Kernel - Fixing very slow CheckMem function

- Blit from AxWin3 is down to 30ms in qemu :)

diff --git a/Kernel/arch/armv7/mm_virt.c b/Kernel/arch/armv7/mm_virt.c
index 64b5d57..d74313f 100644 (file)
--- a/Kernel/arch/armv7/mm_virt.c
+++ b/Kernel/arch/armv7/mm_virt.c
@@ -391,6 +391,33 @@ void MM_SetFlags(tVAddr VAddr, Uint Flags, Uint Mask)
        MM_int_SetPageInfo(VAddr, &pi);
 }
 
+int MM_IsValidBuffer(tVAddr Addr, size_t Size)
+{
+       tMM_PageInfo    pi;
+        int    bUser = 0;
+       
+       Size += Addr & (PAGE_SIZE-1);
+       Addr &= ~(PAGE_SIZE-1);
+
+       if( MM_int_GetPageInfo(Addr, &pi) )     return 0;
+       Addr += PAGE_SIZE;
+
+       if(pi.AP != AP_KRW_ONLY && pi.AP != AP_KRO_ONLY)
+               bUser = 1;
+
+       while( Size >= PAGE_SIZE )
+       {
+               if( MM_int_GetPageInfo(Addr, &pi) )
+                       return 0;
+               if(bUser && (pi.AP == AP_KRW_ONLY || pi.AP == AP_KRO_ONLY))
+                       return 0;
+               Addr += PAGE_SIZE;
+               Size -= PAGE_SIZE;
+       }
+       
+       return 1;
+}
+
 int MM_Map(tVAddr VAddr, tPAddr PAddr)
 {
        tMM_PageInfo    pi = {0};
@@ -923,6 +950,8 @@ void MM_DumpTables(tVAddr Start, tVAddr End)
         int    i = 0, inRange=0;
        
        pi_old.Size = 0;
+       pi_old.AP = 0;
+       pi_old.PhysAddr = 0;
 
        Debug("Page Table Dump (%p to %p):", Start, End);
        range_start = Start;

diff --git a/Kernel/arch/armv7/proc.c b/Kernel/arch/armv7/proc.c
index 54951e1..b5a867c 100644 (file)
--- a/Kernel/arch/armv7/proc.c
+++ b/Kernel/arch/armv7/proc.c
@@ -13,7 +13,7 @@
 extern tThread gThreadZero;
 extern void    SwitchTask(Uint32 NewSP, Uint32 *OldSP, Uint32 NewIP, Uint32 *OldIP, Uint32 MemPtr);
 extern void    KernelThreadHeader(void);       // Actually takes args on stack
-extern void    Proc_int_DropToUser(Uint32 IP, Uint32 SP) NORETURN;
+extern void    Proc_int_DropToUser(Uint32 IP, Uint32 SP) NORETURN __attribute__((long_call));
 extern Uint32  Proc_int_SwapUserSP(Uint32 NewSP);
 extern Uint32  Proc_CloneInt(Uint32 *SP, Uint32 *MemPtr);
 extern tVAddr  MM_NewKStack(int bGlobal);      // TODO: Move out into a header
@@ -91,12 +91,7 @@ void Proc_StartUser(Uint Entrypoint, Uint Base, int ArgC, char **ArgV, int DataS
        
        // Drop to user code
        Log_Debug("Proc", "Proc_int_DropToUser(%p, %p)", Entrypoint, usr_sp);
-
-       // Needed to get around relocation truncation
-       {
-               void    (*drop)(Uint32, Uint32) NORETURN = Proc_int_DropToUser;
-               drop(Entrypoint, (Uint32)usr_sp);
-       }
+       Proc_int_DropToUser(Entrypoint, (Uint32)usr_sp);
 }
 
 void Proc_ClearThread(tThread *Thread)

