//////////////////////////////////////////////////////////////////////////////// // // Microsoft Research Singularity // // Copyright (c) Microsoft Corporation. All rights reserved. // // File: VMManager.cs -- Virtual Memory Manager // // Notes: // // This module manages page-translation tables for virtual memory // using System; using System.Runtime.CompilerServices; using System.Runtime.InteropServices; using System.Threading; using Microsoft.Singularity; using Microsoft.Singularity.Hal; using Microsoft.Singularity.Isal.IX; using PageType = System.GCs.PageType; namespace Microsoft.Singularity.Memory { [NoCCtor] [CLSCompliant(false)] [AccessedByRuntime("referenced from halkd.cpp")] public class VMManager { // // NOTES // // Currently, this class is hand-written to deal with 4KB paging // on the AMD x64 in 32-bit mode, with PAE enabled. The // three-tier structure of the descriptor tables is baked into // the code; we will need to factor and expand this module to // deal with 64-bit mode, or with a different processor. // // This code is also written on the assumption that the user/kernel // boundary is 1GB-aligned, such that one, two or three of the // top-level Page-Directory-Pointer-Table (PDPE) entries fully // describe the kernel space. // // Obviously, memory pages are used to describe the mapping of // virtual memory to physical memory. This gets a little confusing. // We want there to be a mapping for the pages that contain // the page tables (!) so we have virtual addresses we can // use to manipulate the page tables! // // The kernel's page table pages must always be mapped, but the // page tables for the user space must get switched when // we change address spaces. Therefore, the pages used to // describe the mappings for the kernel's space live at a // certain virtual address (kernelPageTableBase), and the // pages that describe the mapping for user-space live at // an address in the user-space (userPageTableBase). This way, // the user's page table mappings get swapped along with the // rest of the user address space. // // A minimally usable address space must include a PDPT, and // PDTs and PTs sufficient to be able to map in all page- // table structures necessary to describe the space, including // user-space if the user-space range will be used. // // Additionally, it is a tricky stunt to set up a new address // space when paging is already turned on, since it is difficult // to "talk about" the page-table pages for the new space. // There are two pages of reserved VM space at scratchAddress1/2 // that are used to temporarily map in single pages of // "foreign" memory so we can build a new address space without // having to turn off paging. // // These are the PDPT entries for the kernel space. Whether we // are using one, two or three of them is determined by // Platform.KERNEL_BOUNDARY private static PhysicalAddress kernelPDPE1; private static PhysicalAddress kernelPDPE2; private static PhysicalAddress kernelPDPE3; // How many of the kernelPDPEs are we using? private static uint numKernelPDPEs; // The address space we create during boot private static AddressSpace BootstrapSpace; // Here is the limit below which memory is identity-mapped private static UIntPtr identityMapLimit; // Here is the virtual address where the page-table pages that // describe the kernel's address space are always accessible private static UIntPtr kernelPageTableBase; private static UIntPtr kernelPageTableLimit; private static UIntPtr kernelPDTOffset; // Here is the same thing for the user space. This address is // above KERNEL_BOUNDARY. private static UIntPtr userPageTableBase; private static UIntPtr userPageTableLimit; private static UIntPtr userPDTOffset; // The PDPT always lives in the user space, so its mapping gets // switched naturally when we change spaces private unsafe static PDPT* pdptAddr; // Here are two page of VM space we keep available for scratch use. // We must take care to use them atomically! private static UIntPtr scratchAddress1; private static UIntPtr scratchAddress2; private static SpinLock scratchSpin; // Here is where we map the uncached-memory window that // corresponds to the high-memory range that we use to // communicate with hardware (like the APIC) private static UIntPtr uncachedWindowBase; private static UIntPtr uncachedWindowSize; // Here is where we have situated the bootRecord private static UIntPtr bootRecordBase; private static UIntPtr bootRecordLimit; // Here is where we have situated the executable image that // contains the kernel (where it gets put by the hal loader) private static UIntPtr kernelImageBase; private static UIntPtr kernelImageLimit; // Here is where we have situated the executable images that // are actually in high physical memory (where they get put // by the boot loader) private static UIntPtr miniDumpBase; private static UIntPtr miniDumpLimit; // Here is where we remap areas of physical memory that are // reported as reserved by the SMAP (in the order they appear // in the SMAP) private static UIntPtr smapRemapBase; private static UIntPtr smapRemapSize; // Here are the bounds of the VM space that the kernel should use // for its GC heap private static UIntPtr kernelHeapBase; private static UIntPtr kernelHeapLimit; // Same for user-land private static UIntPtr userHeapBase; private static UIntPtr userHeapLimit; // A protective lock for the kernel's page mappings private static SmartSpinlock KernelMappingLock; ///////////////////////////////////// // CONSTANTS ///////////////////////////////////// // Each PT describes 512 4K pages private const uint SpacePerPT = 4 * 1024 * 512; // 2MB // Each PDT describes 512 PTs private const uint SpacePerPDT = 512 * SpacePerPT; // 1GB ///////////////////////////////////// // PROPERTIES ///////////////////////////////////// internal static UIntPtr KernelHeapBase { get { return kernelHeapBase; } } internal static UIntPtr KernelHeapLimit { get { return kernelHeapLimit; } } internal static UIntPtr UserHeapBase { get { return userHeapBase; } } internal static UIntPtr UserHeapLimit { get { return userHeapLimit; } } ///////////////////////////////////// // PUBLIC METHODS ///////////////////////////////////// // // This must be called with paging TURNED OFF, and it has the // side effect of TURNING PAGING ON! // // // The virtual address map used by VMManager is set up in Initialize // // Part of the layout is inheirated from MemoryManager who (ignoring SMAP // complications) has set up a range for the pageBlocks and a physicalHeap // for I/O memory below the 0xFFFFFF threshold. This function picks up // from there and lays out the rest of the virtual address space. // // TODO: Create region structure and use it here to interleave fixed address // blocks and SMAP blocks // TODO: Better checks for overlapping range requests internal static unsafe void Initialize() { Platform p = Platform.ThePlatform; uint numKernelPTs = Platform.KERNEL_BOUNDARY / SpacePerPT; UIntPtr userSpaceSize = Platform.MAX_VIRTUAL_ADDR - Platform.KERNEL_BOUNDARY + 1; UIntPtr nextVirtualAddr = MemoryManager.IOMemoryBase.Value + Platform.IO_MEMORY_SIZE; kernelPDTOffset = MemoryManager.BytesFromPages(numKernelPTs); numKernelPDPEs = (uint)(Platform.KERNEL_BOUNDARY / SpacePerPDT); scratchSpin = new SpinLock(SpinLock.Types.VMManager); // How big is the kernel space? DebugStub.Assert(MemoryManager.BytesNotAligned(Platform.KERNEL_BOUNDARY,SpacePerPDT) == 0); DebugStub.Assert(numKernelPDPEs <= 3); // First off locate the kernel and bootRecord which must live // in identity mapped memory if (p.KernelDllSize != 0) { kernelImageBase = MemoryManager.PagePad(p.KernelDllBase); kernelImageLimit = MemoryManager.PagePad(p.KernelDllBase + p.KernelDllSize); DebugStub.WriteLine("Executable Image at 0x{0:x8} - 0x{1:x8}.", __arglist(kernelImageBase, kernelImageLimit)); if (nextVirtualAddr < kernelImageLimit) { nextVirtualAddr = kernelImageLimit; } } if (p.BootAllocatedMemorySize != 0) { bootRecordBase = MemoryManager.PagePad(p.BootAllocatedMemory); bootRecordLimit = MemoryManager.PagePad( p.BootAllocatedMemory + p.BootAllocatedMemorySize); DebugStub.WriteLine("BootRecord at 0x{0:x8} - 0x{1:x8}.", __arglist(bootRecordBase, bootRecordLimit)); if (nextVirtualAddr < bootRecordLimit) { nextVirtualAddr = bootRecordLimit; } } // Now figure out where the kernel page-table pages will live. // We'll stick it immediately above the identity mapped range, on // a page boundary. identityMapLimit = MemoryManager.PagePad(nextVirtualAddr); // numKernelPDPEs == the number of PDTs, by definition kernelPageTableBase = nextVirtualAddr; nextVirtualAddr = kernelPageTableLimit = MemoryManager.PagePad(kernelPageTableBase + MemoryManager.BytesFromPages(numKernelPTs + numKernelPDPEs)); DebugStub.WriteLine("KernelPageTable initialized at 0x{0:x8} - 0x{1:x8}.", __arglist(kernelPageTableBase, kernelPageTableLimit)); // Locate the uncached window immediately after the kernel's page-table space uncachedWindowBase = nextVirtualAddr; uncachedWindowSize = Platform.UNCACHED_MAPPED; nextVirtualAddr = MemoryManager.PagePad(uncachedWindowBase + uncachedWindowSize); DebugStub.WriteLine("Uncached Window at 0x{0:x8} - 0x{1:x8}.", __arglist(uncachedWindowBase, uncachedWindowBase + uncachedWindowSize)); // Now accomidate the minidump containing the executable cache. if (p.MiniDumpBase != p.MiniDumpLimit) { miniDumpBase = nextVirtualAddr; nextVirtualAddr = miniDumpLimit = MemoryManager.PagePad(miniDumpBase + p.MiniDumpLimit - p.MiniDumpBase); DebugStub.WriteLine("MiniDump at 0x{0:x8} - 0x{1:x8}.", __arglist(miniDumpBase, miniDumpLimit)); } // Locate the SMAP window next smapRemapBase = nextVirtualAddr; smapRemapSize = 0; // Calculate its size SMAPINFO* smap = (SMAPINFO*)p.Smap32; for (uint i = 0; i < p.SmapCount; i++) { if ((smap[i].type != (ulong)SMAPINFO.AddressType.Free) && (smap[i].addr >= identityMapLimit)) { smapRemapSize += smap[i].size; } } if (smapRemapSize != 0) { DebugStub.WriteLine("SMAP reserve at 0x{0:x8} - 0x{1:x8}.", __arglist(smapRemapBase, smapRemapBase+smapRemapSize)); nextVirtualAddr = MemoryManager.PagePad(smapRemapBase + smapRemapSize); } // Locate the one-page scratch window immediately above the smap area scratchAddress1 = nextVirtualAddr; scratchAddress2 = scratchAddress1 + MemoryManager.PageSize; nextVirtualAddr = scratchAddress2 + MemoryManager.PageSize; DebugStub.WriteLine("scratch#1 at 0x{0:x8}, scratch#2 at 0x{1:x8}.", __arglist(scratchAddress1, scratchAddress2)); // That's all for our fixed low-memory structures kernelHeapBase = nextVirtualAddr; kernelHeapLimit = Platform.KERNEL_BOUNDARY; DebugStub.Assert(kernelHeapLimit > kernelHeapBase); DebugStub.WriteLine("KernelHeap initialized at 0x{0:x8} - 0x{1:x8}.", __arglist(kernelHeapBase, kernelHeapLimit)); // Stick the user-space page-table mapping right at the // beginning of the user's address space userPageTableBase = Platform.KERNEL_BOUNDARY; uint numPTs = (uint)(userSpaceSize / SpacePerPT); uint numPDTs = (uint)(userSpaceSize / SpacePerPDT); userPDTOffset = MemoryManager.BytesFromPages(numPTs); pdptAddr = (PDPT*)(userPageTableBase + MemoryManager.BytesFromPages(numPTs + numPDTs)); // Count the PDPT as part of the user's page table so we don't // step on the PDPT mapping later userPageTableLimit = (UIntPtr)pdptAddr + MemoryManager.PageSize; DebugStub.WriteLine("UserPageTable initialized at 0x{0:x8} - 0x{1:x8}.", __arglist(userPageTableBase, userPageTableLimit)); // Define the limits of general-purpose user space userHeapBase = userPageTableLimit; userHeapLimit = Platform.MAX_VIRTUAL_ADDR; userHeapLimit = MemoryManager.Trunc(userHeapLimit, MemoryManager.PageSize); DebugStub.WriteLine("UserHeap initialized at 0x{0:x8} - 0x{1:x8}.", __arglist(userHeapBase, userHeapLimit)); // Now we can create the kernel's address space! BootstrapSpace = CreateBootstrapSpace(); // TURN