[sir_richard] 45280: [NTOS]: Rework the trap macros again. This time we needed some added complexity because of the need for runtime patching at boot. We got away with it nicely and basically implement a simple "if" in ASM, which gives us a deterministic set of instructions (vs the compiler). We then patch if if needed (see next bullet). [NTOS]: Support systems without SYSENTER. Nothing was actually disabling the SYSEXIT mechanism recently enabled. Now IRET is always used unless a SYSENTER machi

Author: sir_richard
Date: Wed Jan 27 04:05:10 2010
New Revision: 45280

URL: http://svn.reactos.org/svn/reactos?rev=45280&view=rev
Log:
[NTOS]: Rework the trap macros again. This time we needed some added complexity because of the need for runtime patching at boot. We got away with it nicely and basically implement a simple "if" in ASM, which gives us a deterministic set of instructions (vs the compiler). We then patch if if needed (see next bullet).
[NTOS]: Support systems without SYSENTER. Nothing was actually disabling the SYSEXIT mechanism recently enabled. Now IRET is always used unless a SYSENTER machine is discovered, in which case the jmp to the IRET code is patched with a jmp to SYSEXIT code.
[PERF]: Set WP0 bit in CR0 later, thus allowing the existing runtime patches (and this new one) to function without requiring MmSetPageProtect. Saves TLB flushes and page table walking/mapping on boot.
Left in some debug prints to see what buildbot reports...will remove if it survives.

Modified:
    trunk/reactos/ntoskrnl/include/internal/i386/ke.h
    trunk/reactos/ntoskrnl/include/internal/trap_x.h
    trunk/reactos/ntoskrnl/ke/i386/cpu.c
    trunk/reactos/ntoskrnl/ke/i386/kiinit.c

Modified: trunk/reactos/ntoskrnl/include/internal/i386/ke.h
URL: http://svn.reactos.org/svn/reactos/trunk/reactos/ntoskrnl/include/internal/i...
==============================================================================
--- trunk/reactos/ntoskrnl/include/internal/i386/ke.h [iso-8859-1] (original)
+++ trunk/reactos/ntoskrnl/include/internal/i386/ke.h [iso-8859-1] Wed Jan 27 04:05:10 2010
@@ -456,6 +456,9 @@
 extern VOID __cdecl CopyParams(VOID);
 extern VOID __cdecl ReadBatch(VOID);
 extern VOID __cdecl FrRestore(VOID);
+extern CHAR KiSystemCallExitBranch[];
+extern CHAR KiSystemCallExit[];
+extern CHAR KiSystemCallExit2[];

//
 // Trap Macros

Modified: trunk/reactos/ntoskrnl/include/internal/trap_x.h
URL: http://svn.reactos.org/svn/reactos/trunk/reactos/ntoskrnl/include/internal/t...
==============================================================================
--- trunk/reactos/ntoskrnl/include/internal/trap_x.h [iso-8859-1] (original)
+++ trunk/reactos/ntoskrnl/include/internal/trap_x.h [iso-8859-1] Wed Jan 27 04:05:10 2010
@@ -190,7 +190,7 @@
 //
 VOID
 FORCEINLINE
-DECLSPEC_NORETURN
+/* Do not mark this as DECLSPEC_NORETURN because possibly executing code follows it! */
 KiSystemCallReturn(IN PKTRAP_FRAME TrapFrame)
 {
     /* Restore nonvolatiles, EAX, and do a "jump" back to the kernel caller */
@@ -205,6 +205,7 @@
         "movl %c[e](%%esp), %%edx\n"
         "addl $%c[v],%%esp\n" /* A WHOLE *KERNEL* frame since we're not IRET'ing */
         "jmp *%%edx\n"
+        ".globl _KiSystemCallExit2\n_KiSystemCallExit2:\n"
         :
         : "r"(TrapFrame),
           [b] "i"(KTRAP_FRAME_EBX),
@@ -216,7 +217,6 @@
           [v] "i"(KTRAP_FRAME_ESP)
         : "%esp"
     );
-    UNREACHABLE;
 }

VOID
@@ -227,6 +227,7 @@
     /* Regular interrupt exit, but we only restore EAX as a volatile */
     __asm__ __volatile__
     (
+        ".globl _KiSystemCallExit\n_KiSystemCallExit:\n"
         "movl %0, %%esp\n"
         "movl %c[b](%%esp), %%ebx\n"
         "movl %c[s](%%esp), %%esi\n"
@@ -405,6 +406,29 @@
     asm volatile(".byte 0xC4\n.byte 0xC4\n");
 }