diff --git a/Kernel/arch/x86/lib.c b/Kernel/arch/x86/lib.c
index db05d93..5c2bf64 100644 (file)
--- a/Kernel/arch/x86/lib.c
+++ b/Kernel/arch/x86/lib.c
@@ -334,29 +334,30 @@ void *memcpy(void *Dest, const void *Src, size_t Num)
        #if 1
        else if( Num > 128 && (dst & 15) == (src & 15) )
        {
+               char    tmp[16+15];     // Note, this is a hack to save/restor xmm0
                 int    count = 16 - (dst & 15);
 //             Debug("\nmemcpy:Num=0x%x by %p (SSE)", Num, __builtin_return_address(0));
                if( count < 16 )
                {
                        Num -= count;
-//                     Debug("dst = %p, src = %p, count = %i (Num=0x%x) (head)", dst, src, count, Num);
                        __asm__ __volatile__ ("rep movsb" : "=D"(dst),"=S"(src): "0"(dst), "1"(src), "c"(count));
                }
                
                count = Num / 16;
-//             Debug("dst = %p, src = %p, count = %i (bulk)", dst, src, count);
                __asm__ __volatile__ (
+                       "movdqa 0(%5), %%xmm0;\n\t"
                        "1:\n\t"
                        "movdqa 0(%1), %%xmm0;\n\t"
                        "movdqa %%xmm0, 0(%0);\n\t"
                        "add $16,%0;\n\t"
                        "add $16,%1;\n\t"
                        "loop 1b;\n\t"
-                       : "=r"(dst),"=r"(src) : "0"(dst), "1"(src), "c"(count)
+                       "movdqa %%xmm0, 0(%5);\n\t"
+                       : "=r"(dst),"=r"(src)
+                       : "0"(dst), "1"(src), "c"(count), "r" (((tVAddr)tmp+15)&~15)
                        );
 
                count = Num & 15;
-//             Debug("dst = %p, src = %p, count = %i (tail)", dst, src, count);
                if(count)
                        __asm__ __volatile__ ("rep movsb" :: "D"(dst), "S"(src), "c"(count));
        }
@@ -372,6 +373,7 @@ void *memcpy(void *Dest, const void *Src, size_t Num)
        }
        return Dest;
 }
+
 /**
  * \fn void *memcpyd(void *Dest, const void *Src, size_t Num)
  * \brief Copy \a Num DWORDs from \a Src to \a Dest

diff --git a/Kernel/arch/x86/mm_virt.c b/Kernel/arch/x86/mm_virt.c
index e616fc7..e2442ae 100644 (file)
--- a/Kernel/arch/x86/mm_virt.c
+++ b/Kernel/arch/x86/mm_virt.c
@@ -827,6 +827,45 @@ Uint MM_GetFlags(tVAddr VAddr)
        return ret;
 }
 
+/**
+ * \brief Check if the provided buffer is valid
+ * \return Boolean valid
+ */
+int MM_IsValidBuffer(tVAddr Addr, size_t Size)
+{
+        int    bIsUser;
+        int    dir, tab;
+
+       Size += Addr & (PAGE_SIZE-1);
+       Addr &= ~(PAGE_SIZE-1);
+
+       dir = Addr >> 22;
+       tab = Addr >> 12;
+       
+//     Debug("Addr = %p, Size = 0x%x, dir = %i, tab = %i", Addr, Size, dir, tab);
+
+       if( !(gaPageDir[dir] & 1) )     return 0;
+       if( !(gaPageTable[tab] & 1) )   return 0;
+       
+       bIsUser = !!(gaPageTable[tab] & PF_USER);
+
+       while( Size >= PAGE_SIZE )
+       {
+               if( (tab & 1023) == 0 )
+               {
+                       dir ++;
+                       if( !(gaPageDir[dir] & 1) )     return 0;
+               }
+               
+               if( !(gaPageTable[tab] & 1) )   return 0;
+               if( bIsUser && !(gaPageTable[tab] & PF_USER) )  return 0;
+
+               tab ++;
+               Size -= PAGE_SIZE;
+       }
+       return 1;
+}
+
 /**
  * \fn tPAddr MM_DuplicatePage(tVAddr VAddr)
  * \brief Duplicates a virtual page to a physical one

diff --git a/Kernel/arch/x86_64/mm_virt.c b/Kernel/arch/x86_64/mm_virt.c
index c7b7a22..8b56043 100644 (file)
--- a/Kernel/arch/x86_64/mm_virt.c
+++ b/Kernel/arch/x86_64/mm_virt.c
@@ -8,6 +8,7 @@
 #include <mm_virt.h>
 #include <threads_int.h>
 #include <proc.h>
+#include <hal_proc.h>
 