ON PAGING! // // Save the kernel's operating Pdpt value so that non-boot CPU's // may have the proper map set by the boot loader in undump.cpp // Platform nativePlatform = Platform.ThePlatform; // temporary workaround for Bartok bug AddressSpace b = BootstrapSpace; nativePlatform.KernelPdpt64 = b.PdptPage.Value; Processor.EnablePaging(BootstrapSpace); #if DEBUG // Sanity checks DebugStub.Assert(IsUserDescriptorPageMapped((UIntPtr)pdptAddr)); DebugStub.Assert(((*pdptAddr)[0])->Valid); // We expect there to be sufficient PTs to describe the entire // kernel page table for (UIntPtr step = MemoryManager.Trunc(kernelPageTableBase, SpacePerPT); step < MemoryManager.Pad(kernelPageTableLimit, SpacePerPT); step += SpacePerPT) { PT* pt = GetPT(step); DebugStub.Assert(IsKernelDescriptorPageMapped((UIntPtr)pt)); } // Obviously, the PDT for this space should be mapped too DebugStub.Assert( IsKernelDescriptorPageMapped( (UIntPtr)GetPDT( MemoryManager.Trunc(kernelPageTableBase, SpacePerPDT)))); // Same checks for user-land for (UIntPtr step = MemoryManager.Trunc(userPageTableBase, SpacePerPT); step < MemoryManager.Pad(userPageTableLimit, SpacePerPT); step += SpacePerPT) { PT* pt = GetPT(step); DebugStub.Assert(IsUserDescriptorPageMapped((UIntPtr)pt)); } DebugStub.Assert( IsUserDescriptorPageMapped( (UIntPtr)GetPDT( MemoryManager.Trunc(userPageTableBase, SpacePerPDT)))); #endif DebugStub.WriteLine("VMManager initialized with {0} physical pages still free", __arglist(PhysicalPages.GetFreePageCount())); } internal static void Finalize() { Platform nativePlatform = Platform.ThePlatform; if (nativePlatform != null) { // Restore the PDPT that the boot loader created Processor.ChangeAddressSpace(new AddressSpace(new PhysicalAddress(nativePlatform.Pdpt32))); } } internal static void PostGCInitialize() { // Create a spinlock to protect the kernel's page-mapping // structures KernelMappingLock = new SmartSpinlock(SpinLock.Types.VMKernelMapping); } // // This gets called during the boot sequence when we're ready to // start using user-space for the first time. We flesh out our // bootstrap address space with a user range and hand it out for the // first time. // internal static AddressSpace GetBootstrapSpace() { return BootstrapSpace; } // // Call to create a whole new address space from scratch // public static AddressSpace CreateNewAddressSpace() { AddressSpace retval = CreateNewKernelSpace(); InitializeUserRange(retval); return retval; } // // Writes a new virtual -> physical address into the page table, // ensuring all necessary PDTs and PTs are allocated and mapped // // The caller must specify the protection domain that owns the mapped // pages, for locking purposes. When mapping pages into the kernel // range this parameter can be null. // // What are the "locking purposes"? Routines that may *add* nodes // to the page-mapping tree need to be protected against concurrency. // public static unsafe void MapPage(PhysicalAddress physAddr, UIntPtr virtualAddr, ProtectionDomain inDomain) { DebugStub.Assert(MemoryManager.IsPageAligned(virtualAddr)); DebugStub.Assert(!IsPageMapped(virtualAddr), "Attempt to map over a mapped page"); SmartSpinlock mappingLock = null; if (virtualAddr < Platform.KERNEL_BOUNDARY) { mappingLock = KernelMappingLock; // Might be null during early boot } else { DebugStub.Assert(inDomain != null); mappingLock = inDomain.UserMappingLock; DebugStub.Assert(mappingLock != null); DebugStub.Assert(inDomain.AddressSpace == Processor.GetCurrentAddressSpace()); } bool iflag = false; if (mappingLock != null) { iflag = mappingLock.Lock(); } try { MapPageNoLock(physAddr, virtualAddr); } finally { if (mappingLock != null) { mappingLock.Unlock(iflag); } } } private static unsafe void MapPageNoLock(PhysicalAddress physAddr, UIntPtr virtualAddr) { DebugStub.Assert(MemoryManager.IsPageAligned(virtualAddr)); DebugStub.Assert(!IsPageMapped(virtualAddr), "Attempt to map over a mapped page"); EnsurePTE(virtualAddr); MapPageInternal(physAddr, virtualAddr); } // // The caller must specify the protection domain that owns the mapped // pages, for locking purposes. When mapping pages into the kernel // range this parameter can be null. // // What are the "locking purposes"? Routines that may *add* nodes // to the page-mapping tree need to be protected against concurrency. // public static unsafe void MapRange(PhysicalAddress physAddr, UIntPtr virtualAddr, UIntPtr numPages, ProtectionDomain inDomain) { DebugStub.Assert(MemoryManager.IsPageAligned(virtualAddr)); ulong stepPhys = physAddr.Value; UIntPtr stepVirt = virtualAddr; UIntPtr limitAddr = virtualAddr + MemoryManager.BytesFromPages(numPages); SmartSpinlock mappingLock = null; if (limitAddr <= Platform.KERNEL_BOUNDARY) { mappingLock = KernelMappingLock; // Might be null during early boot } else { DebugStub.Assert(inDomain != null); mappingLock = inDomain.UserMappingLock; DebugStub.Assert(mappingLock != null); DebugStub.Assert(inDomain.AddressSpace == Processor.GetCurrentAddressSpace()); } bool iflag = false; if (mappingLock != null) { iflag = mappingLock.Lock(); } try { for (UIntPtr i = 0; i < numPages; i++) { MapPageNoLock(new PhysicalAddress(stepPhys), stepVirt); stepPhys += MemoryManager.PageSize; stepVirt += MemoryManager.PageSize; } } finally { if (mappingLock != null) { mappingLock.Unlock(iflag); } } } // // Unmaps an existing virtual-memory page. // Does NOT release the underlying physical page // (whose address is returned) // public static unsafe PhysicalAddress UnmapPage(UIntPtr virtualAddr) { DebugStub.Assert(MemoryManager.IsPageAligned(virtualAddr)); PTE* pte = GetPTE(virtualAddr); if (pte->Valid) { PhysicalAddress phys = new PhysicalAddress(pte->Address); pte->Valid = false; pte->Address = 0; Processor.InvalidateTLBEntry(virtualAddr); return phys; } else { DebugStub.Assert(false, "Attempt to unmap an unmapped page"); return PhysicalAddress.Null; } } public static unsafe