+VOID
+FORCEINLINE
+KiUserSystemCall(IN PKTRAP_FRAME TrapFrame)
+{
+    /*
+     * Kernel call or user call?
+     *
+     * This decision is made in inlined assembly because we need to patch
+     * the relative offset of the user-mode jump to point to the SYSEXIT
+     * routine if the CPU supports it. The only way to guarantee that a
+     * relative jnz/jz instruction is generated is to force it with the
+     * inline assembler.
+     */
+    asm volatile
+    (
+        "test $1, %0\n" /* MODE_MASK */
+        ".globl _KiSystemCallExitBranch\n_KiSystemCallExitBranch:\n"
+        "jnz _KiSystemCallExit\n"
+        :
+        : "r"(TrapFrame->SegCs)
+    );
+}
+        
 //
 // Generic Exit Routine
 //
@@ -502,43 +526,35 @@
     /* Check for system call -- a system call skips volatiles! */
     if (__builtin_expect(SkipBits.SkipVolatiles, 0)) /* More INTs than SYSCALLs */
     {
-        /* Kernel call or user call? */
-        if (__builtin_expect(KiUserTrap(TrapFrame), 1)) /* More Ring 3 than 0 */
+        /* User or kernel call? */
+        KiUserSystemCall(TrapFrame);
+        
+        /* Restore EFLags */
+        __writeeflags(TrapFrame->EFlags);
+            
+        /* Call is kernel, so do a jump back since this wasn't a real INT */
+        KiSystemCallReturn(TrapFrame);
+
+        /* If we got here, this is SYSEXIT: are we stepping code? */
+        if (!(TrapFrame->EFlags & EFLAGS_TF))
         {
-            /* Is SYSENTER supported and/or enabled, or are we stepping code? */
-            if (__builtin_expect((KiFastSystemCallDisable) ||
-                                 (TrapFrame->EFlags & EFLAGS_TF), 0))
-            {
-                /* Exit normally */
-                KiSystemCallTrapReturn(TrapFrame);
-            }
-            else
-            {
-                /* Restore user FS */
-                Ke386SetFs(KGDT_R3_TEB | RPL_MASK);
-                
-                /* Remove interrupt flag */
-                TrapFrame->EFlags &= ~EFLAGS_INTERRUPT_MASK;
-                __writeeflags(TrapFrame->EFlags);
-                
-                /* Exit through SYSEXIT */
-                KiSystemCallSysExitReturn(TrapFrame);
-            }
+            /* Restore user FS */
+            Ke386SetFs(KGDT_R3_TEB | RPL_MASK);
+
+            /* Remove interrupt flag */
+            TrapFrame->EFlags &= ~EFLAGS_INTERRUPT_MASK;
+            __writeeflags(TrapFrame->EFlags);
+
+            /* Exit through SYSEXIT */
+            KiSystemCallSysExitReturn(TrapFrame);
         }
-        else
-        {
-            /* Restore EFLags */
-            __writeeflags(TrapFrame->EFlags);
-            
-            /* Call is kernel, so do a jump back since this wasn't a real INT */
-            KiSystemCallReturn(TrapFrame);
-        }  
-    }
-    else
-    {
-        /* Return from interrupt */
-        KiTrapReturn(TrapFrame);
-    }
+        
+        /* Exit through IRETD, either due to debugging or due to lack of SYSEXIT */
+        KiSystemCallTrapReturn(TrapFrame);
+    }
+    
+    /* Return from interrupt */
+    KiTrapReturn(TrapFrame);
 }

//

Modified: trunk/reactos/ntoskrnl/ke/i386/cpu.c
URL: http://svn.reactos.org/svn/reactos/trunk/reactos/ntoskrnl/ke/i386/cpu.c?rev=...
==============================================================================
--- trunk/reactos/ntoskrnl/ke/i386/cpu.c [iso-8859-1] (original)
+++ trunk/reactos/ntoskrnl/ke/i386/cpu.c [iso-8859-1] Wed Jan 27 04:05:10 2010
@@ -45,6 +45,15 @@
 BOOLEAN KiI386PentiumLockErrataPresent;
 BOOLEAN KiSMTProcessorsPresent;

+/* The distance between SYSEXIT and IRETD return modes */
+UCHAR KiSystemCallExitAdjust;
+
+/* The offset that was applied -- either 0 or the value above */
+UCHAR KiSystemCallExitAdjusted;
+
+/* Whether the adjustment was already done once */
+BOOLEAN KiFastCallCopyDoneOnce;
+
 /* Flush data */
 volatile LONG KiTbFlushTimeStamp;

