);
offset += size;
}
for (; offset < data.Length; offset += size)
{
ulong pa = Translate(va + (ulong)offset);
size = Math.Min(data.Length - offset, (int)PageSize);
Physical.GetSpan(pa, size, tracked).CopyTo(data.Slice(offset, size));
}
}
///
/// Gets a writable region from GPU mapped memory.
///
/// Start address of the range
/// Size in bytes to be range
/// True if write tracking is triggered on the span
/// A writable region with the data at the specified memory location
public WritableRegion GetWritableRegion(ulong va, int size, bool tracked = false)
{
if (IsContiguous(va, size))
{
return Physical.GetWritableRegion(Translate(va), size, tracked);
}
else
{
IMemoryOwner memoryOwner = ByteMemoryPool.Rent(size);
GetSpan(va, size).CopyTo(memoryOwner.Memory.Span);
return new WritableRegion(this, va, memoryOwner, tracked);
}
}
///
/// Writes data to GPU mapped memory.
///
/// Type of the data
/// GPU virtual address to write the value into
/// The value to be written
public void Write(ulong va, T value) where T : unmanaged
{
Write(va, MemoryMarshal.Cast(MemoryMarshal.CreateSpan(ref value, 1)));
}
///
/// Writes data to GPU mapped memory.
///
/// GPU virtual address to write the data into
/// The data to be written
public void Write(ulong va, ReadOnlySpan data)
{
WriteImpl(va, data, Physical.Write);
}
///
/// Writes data to GPU mapped memory, destined for a tracked resource.
///
/// GPU virtual address to write the data into
/// The data to be written
public void WriteTrackedResource(ulong va, ReadOnlySpan data)
{
WriteImpl(va, data, Physical.WriteTrackedResource);
}
///
/// Writes data to GPU mapped memory without write tracking.
///
/// GPU virtual address to write the data into
/// The data to be written
public void WriteUntracked(ulong va, ReadOnlySpan data)
{
WriteImpl(va, data, Physical.WriteUntracked);
}
private delegate void WriteCallback(ulong address, ReadOnlySpan data);
///
/// Writes data to possibly non-contiguous GPU mapped memory.
///
/// GPU virtual address of the region to write into
/// Data to be written
/// Write callback
private void WriteImpl(ulong va, ReadOnlySpan data, WriteCallback writeCallback)
{
if (IsContiguous(va, data.Length))
{
writeCallback(Translate(va), data);
}
else
{
int offset = 0, size;
if ((va & PageMask) != 0)
{
ulong pa = Translate(va);
size = Math.Min(data.Length, (int)PageSize - (int)(va & PageMask));
writeCallback(pa, data[..size]);
offset += size;
}
for (; offset < data.Length; offset += size)
{
ulong pa = Translate(va + (ulong)offset);
size = Math.Min(data.Length - offset, (int)PageSize);
writeCallback(pa, data.Slice(offset, size));
}
}
}
///
/// Runs remap actions that are added to an unmap event.
/// These must run after the mapping completes.
///
/// Event with remap actions
private static void RunRemapActions(UnmapEventArgs e)
{
if (e.RemapActions != null)
{
foreach (Action action in e.RemapActions)
{
action();
}
}
}
///
/// Maps a given range of pages to the specified CPU virtual address.
///
///
/// All addresses and sizes must be page aligned.
///
/// CPU virtual address to map into
/// GPU virtual address to be mapped
/// Size in bytes of the mapping
/// Kind of the resource located at the mapping
public void Map(ulong pa, ulong va, ulong size, PteKind kind)
{
lock (_pageTable)
{
UnmapEventArgs e = new(va, size);
MemoryUnmapped?.Invoke(this, e);
for (ulong offset = 0; offset < size; offset += PageSize)
{
SetPte(va + offset, PackPte(pa + offset, kind));
}
RunRemapActions(e);
}
}
///
/// Unmaps a given range of pages at the specified GPU virtual memory region.