void UnmapRange(UIntPtr virtualAddr, UIntPtr numPages) { DebugStub.Assert(MemoryManager.IsPageAligned(virtualAddr)); UIntPtr stepVirt = virtualAddr; for (UIntPtr i = 0; i < numPages; i++) { UnmapPage(stepVirt); stepVirt += MemoryManager.PageSize; } } // // Returns the physical address that a given virtual address is // currently mapped to (or PhysicalAddress.Null if none) // public static unsafe PhysicalAddress GetPageMapping(UIntPtr virtualAddr) { DebugStub.Assert(MemoryManager.IsPageAligned(virtualAddr)); PTE* pte = GetPTE(virtualAddr); if (pte->Valid) { return new PhysicalAddress(pte->Address); } else { return PhysicalAddress.Null; } } // Examines the page-table structures to determine whether the page // at pageAddr is mapped or not. // // REVIEW: use a processor instruction instead to get a more definitive // answer? [AccessedByRuntime("referenced from halkd.cpp")] public static unsafe bool IsPageMapped(UIntPtr pageAddr) { DebugStub.Assert(MemoryManager.IsPageAligned(pageAddr)); if ((pageAddr >= kernelPageTableBase) && (pageAddr < kernelPageTableLimit)) { // This is the address of a kernel page-table page. // There should always be sufficient PTs to describe these. return IsKernelDescriptorPageMapped(pageAddr); } else if ((pageAddr >= userPageTableBase) && (pageAddr < userPageTableLimit)) { // User-land page-table page return IsUserDescriptorPageMapped(pageAddr); } else { // This page isn't part of the page table, so its // PT may be missing. Check that. PT* pt = GetPT(MemoryManager.Trunc(pageAddr, SpacePerPT)); if (pageAddr >= Platform.KERNEL_BOUNDARY) { if (!IsUserDescriptorPageMapped((UIntPtr)pt)) { // The PT for this address isn't even mapped, // so the address itself can't be mapped! return false; } } else { if (!IsKernelDescriptorPageMapped((UIntPtr)pt)) { return false; } } // There's a PT but it might say that this page isn't // mapped. PTE* pte = GetPTE(pageAddr); return pte->Valid; } } // // A complement to TranslatePhysicalRange. Returns true if // the virtual address provided falls into one of the // virtual ranges we have reserved for permanent mappings // of hardware. // internal static unsafe bool IsPermaMapped(UIntPtr virtualAddr, UIntPtr limitAddr) { if (limitAddr <= identityMapLimit) { return true; } if ((virtualAddr >= smapRemapBase) && (limitAddr <= smapRemapBase + smapRemapSize)) { return true; } if ((virtualAddr >= uncachedWindowBase) && (limitAddr <= uncachedWindowBase + uncachedWindowSize)) { return true; } return false; } // // This call returns non-null if the physical address provided // is already well-known and mapped into a stable area of // memory. // internal static unsafe UIntPtr TranslatePhysicalRange(PhysicalAddress physAddress, UIntPtr numBytes) { Platform p = Platform.ThePlatform; ulong physAddr = physAddress.Value; ulong physLimit = physAddr + (ulong)numBytes; // Some physical pointers are identity-mapped if (physLimit <= identityMapLimit) { return physAddr; } // Is this an address into the executable-image range // created by the loader? if ((physAddr >= p.MiniDumpBase) && (physLimit <= p.MiniDumpLimit)) { // Assume we move the exec image lower in memory DebugStub.Assert(p.MiniDumpBase > miniDumpBase); return (UIntPtr)(physAddr - (Platform.ThePlatform.MiniDumpBase - miniDumpBase)); } // Is this a pointer into the uncached window at the // top of physical memory? ulong uncachedLimit = (ulong)Platform.UNCACHED_PHYSICAL + (ulong)Platform.UNCACHED_MAPPED; if ((physAddr >= Platform.UNCACHED_PHYSICAL) && (physLimit <= uncachedLimit)) { return uncachedWindowBase + (UIntPtr)(physAddr - Platform.UNCACHED_PHYSICAL); } // Is this an address into an SMAP-reserved area? SMAPINFO *smap = (SMAPINFO*)p.Smap32; UIntPtr stepVirtual = smapRemapBase; for (uint i = 0; i < p.SmapCount; i++) { if ((smap[i].type != (ulong)SMAPINFO.AddressType.Free) && (smap[i].addr >= identityMapLimit)) { if ((physAddr >= smap[i].addr) && (physLimit <= smap[i].addr + smap[i].size)) { // Match! return stepVirtual + (physAddr - smap[i].addr); } stepVirtual += smap[i].size; } } // No match return UIntPtr.Zero; } // // Finds a range of virtual space in the provided address range // and maps in the requested physical range on the fly // internal static UIntPtr MapPhysicalMemory(VirtualMemoryRange range, Process process, PhysicalAddress physStart, UIntPtr numBytes) { ulong startPhysPage = MemoryManager.Trunc(physStart.Value, MemoryManager.PageSize); ulong physLimit = MemoryManager.PagePad(physStart.Value + (ulong)numBytes); UIntPtr numPages = MemoryManager.PagesFromBytes((UIntPtr)(physLimit - startPhysPage)); // Reserve an area to accomplish the mapping in UIntPtr mapPoint = range.Reserve(numPages, process, 0, PageType.System); DebugStub.Assert(mapPoint != UIntPtr.Zero, "Out of virtual memory trying to map a physical range"); // Perform the mapping MapRange(new PhysicalAddress(startPhysPage), mapPoint, numPages, range.ParentDomain); // Take care to point into the mapped virtual memory properly return mapPoint + (UIntPtr)MemoryManager.BytesNotAligned( physStart.Value, MemoryManager.PageSize); } // // Reverses the operation of MapPhysicalMemory // internal static void UnmapPhysicalMemory(VirtualMemoryRange range, Process process, UIntPtr startAddr, UIntPtr limitAddr) { UIntPtr startPageAddr = MemoryManager.Trunc(startAddr, MemoryManager.PageSize); UIntPtr limitPageAddr = MemoryManager.PagePad(limitAddr); uint numPages = (uint)MemoryManager.PagesFromBytes(limitPageAddr - startPageAddr); // First undo the mapping UnmapRange(startPageAddr, numPages); // Now mark all the pages as free range.Unreserve(startPageAddr, numPages, process); } ///////////////////////////////////// // PRIVATE BOOTSTRAP METHODS ///////////////////////////////////// // This function sets up the initial memory space from scratch. // It assumes paging is NOT on. private static unsafe AddressSpace CreateBootstrapSpace() { // // The kernel's address space