X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FAdjustor.c;h=c6729163171fb5066c78d92bc67e61a136c6352f;hb=4b17256112ecbee54dd4437f29c711bbfae239c7;hp=ed3cb507dbea8d27395b4461b791ee0b34b6272d;hpb=3b6a5af52cff72d77600b848d6d851f71b162bc2;p=ghc-hetmet.git diff --git a/ghc/rts/Adjustor.c b/ghc/rts/Adjustor.c index ed3cb50..c672916 100644 --- a/ghc/rts/Adjustor.c +++ b/ghc/rts/Adjustor.c @@ -34,46 +34,73 @@ An adjustor thunk differs from a C function pointer in one respect: when the code is through with it, it has to be freed in order to release Haskell and C resources. Failure to do so result in memory leaks on both the C and Haskell side. - */ + #include "PosixSource.h" #include "Rts.h" +#include "RtsExternal.h" #include "RtsUtils.h" -#include "RtsFlags.h" - #include +#if defined(_WIN32) +#include +#endif + +#if defined(openbsd_TARGET_OS) +#include +#include +#include + +/* no C99 header stdint.h on OpenBSD? */ +typedef unsigned long my_uintptr_t; +#endif + /* Heavily arch-specific, I'm afraid.. */ +/* + * Allocate len bytes which are readable, writable, and executable. + * + * ToDo: If this turns out to be a performance bottleneck, one could + * e.g. cache the last VirtualProtect/mprotect-ed region and do + * nothing in case of a cache hit. + */ +static void* +mallocBytesRWX(int len) +{ + void *addr = stgMallocBytes(len, "mallocBytesRWX"); +#if defined(i386_TARGET_ARCH) && defined(_WIN32) + /* This could be necessary for processors which distinguish between READ and + EXECUTE memory accesses, e.g. Itaniums. */ + DWORD dwOldProtect = 0; + if (VirtualProtect (addr, len, PAGE_EXECUTE_READWRITE, &dwOldProtect) == 0) { + barf("mallocBytesRWX: failed to protect 0x%p; error=%lu; old protection: %lu\n", + addr, (unsigned long)GetLastError(), (unsigned long)dwOldProtect); + } +#elif defined(openbsd_TARGET_OS) + /* malloced memory isn't executable by default on OpenBSD */ + my_uintptr_t pageSize = sysconf(_SC_PAGESIZE); + my_uintptr_t mask = ~(pageSize - 1); + my_uintptr_t startOfFirstPage = ((my_uintptr_t)addr ) & mask; + my_uintptr_t startOfLastPage = ((my_uintptr_t)addr + len - 1) & mask; + my_uintptr_t size = startOfLastPage - startOfFirstPage + pageSize; + if (mprotect((void*)startOfFirstPage, (size_t)size, PROT_EXEC | PROT_READ | PROT_WRITE) != 0) { + barf("mallocBytesRWX: failed to protect 0x%p\n", addr); + } +#endif + return addr; +} + #if defined(i386_TARGET_ARCH) -/* Now here's something obscure for you: - - When generating an adjustor thunk that uses the C calling - convention, we have to make sure that the thunk kicks off - the process of jumping into Haskell with a tail jump. Why? - Because as a result of jumping in into Haskell we may end - up freeing the very adjustor thunk we came from using - freeHaskellFunctionPtr(). Hence, we better not return to - the adjustor code on our way out, since it could by then - point to junk. - - The fix is readily at hand, just include the opcodes - for the C stack fixup code that we need to perform when - returning in some static piece of memory and arrange - to return to it before tail jumping from the adjustor thunk. - - For this to work we make the assumption that bytes in .data - are considered executable. -*/ -static unsigned char __obscure_ccall_ret_code [] = - { 0x83, 0xc4, 0x04 /* addl $0x4, %esp */ - , 0xc3 /* ret */ - }; +static unsigned char *obscure_ccall_ret_code; #endif #if defined(alpha_TARGET_ARCH) /* To get the definition of PAL_imb: */ -#include +# if defined(linux_TARGET_OS) +# include +# else +# include +# endif #endif #if defined(ia64_TARGET_ARCH) @@ -128,7 +155,8 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) : ff e0 jmp %eax # and jump to it. # the callee cleans up the stack */ - if ((adjustor = stgMallocBytes(14, "createAdjustor")) != NULL) { + adjustor = mallocBytesRWX(14); + { unsigned char *const adj_code = (unsigned char *)adjustor; adj_code[0x00] = (unsigned char)0x58; /* popl %eax */ @@ -155,7 +183,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) <00>: 68 ef be ad de pushl $0xdeadbeef # constant is large enough to # hold a StgStablePtr <05>: b8 fa ef ff 00 movl $0x00ffeffa, %eax # load up wptr - <0a>: 68 ef be ad de pushl $__obscure_ccall_ret_code # push the return address + <0a>: 68 ef be ad de pushl $obscure_ccall_ret_code # push the return address <0f>: ff e0 jmp *%eax # jump to wptr The ccall'ing version is a tad different, passing in the return @@ -163,7 +191,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) via the stable pointer.) (The auto-generated C stub is in on this game, don't worry :-) - See the comment next to __obscure_ccall_ret_code why we need to + See the comment next to obscure_ccall_ret_code why we need to perform a tail jump instead of a call, followed by