@@ -801,16 +810,92 @@

VOID
 NTAPI
+KiDisableFastSyscallReturn(VOID)
+{
+    /* Was it applied? */
+    if (KiSystemCallExitAdjusted)
+    {        
+        /* Restore the original value */
+        KiSystemCallExitBranch[1] = KiSystemCallExitBranch[1] - KiSystemCallExitAdjusted;
+                
+        /* It's not adjusted anymore */
+        KiSystemCallExitAdjusted = FALSE;
+    }
+}
+
+VOID
+NTAPI
+KiEnableFastSyscallReturn(VOID)
+{
+    /* Check if the patch has already been done */
+    if ((KiSystemCallExitAdjusted == KiSystemCallExitAdjust) &&
+        (KiFastCallCopyDoneOnce))
+    {
+        DPRINT1("SYSEXIT Code Patch was already done!\n");
+        return;
+    }
+    
+    /* Make sure the offset is within the distance of a Jxx SHORT */
+    if ((KiSystemCallExitBranch[1] - KiSystemCallExitAdjust) < 0x80)
+    {
+        /* Remove any existing code patch */
+        DPRINT1("Correct SHORT size found\n");
+        KiDisableFastSyscallReturn();
+        
+        /* We should have a JNZ there */
+        ASSERT(KiSystemCallExitBranch[0] == 0x75);
+
+        /* Do the patch */        
+        DPRINT1("Current jump offset: %lx\n", KiSystemCallExitBranch[1]);
+        KiSystemCallExitAdjusted = KiSystemCallExitAdjust;
+        KiSystemCallExitBranch[1] -= KiSystemCallExitAdjusted;
+        DPRINT1("New jump offset: %lx\n", KiSystemCallExitBranch[1]);
+        
+        /* Remember that we've done it */
+        KiFastCallCopyDoneOnce = TRUE;
+    }
+    else
+    {
+        /* This shouldn't happen unless we've messed the macros up */
+        DPRINT1("Your compiled kernel is broken!\n");
+        DbgBreakPoint();
+    }
+}
+
+VOID
+NTAPI
 KiRestoreFastSyscallReturnState(VOID)
 {
-    /* FIXME: NT has support for SYSCALL, IA64-SYSENTER, etc. */
-
     /* Check if the CPU Supports fast system call */
     if (KeFeatureBits & KF_FAST_SYSCALL)
     {
+        /* Check if it has been disabled */
+        if (!KiFastSystemCallDisable)
+        {
+            /* KiSystemCallExit2 should come BEFORE KiSystemCallExit */
+            DPRINT1("Exit2: %p Exit1: %p\n", KiSystemCallExit2, KiSystemCallExit);
+            ASSERT(KiSystemCallExit2 < KiSystemCallExit);
+            
+            /* It's enabled, so we'll have to do a code patch */
+            KiSystemCallExitAdjust = KiSystemCallExit - KiSystemCallExit2;
+            DPRINT1("SYSENTER Capable Machine. Jump Offset Delta: %lx\n", KiSystemCallExitAdjust);
+        }
+        else
+        {
+            /* Disable fast system call */
+            KeFeatureBits &= ~KF_FAST_SYSCALL;
+        }
+    }
+    
+    /* Now check if all CPUs support fast system call, and the registry allows it */
+    if (KeFeatureBits & KF_FAST_SYSCALL)
+    {
         /* Do an IPI to enable it */
         KeIpiGenericCall(KiLoadFastSyscallMachineSpecificRegisters, 0);
     }
+    
+    /* Perform the code patch that is required */
+    KiEnableFastSyscallReturn();
 }

ULONG_PTR

Modified: trunk/reactos/ntoskrnl/ke/i386/kiinit.c
URL: http://svn.reactos.org/svn/reactos/trunk/reactos/ntoskrnl/ke/i386/kiinit.c?r...
==============================================================================
--- trunk/reactos/ntoskrnl/ke/i386/kiinit.c [iso-8859-1] (original)
+++ trunk/reactos/ntoskrnl/ke/i386/kiinit.c [iso-8859-1] Wed Jan 27 04:05:10 2010
@@ -31,7 +31,6 @@
 NTAPI
 KiInitMachineDependent(VOID)
 {
-    ULONG Protect;
     ULONG CpuCount;
     BOOLEAN FbCaching = FALSE;
     NTSTATUS Status;
@@ -153,12 +152,7 @@
             /* FIXME: Implement and enable XMM Page Zeroing for Mm */

/* Patch the RtlPrefetchMemoryNonTemporal routine to enable it */
-            Protect = MmGetPageProtect(NULL, RtlPrefetchMemoryNonTemporal);
-            MmSetPageProtect(NULL,
-                             RtlPrefetchMemoryNonTemporal,
-                             Protect | PAGE_IS_WRITABLE);
             *(PCHAR)RtlPrefetchMemoryNonTemporal = 0x90;
-            MmSetPageProtect(NULL, RtlPrefetchMemoryNonTemporal, Protect);
         }
     }

@@ -320,6 +314,9 @@
         /* FIXME: TODO */
         DPRINT1("ISR Time Limit not yet supported\n");
     }
+    
+    /* Set CR0 features based on detected CPU */
+    KiSetCR0Bits();
 }

VOID
@@ -400,9 +397,6 @@

/* Detect and set the CPU Type */
     KiSetProcessorType();
-
-    /* Set CR0 features based on detected CPU */
-    KiSetCR0Bits();

/* Check if an FPU is present */
     NpxPresent = KiIsNpxPresent();