serves as the template for all other // address spaces, because everything it maps must also be mapped // in other spaces. // // We need mapping for a bunch of things before we can turn on // paging. Here's the list: // // BELOW KERNEL_BOUNDARY: // // Zero through BootInfo.PHYSICAL_DISABLED is always kept unmapped // // Memory from PhysicalBase to the top of the I/O heap is // identity-mapped (we could probably stand to punch some holes // in this). // // The kernel image and the boot record need to be identity mapped. // // Next comes the page-table area, starting at kernelPageTableBase. // It is large enough to cover all the pages that could be // needed to describe the kernel's address space. // // Next comes a window that maps the executable images downloaded // by the boot loader (these usually get stuck in high memory) // into the kernel's space, which is a more convenient location. // // Mappings above KERNEL_BOUNDARY are free for use by processes, // except that from UNCACHED_PHYSICAL to the top (4GB) is marked // uncached. // Platform p = Platform.ThePlatform; // To get started, we need a PDPT. PhysicalAddress phys = PhysicalPages.AllocPage(); DebugStub.Assert(phys != PhysicalAddress.Null); PDPT* pdpt = (PDPT*)phys.Value; pdpt->Init(); DebugStub.WriteLine("Allocated PDPT at 0x{0:x8}.", __arglist(phys.Value)); // Start out mapping the entire identityMapped region as writable, not writethrough, // with caching enabled // TODO: Can't use p.PhysicalBase because hypervisor is currently using pages below // that threshold. s/b fix hypervisor PhysicalBase reporting. BootstrapMapRange(pdpt, 0, identityMapLimit, new PhysicalAddress(0), true, // writeable false, // not write-through false); // don't disable caching // Map the uncached area of high memory down into the kernel space BootstrapMapRange(pdpt, uncachedWindowBase, MemoryManager.PagePad(uncachedWindowBase + uncachedWindowSize), new PhysicalAddress(Platform.UNCACHED_PHYSICAL), true, // writeable true, // writethrough true); // disable caching BootstrapMapRange(pdpt, miniDumpBase, MemoryManager.PagePad(miniDumpLimit), new PhysicalAddress(p.MiniDumpBase), true, // writeable false, // not writethrough false); // don't disable caching // Walk the SMAP and remap any reserved areas into the smap window SMAPINFO* smap = (SMAPINFO*)p.Smap32; UIntPtr stepVirtual = smapRemapBase; for (uint i = 0; i < p.SmapCount; i++) { if ((smap[i].type != (ulong)SMAPINFO.AddressType.Free) && (smap[i].addr >= identityMapLimit)) { DebugStub.Assert(MemoryManager.IsPageAligned(smap[i].size)); BootstrapMapRange(pdpt, stepVirtual, stepVirtual + smap[i].size, new PhysicalAddress(smap[i].addr), true, // writeable true, // write-through, true); // disable caching stepVirtual += smap[i].size; } } DebugStub.Assert((stepVirtual - smapRemapBase) == smapRemapSize); // Make sure we have committed a PDT for each kernel PDPE slot, so // other address spaces can alias the kernel range sensibly BootstrapEnsurePDTs(pdpt, UIntPtr.Zero, Platform.KERNEL_BOUNDARY); kernelPDPE1 = new PhysicalAddress(((*pdpt)[0])->Address); DebugStub.Assert(kernelPDPE1 != PhysicalAddress.Null); if (numKernelPDPEs >= 2) { kernelPDPE2 = new PhysicalAddress(((*pdpt)[1])->Address); DebugStub.Assert(kernelPDPE2 != PhysicalAddress.Null); } if (numKernelPDPEs >= 3) { kernelPDPE3 = new PhysicalAddress(((*pdpt)[2])->Address); DebugStub.Assert(kernelPDPE3 != PhysicalAddress.Null); } // Create the mappings for all the page-table structures // we just set up so we can still manipulate them after // we turn on paging! BootstrapMapPageTable(pdpt); return new AddressSpace(new PhysicalAddress((ulong)pdpt)); } // For use during startup. Creates a mapping for all the physical // pages currently being used to describe memory mappings (yes, // this is confusing). // // Assumes paging is NOT on so that physical addresses can be // used as-is private static unsafe void BootstrapMapPageTable(PDPT* pdpt) { BootstrapEnsurePTs(pdpt, kernelPageTableBase, kernelPageTableLimit); BootstrapEnsurePTs(pdpt, userPageTableBase, userPageTableLimit); BootstrapMapPage(pdpt, (UIntPtr)pdptAddr, new PhysicalAddress((ulong)pdpt), true); // First map the kernel page table pages for (uint i = 0; i < numKernelPDPEs; i++) { if (((*pdpt)[i])->Valid) { PDT* pdt = (PDT*)(((*pdpt)[i])->Address); BootstrapMapPage(pdpt, kernelPageTableBase + kernelPDTOffset + MemoryManager.BytesFromPages(i), new PhysicalAddress((ulong)pdt), true); for (uint j = 0; j < 512; j++) { if (((*pdt)[j])->Valid) { PT* pt = (PT*)(((*pdt)[j])->Address); BootstrapMapPage(pdpt, kernelPageTableBase + MemoryManager.BytesFromPages((i * 512) + j), new PhysicalAddress((ulong)pt), true); } } } } // Now map user-space page-table pages for (uint i = 0; i < (4 - numKernelPDPEs); i++) { if (((*pdpt)[numKernelPDPEs + i])->Valid) { PDT* pdt = (PDT*)(((*pdpt)[numKernelPDPEs + i])->Address); BootstrapMapPage(pdpt, userPageTableBase + userPDTOffset + MemoryManager.BytesFromPages(i), new PhysicalAddress((ulong)pdt), true); for (uint j = 0; j < 512; j++) { if (((*pdt)[j])->Valid) { PT* pt = (PT*)(((*pdt)[j])->Address); BootstrapMapPage(pdpt, userPageTableBase + MemoryManager.BytesFromPages((i * 512) + j), new PhysicalAddress((ulong)pt), true); } } } } } // For use during startup. Makes sure the given PDPT points to // enough PDTs (which are allocated as necessary) to describe the // address range specified. // // Assumes paging is NOT on. private static unsafe void BootstrapEnsurePDTs(PDPT* pdpt, UIntPtr startAddr, UIntPtr limitAddr) { limitAddr = MemoryManager.Pad(limitAddr, SpacePerPDT); uint stepPDTidx = (uint)(startAddr / SpacePerPDT); UIntPtr stepAddr = startAddr; while (stepAddr < limitAddr) { if (!