///
/// GPU virtual address to unmap
/// Size in bytes of the region being unmapped
public void Unmap(ulong va, ulong size)
{
lock (_pageTable)
{
// Event handlers are not expected to be thread safe.
UnmapEventArgs e = new(va, size);
MemoryUnmapped?.Invoke(this, e);
for (ulong offset = 0; offset < size; offset += PageSize)
{
SetPte(va + offset, PteUnmapped);
}
RunRemapActions(e);
}
}
///
/// Checks if a region of GPU mapped memory is contiguous.
///
/// GPU virtual address of the region
/// Size of the region
/// True if the region is contiguous, false otherwise
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private bool IsContiguous(ulong va, int size)
{
if (!ValidateAddress(va) || GetPte(va) == PteUnmapped)
{
return false;
}
ulong endVa = (va + (ulong)size + PageMask) & ~PageMask;
va &= ~PageMask;
int pages = (int)((endVa - va) / PageSize);
for (int page = 0; page < pages - 1; page++)
{
if (!ValidateAddress(va + PageSize) || GetPte(va + PageSize) == PteUnmapped)
{
return false;
}
if (Translate(va) + PageSize != Translate(va + PageSize))
{
return false;
}
va += PageSize;
}
return true;
}
///
/// Gets the physical regions that make up the given virtual address region.
///
/// Virtual address of the range
/// Size of the range
/// Multi-range with the physical regions
public MultiRange GetPhysicalRegions(ulong va, ulong size)
{
if (IsContiguous(va, (int)size))
{
return new MultiRange(Translate(va), size);
}
ulong regionStart = Translate(va);
ulong regionSize = Math.Min(size, PageSize - (va & PageMask));
ulong endVa = va + size;
ulong endVaRounded = (endVa + PageMask) & ~PageMask;
va &= ~PageMask;
int pages = (int)((endVaRounded - va) / PageSize);
var regions = new List();
for (int page = 0; page < pages - 1; page++)
{
ulong currPa = Translate(va);
ulong newPa = Translate(va + PageSize);
if ((currPa != PteUnmapped || newPa != PteUnmapped) && currPa + PageSize != newPa)
{
regions.Add(new MemoryRange(regionStart, regionSize));
regionStart = newPa;
regionSize = 0;
}
va += PageSize;
regionSize += Math.Min(endVa - va, PageSize);
}
if (regions.Count == 0)
{
return new MultiRange(regionStart, regionSize);
}
regions.Add(new MemoryRange(regionStart, regionSize));
return new MultiRange(regions.ToArray());
}
///
/// Checks if a given GPU virtual memory range is mapped to the same physical regions
/// as the specified physical memory multi-range.
///
/// Physical memory multi-range
/// GPU virtual memory address
/// True if the virtual memory region is mapped into the specified physical one, false otherwise
public bool CompareRange(MultiRange range, ulong va)
{
va &= ~PageMask;
for (int i = 0; i < range.Count; i++)
{
MemoryRange currentRange = range.GetSubRange(i);
if (currentRange.Address != PteUnmapped)
{
ulong address = currentRange.Address & ~PageMask;
ulong endAddress = (currentRange.EndAddress + PageMask) & ~PageMask;
while (address < endAddress)
{
if (Translate(va) != address)
{
return false;
}
va += PageSize;
address += PageSize;
}
}
else
{
ulong endVa = va + (((currentRange.Size) + PageMask) & ~PageMask);
while (va < endVa)
{
if (Translate(va) != PteUnmapped)
{
return false;
}
va += PageSize;
}
}
}
return true;
}
///
/// Validates a GPU virtual address.
///
/// Address to validate
/// True if the address is valid, false otherwise
private static bool ValidateAddress(ulong va)
{
return va < (1UL << AddressSpaceBits);
}
///
/// Checks if a given page is mapped.
///
/// GPU virtual address of the page to check
/// True if the page is mapped, false otherwise
public bool IsMapped(ulong va)
{
return Translate(va) != PteUnmapped;
}
///
/// Translates a GPU virtual address to a CPU virtual address.