 // === CONSTANTS ===
 #define PHYS_BITS      52      // TODO: Move out
@@ -64,13 +65,13 @@ void        MM_InitVirt(void);
 void   MM_int_ClonePageEnt( Uint64 *Ent, void *NextLevel, tVAddr Addr, int bTable );
  int   MM_PageFault(tVAddr Addr, Uint ErrorCode, tRegs *Regs);
 void   MM_int_DumpTablesEnt(tVAddr RangeStart, size_t Length, tPAddr Expected);
-void   MM_DumpTables(tVAddr Start, tVAddr End);
+//void MM_DumpTables(tVAddr Start, tVAddr End);
  int   MM_GetPageEntryPtr(tVAddr Addr, BOOL bTemp, BOOL bAllocate, BOOL bLargePage, tPAddr **Pointer);
  int   MM_MapEx(tVAddr VAddr, tPAddr PAddr, BOOL bTemp, BOOL bLarge);
 // int MM_Map(tVAddr VAddr, tPAddr PAddr);
 void   MM_Unmap(tVAddr VAddr);
 void   MM_int_ClearTableLevel(tVAddr VAddr, int LevelBits, int MaxEnts);
-void   MM_ClearUser(void);
+//void MM_ClearUser(void);
  int   MM_GetPageEntry(tVAddr Addr, tPAddr *Phys, Uint *Flags);
 
 // === GLOBALS ===
@@ -669,6 +670,60 @@ Uint MM_GetFlags(tVAddr VAddr)
        return ret;
 }
 
+/**
+ * \brief Check if the provided buffer is valid
+ * \return Boolean valid
+ */
+int MM_IsValidBuffer(tVAddr Addr, size_t Size)
+{
+        int    bIsUser;
+        int    pml4, pdp, dir, tab;
+
+       Size += Addr & (PAGE_SIZE-1);
+       Addr &= ~(PAGE_SIZE-1);
+       Addr &= ((1UL << 48)-1);        // Clap to address space
+
+       pml4 = Addr >> 39;
+       pdp = Addr >> 30;
+       dir = Addr >> 21;
+       tab = Addr >> 12;
+       
+//     Debug("Addr = %p, Size = 0x%x, dir = %i, tab = %i", Addr, Size, dir, tab);
+
+       if( !(PAGEMAPLVL4(pml4) & 1) )  return 0;
+       if( !(PAGEDIRPTR(pdp) & 1) )    return 0;
+       if( !(PAGEDIR(dir) & 1) )       return 0;
+       if( !(PAGETABLE(tab) & 1) )     return 0;
+       
+       bIsUser = !!(PAGETABLE(tab) & PF_USER);
+
+       while( Size >= PAGE_SIZE )
+       {
+               if( (tab & 511) == 0 )
+               {
+                       dir ++;
+                       if( ((dir >> 9) & 511) == 0 )
+                       {
+                               pdp ++;
+                               if( ((pdp >> 18) & 511) == 0 )
+                               {
+                                       pml4 ++;
+                                       if( !(PAGEMAPLVL4(pml4) & 1) )  return 0;
+                               }
+                               if( !(PAGEDIRPTR(pdp) & 1) )    return 0;
+                       }
+                       if( !(PAGEDIR(dir) & 1) )       return 0;
+               }
+               
+               if( !(PAGETABLE(tab) & 1) )   return 0;
+               if( bIsUser && !(PAGETABLE(tab) & PF_USER) )    return 0;
+
+               tab ++;
+               Size -= PAGE_SIZE;
+       }
+       return 1;
+}
+
 // --- Hardware Mappings ---
 /**
  * \brief Map a range of hardware pages