((*pdpt)[stepPDTidx])->Valid) { PhysicalAddress newPDT = PhysicalPages.AllocPage(); ((PDT*)newPDT.Value)->Init(); ((*pdpt)[stepPDTidx])->Address = newPDT.Value; ((*pdpt)[stepPDTidx])->Valid = true; } stepPDTidx++; stepAddr += SpacePerPDT; } } // For use during startup. Makes sure there are enough PTs // (which are allocated as necessary) to describe the address // range specified. // // Assumes paging is NOT on. private static unsafe void BootstrapEnsurePTs(PDPT* pdpt, UIntPtr startAddr, UIntPtr limitAddr) { limitAddr = MemoryManager.Pad(limitAddr, SpacePerPT); BootstrapEnsurePDTs(pdpt, startAddr, limitAddr); uint stepPDTidx = (uint)(startAddr / SpacePerPDT); UIntPtr stepAddr = startAddr; while (stepAddr < limitAddr) { DebugStub.Assert(((*pdpt)[stepPDTidx])->Valid); // ensured above PDT* pdt = (PDT*)(((*pdpt)[stepPDTidx])->Address); UIntPtr offsetInPDT = MemoryManager.BytesNotAligned(stepAddr, SpacePerPDT); uint stepPTidx = (uint)(offsetInPDT / SpacePerPT); while ((stepAddr < limitAddr) && (stepPTidx < 512)) { if (!((*pdt)[stepPTidx])->Valid) { PhysicalAddress newPT = PhysicalPages.AllocPage(); ((PT*)newPT.Value)->Init(); ((*pdt)[stepPTidx])->Address = newPT.Value; ((*pdt)[stepPTidx])->User = true; // Ring3 ((*pdt)[stepPTidx])->Valid = true; // Up to each page to set its writability ((*pdt)[stepPTidx])->Writeable = true; } stepPTidx++; stepAddr += SpacePerPT; } stepPDTidx++; } } // For use during startup. Map a range of addresses. // // Assumes paging is NOT on. private static unsafe void BootstrapMapRange(PDPT* pdpt, UIntPtr startAddr, UIntPtr limitAddr, PhysicalAddress physicalStart, bool writeable, bool writeThrough, bool cacheDisable) { DebugStub.Assert(MemoryManager.IsPageAligned(startAddr)); DebugStub.Assert(MemoryManager.IsPageAligned(limitAddr)); DebugStub.Assert(MemoryManager.IsPageAligned(physicalStart.Value)); BootstrapEnsurePTs(pdpt, startAddr, limitAddr); uint stepPDTidx = (uint)(startAddr / SpacePerPDT); UIntPtr stepAddr = startAddr; ulong stepPhysical = physicalStart.Value; while ((stepAddr < limitAddr) && (stepPDTidx < 4)) { DebugStub.Assert(((*pdpt)[stepPDTidx])->Valid); // ensured above PDT* pdt = (PDT*)((*pdpt)[stepPDTidx])->Address; UIntPtr offsetInPDT = MemoryManager.BytesNotAligned(stepAddr, SpacePerPDT); DebugStub.Assert(MemoryManager.IsPageAligned(offsetInPDT)); uint stepPTidx = (uint)(offsetInPDT / SpacePerPT); while ((stepAddr < limitAddr) && (stepPTidx < 512)) { DebugStub.Assert(((*pdt)[stepPTidx])->Valid); // ensured above PT* pt = (PT*)(((*pdt)[stepPTidx])->Address); UIntPtr offsetInPT = MemoryManager.BytesNotAligned(stepAddr, SpacePerPT); DebugStub.Assert(MemoryManager.IsPageAligned(offsetInPT)); uint pageIndex = (uint)MemoryManager.PagesFromBytes(offsetInPT); while ((stepAddr < limitAddr) && (pageIndex < 512)) { // allow overlapping writes for convience... //DebugStub.Assert(!((*pt)[pageIndex])->Valid); ((*pt)[pageIndex])->Address = stepPhysical; ((*pt)[pageIndex])->User = true; // Ring3 ((*pt)[pageIndex])->Valid = true; ((*pt)[pageIndex])->Writeable = writeable; ((*pt)[pageIndex])->Global = stepAddr < Platform.KERNEL_BOUNDARY; ((*pt)[pageIndex])->WriteThrough = writeThrough; ((*pt)[pageIndex])->CacheDisable = cacheDisable; stepPhysical += MemoryManager.PageSize; pageIndex++; stepAddr += MemoryManager.PageSize; } stepPTidx++; } stepPDTidx++; } DebugStub.Assert(stepAddr >= limitAddr); } // For use during startup. Maps a single page of memory. // // Assumes paging is NOT on. private static unsafe void BootstrapMapPage(PDPT* pdpt, UIntPtr pageAddr, PhysicalAddress physPage, bool writeable) { DebugStub.Assert(MemoryManager.IsPageAligned(pageAddr)); DebugStub.Assert(MemoryManager.IsPageAligned(physPage.Value)); BootstrapEnsurePTs(pdpt, pageAddr, pageAddr + MemoryManager.PageSize); uint pdtIndex = (uint)(pageAddr / SpacePerPDT); UIntPtr offsetInPDT = MemoryManager.BytesNotAligned(pageAddr, SpacePerPDT); DebugStub.Assert(((*pdpt)[pdtIndex])->Valid); // ensured above PDT* pdt = (PDT*)((*pdpt)[pdtIndex])->Address; uint ptIndex = (uint)(offsetInPDT / SpacePerPT); UIntPtr offsetInPT = MemoryManager.BytesNotAligned(offsetInPDT, SpacePerPT); DebugStub.Assert(((*pdt)[ptIndex])->Valid); // ensured above; DebugStub.Assert(((*pdt)[ptIndex])->User); // Ring3 PT* pt = (PT*)(((*pdt)[ptIndex])->Address); uint pageIndex = (uint)MemoryManager.PagesFromBytes(offsetInPT); DebugStub.Assert(!((*pt)[pageIndex])->Valid); ((*pt)[pageIndex])->Address = physPage.Value; ((*pt)[pageIndex])->User = true; // Ring3 ((*pt)[pageIndex])->Writeable = writeable; ((*pt)[pageIndex])->Global = pageAddr < Platform.KERNEL_BOUNDARY; ((*pt)[pageIndex])->Valid = true; } ///////////////////////////////////// // PRIVATE METHODS, POST-BOOTSTRAP ///////////////////////////////////// // // Create a new address space that just maps the (existing) kernel // range. // private static unsafe AddressSpace CreateNewKernelSpace() { // REVIEW: it's regrettable that we have to lock // for access to the scratch area. Maybe we should try to // reserve two pages of kernel-space memory instead? bool iflag = ScratchLock(); try { // Get a page to use as a PDPT and map it PhysicalAddress physPage = PhysicalPages.AllocPage(); DebugStub.Assert(physPage != PhysicalAddress.Null); MapPageInternal(physPage, scratchAddress1); PDPT* pdpt = (PDPT*)scratchAddress1; pdpt->Init(); ((*pdpt)[0])->Address = kernelPDPE1.Value; ((*pdpt)[0])->Valid = true; // This is klunky but we can't have static arrays... if (numKernelPDPEs >= 2) { ((*pdpt)[1])->Address = kernelPDPE2.Value; ((*pdpt)[1])->Valid = true; } if (numKernelPDPEs >= 3) { ((*pdpt)[2])->Address = kernelPDPE3.Value; ((*pdpt)[2])->Valid = true; } DebugStub.Assert(numKernelPDPEs < 4); PhysicalAddress sanity = UnmapPage(scratchAddress1); DebugStub.Assert(sanity == physPage); return new AddressSpace(physPage); } finally { ScratchUnlock(iflag); } } // // Set up a kernel-only address space so its user range can be used. // private static unsafe void InitializeUserRange(AddressSpace space) { // Our task here is to set up enough PTs to cover the // user range's page-table pages. This gets confusing because // we also want to write mappings