///
/// GPU virtual address to be translated
/// CPU virtual address, or if unmapped
public ulong Translate(ulong va)
{
if (!ValidateAddress(va))
{
return PteUnmapped;
}
ulong pte = GetPte(va);
if (pte == PteUnmapped)
{
return PteUnmapped;
}
return UnpackPaFromPte(pte) + (va & PageMask);
}
///
/// Translates a GPU virtual address to a CPU virtual address on the first mapped page of memory
/// on the specified region.
/// If no page is mapped on the specified region, is returned.
///
/// GPU virtual address to be translated
/// Size of the range to be translated
/// CPU virtual address, or if unmapped
public ulong TranslateFirstMapped(ulong va, ulong size)
{
if (!ValidateAddress(va))
{
return PteUnmapped;
}
ulong endVa = va + size;
ulong pte = GetPte(va);
for (; va < endVa && pte == PteUnmapped; va += PageSize - (va & PageMask))
{
pte = GetPte(va);
}
if (pte == PteUnmapped)
{
return PteUnmapped;
}
return UnpackPaFromPte(pte) + (va & PageMask);
}
///
/// Translates a GPU virtual address and returns the number of bytes that are mapped after it.
///
/// GPU virtual address to be translated
/// Maximum size in bytes to scan
/// Number of bytes, 0 if unmapped
public ulong GetMappedSize(ulong va, ulong maxSize)
{
if (!ValidateAddress(va))
{
return 0;
}
ulong startVa = va;
ulong endVa = va + maxSize;
ulong pte = GetPte(va);
while (pte != PteUnmapped && va < endVa)
{
va += PageSize - (va & PageMask);
pte = GetPte(va);
}
return Math.Min(maxSize, va - startVa);
}
///
/// Gets the kind of a given memory page.
/// This might indicate the type of resource that can be allocated on the page, and also texture tiling.
///
/// GPU virtual address
/// Kind of the memory page
public PteKind GetKind(ulong va)
{
if (!ValidateAddress(va))
{
return PteKind.Invalid;
}
ulong pte = GetPte(va);
if (pte == PteUnmapped)
{
return PteKind.Invalid;
}
return UnpackKindFromPte(pte);
}
///
/// Gets the Page Table entry for a given GPU virtual address.
///
/// GPU virtual address
/// Page table entry (CPU virtual address)
private ulong GetPte(ulong va)
{
ulong l0 = (va >> PtLvl0Bit) & PtLvl0Mask;
ulong l1 = (va >> PtLvl1Bit) & PtLvl1Mask;
if (_pageTable[l0] == null)
{
return PteUnmapped;
}
return _pageTable[l0][l1];
}
///
/// Sets a Page Table entry at a given GPU virtual address.
///
/// GPU virtual address
/// Page table entry (CPU virtual address)
private void SetPte(ulong va, ulong pte)
{
ulong l0 = (va >> PtLvl0Bit) & PtLvl0Mask;
ulong l1 = (va >> PtLvl1Bit) & PtLvl1Mask;
if (_pageTable[l0] == null)
{
_pageTable[l0] = new ulong[PtLvl1Size];
for (ulong index = 0; index < PtLvl1Size; index++)
{
_pageTable[l0][index] = PteUnmapped;
}
}
_pageTable[l0][l1] = pte;
}
///
/// Creates a page table entry from a physical address and kind.
///
/// Physical address
/// Kind
/// Page table entry
private static ulong PackPte(ulong pa, PteKind kind)
{
return pa | ((ulong)kind << 56);
}
///
/// Unpacks kind from a page table entry.
///
/// Page table entry
/// Kind
private static PteKind UnpackKindFromPte(ulong pte)
{
return (PteKind)(pte >> 56);
}
///
/// Unpacks physical address from a page table entry.
///
/// Page table entry
/// Physical address
private static ulong UnpackPaFromPte(ulong pte)
{
return pte & 0xffffffffffffffUL;
}
}
}