diff --git a/Kernel/drv/vterm.c b/Kernel/drv/vterm.c
index 634865b..9488b3e 100644 (file)
--- a/Kernel/drv/vterm.c
+++ b/Kernel/drv/vterm.c
@@ -589,7 +589,8 @@ Uint64 VT_Write(tVFS_Node *Node, Uint64 Offset, Uint64 Length, void *Buffer)
                                }
                                return 0;
                        }
-                       else {
+                       else
+                       {
                                return VFS_WriteAt( giVT_OutputDevHandle, Offset, Length, Buffer );
                        }
                }

diff --git a/Kernel/include/hal_proc.h b/Kernel/include/hal_proc.h
index 0df0556..a189d56 100644 (file)
--- a/Kernel/include/hal_proc.h
+++ b/Kernel/include/hal_proc.h
@@ -80,4 +80,11 @@ extern void  MM_ClearUser(void);
  */
 extern void    MM_DumpTables(tVAddr Start, tVAddr End);
 
+/**
+ * \brief Check if a buffer is valid (and all user if originally user)
+ * \param Addr Base address
+ * \param Size Size of the buffer in bytes
+ * \return Boolean valid (0: invalid, non-0: Valid)
+ */
+extern int     MM_IsValidBuffer(tVAddr Addr, size_t Size);
 #endif

diff --git a/Kernel/lib.c b/Kernel/lib.c
index 73b6432..369b7b0 100644 (file)
--- a/Kernel/lib.c
+++ b/Kernel/lib.c
@@ -3,6 +3,7 @@
  * Common Library Functions
  */
 #include <acess.h>
+#include <hal_proc.h>
 
 // === CONSTANTS ===
 #define        RANDOM_SEED     0xACE55052
@@ -821,35 +822,29 @@ int rand(void)
  */
 int CheckString(const char *String)
 {
-       if( !MM_GetPhysAddr( (tVAddr)String ) )
+       tVAddr  addr;
+        int    bUser;
+
+       addr = (tVAddr)String;
+
+       if( !MM_GetPhysAddr( addr ) )
                return 0;
        
        // Check 1st page
-       if( MM_IsUser( (tVAddr)String ) )
+       bUser = MM_IsUser( addr );
+       
+       while( *(char*)addr )
        {
-               // Traverse String
-               while( *String )
+               if( (addr & (PAGE_SIZE-1)) == 0 )
                {
-                       if( !MM_IsUser( (tVAddr)String ) )
+                       if(bUser && !MM_IsUser(addr) )
                                return 0;
-                       // Increment string pointer
-                       String ++;
-               }
-               return 1;
-       }
-       else if( MM_GetPhysAddr( (tVAddr)String ) )
-       {
-               // Traverse String
-               while( *String )
-               {
-                       if( !MM_GetPhysAddr( (tVAddr)String ) )
+                       if(!bUser && !MM_GetPhysAddr(addr) )
                                return 0;
-                       // Increment string pointer
-                       String ++;
                }
-               return 1;
+               addr ++;
        }
-       return 0;
+       return 1;
 }
 
 /**
@@ -858,32 +853,7 @@ int CheckString(const char *String)
  */
 int CheckMem(const void *Mem, int NumBytes)
 {
-       tVAddr  addr = (tVAddr)Mem;
-
-       if( !MM_GetPhysAddr( addr ) )
-               return 0;
-       
-       if( MM_IsUser( addr ) )
-       {
-               while( NumBytes-- )
-               {
-                       if( !MM_IsUser( addr ) )
-                               return 0;
-                       addr ++;
-               }
-               return 1;
-       }
-       else if( MM_GetPhysAddr( addr ) )
-       {
-               while( NumBytes-- )
-               {
-                       if( !MM_GetPhysAddr( addr ) )
-                               return 0;
-                       addr ++;
-               }
-               return 1;
-       }
-       return 0;
+       return MM_IsValidBuffer( (tVAddr)Mem, NumBytes );
 }
 /* *
  * \}