for the PDT and PTs we // create. // Assume the user-space page-table range doesn't straddle a PDT. DebugStub.Assert((userPageTableBase / SpacePerPDT) == ((userPageTableLimit - 1) / SpacePerPDT)); // REVIEW: it's regrettable that we have to lock // for access to the scratch area. Maybe we should try to // reserve two pages of kernel-space memory instead? bool iflag = ScratchLock(); try { PT* pt = null; // First create the PDT PhysicalAddress newPdt = PhysicalPages.AllocPage(); DebugStub.Assert(newPdt != PhysicalAddress.Null); MapPageInternal(newPdt, scratchAddress1); PDT* pdt = (PDT*)scratchAddress1; pdt->Init(); // We're going to need to make a mapping for the PDT // so it is addressable at userPageTableBase + userPDTOffset. // Figure out what PT and PTE that lands in. UIntPtr bytesIntoPDTRange = MemoryManager.BytesNotAligned( userPageTableBase + userPDTOffset, SpacePerPDT); uint pdtPTIdx = (uint)(bytesIntoPDTRange / SpacePerPT); UIntPtr bytesIntoPTRange = MemoryManager.BytesNotAligned( bytesIntoPDTRange, SpacePerPT); uint pdtPTEIdx = (uint)(bytesIntoPTRange / MemoryManager.PageSize); // Same story for the PDPT, which gets mapped to a user-space // address (pdptAddr). bytesIntoPDTRange = MemoryManager.BytesNotAligned((UIntPtr)pdptAddr, SpacePerPDT); uint pdptPTIdx = (uint)(bytesIntoPDTRange / SpacePerPT); bytesIntoPTRange = MemoryManager.BytesNotAligned( bytesIntoPDTRange, SpacePerPT); uint pdptPTEIdx = (uint)(bytesIntoPTRange / MemoryManager.PageSize); // Now fill in as many PDT slots with new PTs as are // necessary to cover the pagetable range bytesIntoPDTRange = MemoryManager.BytesNotAligned(userPageTableBase, SpacePerPDT); uint startPTIdx = (uint)(bytesIntoPDTRange / SpacePerPT); bytesIntoPDTRange = MemoryManager.BytesNotAligned(userPageTableLimit, SpacePerPDT); bytesIntoPDTRange = MemoryManager.Pad(bytesIntoPDTRange, SpacePerPT); uint limitPTIdx = (uint)(bytesIntoPDTRange / SpacePerPT); uint numPTs = limitPTIdx - startPTIdx; for (uint i = startPTIdx; i < limitPTIdx; i++) { PhysicalAddress newPT = PhysicalPages.AllocPage(); ((*pdt)[i])->Address = newPT.Value; ((*pdt)[i])->User = true; // Ring3 ((*pdt)[i])->Valid = true; ((*pdt)[i])->Writeable = true; // each page sets this individually // Map in the new PT page MapPageInternal(newPT, scratchAddress2, i > startPTIdx); pt = (PT*)scratchAddress2; pt->Init(); if (i == pdtPTIdx) { // This is the PT that the PTE for our new PDT // goes into (yes, that's confusing) ((*pt)[pdtPTEIdx])->Address = newPdt.Value; ((*pt)[pdtPTEIdx])->User = true; // Ring3 ((*pt)[pdtPTEIdx])->Valid = true; ((*pt)[pdtPTEIdx])->Writeable = true; } if (i == pdptPTIdx) { // Same drill for the PDPT ((*pt)[pdptPTEIdx])->Address = space.PdptPage.Value; ((*pt)[pdptPTEIdx])->User = true; // Ring3 ((*pt)[pdptPTEIdx])->Valid = true; ((*pt)[pdptPTEIdx])->Writeable = true; } } // Now, write mappings for all the new PT pages we just // created. Assume here that all the PT page mappings // fit into a single PT. DebugStub.Assert(startPTIdx + numPTs < 512); // Map the PT that will hold our PTEs MapPageInternal(new PhysicalAddress(((*pdt)[startPTIdx])->Address), scratchAddress2, true); pt = (PT*)scratchAddress2; bytesIntoPTRange = MemoryManager.BytesNotAligned(userPageTableBase, SpacePerPT); uint startPTEIdx = (uint)(bytesIntoPTRange / MemoryManager.PageSize); for (uint i = 0; i < numPTs; i++) { ((*pt)[startPTEIdx + i])->Address = ((*pdt)[startPTIdx + i])->Address; DebugStub.Assert(((*pt)[startPTEIdx + i])->Address != 0); ((*pt)[startPTEIdx + i])->User = true; // Ring3 ((*pt)[startPTEIdx + i])->Valid = true; ((*pt)[startPTEIdx + i])->Writeable = true; } PhysicalAddress sanity = UnmapPage(scratchAddress2); DebugStub.Assert(sanity.Value == ((*pdt)[startPTIdx])->Address); // Now flip scratchAddress1 to the PDPT for this space MapPageInternal(space.PdptPage, scratchAddress1, true); PDPT* pdpt = (PDPT*)scratchAddress1; // Write our new PDT into the PDPT uint pdtIdx = (uint)(userPageTableBase / SpacePerPDT); DebugStub.Assert(pdtIdx >= numKernelPDPEs); ((*pdpt)[pdtIdx])->Address = newPdt.Value; ((*pdpt)[pdtIdx])->Valid = true; sanity = UnmapPage(scratchAddress1); DebugStub.Assert(sanity == space.PdptPage); } finally { ScratchUnlock(iflag); } } // Return a pointer to the PDT that describes the address // range beginning at forAddr, allocating and mapping a // new one if necessary. private static unsafe PDT* EnsurePDT(UIntPtr forAddr) { PDPE* pdpe = GetPDPE(forAddr); // must succeed PDT* retval = GetPDT(forAddr); if (!IsPageMapped((UIntPtr)retval)) { // Don't panic; grab a new page to use as a PDT PhysicalAddress newPdt = PhysicalPages.AllocPage(); if (newPdt == PhysicalAddress.Null) { Kernel.Panic("Out of physical pages while trying to allocate a PDT"); return null; } // Activate this page. If this faults, it means // our address space wasn't set up properly with // enough PTs to describe the page table pages. MapPageInternal(newPdt, (UIntPtr)retval); // Now we can address the new page retval->Init(); // Point our parent PDPT here DebugStub.Assert(!pdpe->Valid); pdpe->Address = newPdt.Value; pdpe->Valid = true; } else { DebugStub.Assert(pdpe->Valid); } return retval; } // Return a pointer to the PDE for the PT describing the // memory range beginning at forAddr, allocating and // mapping descriptors as necessary. private static unsafe PDE* EnsurePDE(UIntPtr forAddr) { DebugStub.Assert(MemoryManager.BytesNotAligned(forAddr, SpacePerPT) == 0); UIntPtr bytesIntoPDTRange = MemoryManager.BytesNotAligned(forAddr, SpacePerPDT); EnsurePDT(forAddr - bytesIntoPDTRange); return GetPDE(forAddr); } // Return a pointer to the PT that describes the address // range beginning at forAddr, allocating and mapping a // new one if necessary. private static unsafe PT* EnsurePT(UIntPtr forAddr) { PDE* pde = EnsurePDE(forAddr); PT* retval = GetPT(forAddr); if (!IsPageMapped((UIntPtr)retval)) { // Don't panic; grab a new page to use as a PT PhysicalAddress newPt = PhysicalPages.AllocPage(); if (newPt == PhysicalAddress.Null) { Kernel.Panic("Out of physical pages while trying to allocate a PT"); return null; } // Activate this page. If this faults, it means // our address space wasn't set up properly with // enough PTs to describe the page table pages. MapPageInternal(newPt, (UIntPtr)retval); // Now we can address the new page retval->Init(); // Add ourselves to our parent PDT DebugStub.Assert(!pde->Valid); pde->Address = newPt.Value; pde->User = true; // Ring3 pde->Valid = true; pde->Writeable = true; } else { DebugStub.Assert(pde->Valid); } return retval; } // Return a pointer to the PTE for the given page address, // allocating and mapping descriptors as necessary. private static unsafe PTE* EnsurePTE(UIntPtr pageAddr) { DebugStub.Assert(MemoryManager.IsPageAligned(pageAddr)); UIntPtr bytesIntoPTRange = MemoryManager.BytesNotAligned(pageAddr, SpacePerPT); EnsurePT(pageAddr - bytesIntoPTRange); return GetPTE(pageAddr); } // Return a pointer to the PDT that describes the address // range beginning at forAddr, without ensuring that it is // mapped. private static unsafe PDT* GetPDT(UIntPtr forAddr) { // Make sure forAddr specifies a range that // makes sense DebugStub.Assert(MemoryManager.BytesNotAligned(forAddr, SpacePerPDT) == 0); UIntPtr pdtBase = kernelPageTableBase + kernelPDTOffset; UIntPtr addr = forAddr; if (forAddr >= Platform.KERNEL_BOUNDARY) { // User-space address pdtBase = userPageTableBase + userPDTOffset; addr -= Platform.KERNEL_BOUNDARY; } // Index into the table of PDTs uint pdtIndex = (uint)(addr / SpacePerPDT); return (PDT*)(pdtBase + MemoryManager.BytesFromPages(pdtIndex)); } // Return a pointer to the PT that describes the address // range beginning at forAddr, without ensuring that it is // mapped. private static unsafe PT* GetPT(UIntPtr forAddr) { DebugStub.Assert(MemoryManager.BytesNotAligned(forAddr, SpacePerPT) == 0); UIntPtr ptBase = kernelPageTableBase; UIntPtr addr = forAddr; if (forAddr >= Platform.KERNEL_BOUNDARY) { // User-space address ptBase = userPageTableBase; addr -= Platform.KERNEL_BOUNDARY; } // Index into the table of PTs uint ptIndex = (uint)(addr / SpacePerPT); return (PT*)(ptBase + MemoryManager.BytesFromPages(ptIndex)); } // Return a pointer to the PDPE for the PDT that describes // the memory range beginning at the provided address. // We assume the PDPT is always mapped. private static unsafe PDPE* GetPDPE(UIntPtr forAddr) { DebugStub.Assert(MemoryManager.BytesNotAligned(forAddr, SpacePerPDT) == 0); uint PDTidx = (uint)(forAddr / SpacePerPDT); return (*pdptAddr)[(uint)PDTidx]; } // Return a pointer to the PDE for the PT that describes // the memory range beginning at the provided address, // without ensuring that the PDT is mapped private static unsafe PDE* GetPDE(UIntPtr forAddr) { DebugStub.Assert(MemoryManager.BytesNotAligned(forAddr, SpacePerPT) == 0); UIntPtr bytesIntoPDTRange = MemoryManager.BytesNotAligned(forAddr, SpacePerPDT); uint ptWithinPDT = (uint)(bytesIntoPDTRange / SpacePerPT); PDT* pdt = GetPDT(forAddr - bytesIntoPDTRange); return (*pdt)[(uint)ptWithinPDT]; } // Return a pointer to the PTE for the given page address, // without ensuring that the PT is mapped. private static unsafe PTE* GetPTE(UIntPtr pageAddr) { DebugStub.Assert(MemoryManager.IsPageAligned(pageAddr)); UIntPtr bytesIntoPTRange = MemoryManager.BytesNotAligned(pageAddr, SpacePerPT); uint pageWithinPT = (uint)MemoryManager.PagesFromBytes(bytesIntoPTRange); PT* pt = GetPT(pageAddr - bytesIntoPTRange); return (*pt)[pageWithinPT]; } // This function directly examines the PTE for the virtual // address of a page-table structure (this address should // fall between pageTableBaseAddr and // pageTableBaseAddr + PageTableSize), on the assumption // that sufficient PTs to describe all page-table structures // are always mapped. private static unsafe bool IsKernelDescriptorPageMapped(UIntPtr pageAddr) { // Make sure this address is for a page-table element DebugStub.Assert(MemoryManager.IsPageAligned(pageAddr)); DebugStub.Assert(pageAddr >= kernelPageTableBase); DebugStub.Assert(pageAddr < kernelPageTableLimit); PTE* pte = GetPTE(pageAddr); // If this faults, our address space wasn't set up right return pte->Valid; } // Same for userland private static unsafe bool IsUserDescriptorPageMapped(UIntPtr pageAddr) { // Make sure this address is for a page-table element DebugStub.Assert(MemoryManager.IsPageAligned(pageAddr)); DebugStub.Assert(pageAddr >= userPageTableBase); DebugStub.Assert(pageAddr < userPageTableLimit); PTE* pte = GetPTE(pageAddr); // If this faults, our address space wasn't set up right return pte->Valid; } // // Writes a new virtual -> physical address into the page table, // ASSUMING the right PT is already allocated and mapped. // private static unsafe void MapPageInternal(PhysicalAddress physAddr, UIntPtr virtualAddr, bool expectExistingMapping) { DebugStub.Assert(MemoryManager.IsPageAligned(virtualAddr)); PTE* pte = GetPTE(virtualAddr); DebugStub.Assert(pte->Valid == expectExistingMapping, "Unexpected mapping state"); pte->Address = physAddr.Value; pte->User = true; // Ring3 pte->Valid = true; pte->Writeable = true; pte->Global = virtualAddr < Platform.KERNEL_BOUNDARY; Processor.InvalidateTLBEntry(virtualAddr); } private static unsafe void MapPageInternal(PhysicalAddress physAddr, UIntPtr virtualAddr) { // Expect to not overwrite an existing mapping MapPageInternal(physAddr, virtualAddr, false); } [Inline] [NoStackLinkCheckTrans] private static bool ScratchLock() { bool enabled = Processor.DisableInterrupts(); #if SINGULARITY_MP scratchSpin.Acquire(Thread.CurrentThread); #endif // SINGULARITY_MP return enabled; } [Inline] [NoStackLinkCheckTrans] private static void ScratchUnlock(bool iflag) { #if SINGULARITY_MP scratchSpin.Release(Thread.CurrentThread); #endif // SINGULARITY_MP Processor.RestoreInterrupts(iflag); } } }