some C stack fixup. @@ -172,7 +200,8 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) That's (thankfully) the case here with the restricted set of return types that we support. */ - if ((adjustor = stgMallocBytes(17, "createAdjustor")) != NULL) { + adjustor = mallocBytesRWX(17); + { unsigned char *const adj_code = (unsigned char *)adjustor; adj_code[0x00] = (unsigned char)0x68; /* pushl hptr (which is a dword immediate ) */ @@ -181,8 +210,8 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) adj_code[0x05] = (unsigned char)0xb8; /* movl $wptr, %eax */ *((StgFunPtr*)(adj_code + 0x06)) = (StgFunPtr)wptr; - adj_code[0x0a] = (unsigned char)0x68; /* pushl __obscure_ccall_ret_code */ - *((StgFunPtr*)(adj_code + 0x0b)) = (StgFunPtr)__obscure_ccall_ret_code; + adj_code[0x0a] = (unsigned char)0x68; /* pushl obscure_ccall_ret_code */ + *((StgFunPtr*)(adj_code + 0x0b)) = (StgFunPtr)obscure_ccall_ret_code; adj_code[0x0f] = (unsigned char)0xff; /* jmp *%eax */ adj_code[0x10] = (unsigned char)0xe0; @@ -217,7 +246,8 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) similarly, and local variables should be accessed via %fp, not %sp. In a nutshell: This should work! (Famous last words! :-) */ - if ((adjustor = stgMallocBytes(4*(11+1), "createAdjustor")) != NULL) { + adjustor = mallocBytesRWX(4*(11+1)); + { unsigned long *const adj_code = (unsigned long *)adjustor; adj_code[ 0] = 0x9C23A008UL; /* sub %sp, 8, %sp */ @@ -293,7 +323,8 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for 4 bytes (getting rid of the nop), hence saving memory. [ccshan] */ ASSERT(((StgWord64)wptr & 3) == 0); - if ((adjustor = stgMallocBytes(48, "createAdjustor")) != NULL) { + adjustor = mallocBytesRWX(48); + { StgWord64 *const code = (StgWord64 *)adjustor; code[0] = 0x4610041246520414L; @@ -330,7 +361,8 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for this code, it only works for up to 6 arguments (when floating point arguments are involved, this may be more or less, depending on the exact situation). */ - if ((adjustor = stgMallocBytes(4*13, "createAdjustor")) != NULL) { + adjustor = mallocBytesRWX(4*13); + { unsigned long *const adj_code = (unsigned long *)adjustor; // make room for extra arguments @@ -366,7 +398,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for while(n--) { __asm__ volatile ("dcbf 0,%0\n\tsync\n\ticbi 0,%0" - : : "g" (p)); + : : "r" (p)); p++; } __asm__ volatile ("sync\n\tisync"); @@ -472,7 +504,7 @@ freeHaskellFunctionPtr(void* ptr) #if defined(i386_TARGET_ARCH) if ( *(unsigned char*)ptr != 0x68 && *(unsigned char*)ptr != 0x58 ) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } @@ -484,7 +516,7 @@ freeHaskellFunctionPtr(void* ptr) } #elif defined(sparc_TARGET_ARCH) if ( *(unsigned long*)ptr != 0x9C23A008UL ) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } @@ -492,7 +524,7 @@ freeHaskellFunctionPtr(void* ptr) freeStablePtr(*((StgStablePtr*)((unsigned long*)ptr + 11))); #elif defined(alpha_TARGET_ARCH) if ( *(StgWord64*)ptr != 0xa77b0018a61b0010L ) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } @@ -500,7 +532,7 @@ freeHaskellFunctionPtr(void* ptr) freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x10))); #elif defined(powerpc_TARGET_ARCH) if ( *(StgWord*)ptr != 0x7d0a4378 ) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 4*12))); @@ -509,7 +541,7 @@ freeHaskellFunctionPtr(void* ptr) StgWord64 *code = (StgWord64 *)(fdesc+1); if (fdesc->ip != (StgWord64)code) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } freeStablePtr((StgStablePtr)code[16]); @@ -520,6 +552,42 @@ freeHaskellFunctionPtr(void* ptr) #endif *((unsigned char*)ptr) = '\0'; - free(ptr); + stgFree(ptr); } + +/* + * Function: initAdjustor() + * + * Perform initialisation of adjustor thunk layer (if needed.) + */ +void +initAdjustor(void) +{ +#if defined(i386_TARGET_ARCH) + /* Now here's something obscure for you: + + When generating an adjustor thunk that uses the C calling + convention, we have to make sure that the thunk kicks off + the process of jumping into Haskell with a tail jump. Why? + Because as a result of jumping in into Haskell we may end + up freeing the very adjustor thunk we came from using + freeHaskellFunctionPtr(). Hence, we better not return to + the adjustor code on our way out, since it could by then + point to junk. + + The fix is readily at hand, just include the opcodes + for the C stack fixup code that we need to perform when + returning in some static piece of memory and arrange + to return to it before tail jumping from the adjustor thunk. + */ + + obscure_ccall_ret_code = mallocBytesRWX(4); + + obscure_ccall_ret_code[0x00] = (unsigned char)0x83; /* addl $0x4, %esp */ + obscure_ccall_ret_code[0x01] = (unsigned char)0xc4; + obscure_ccall_ret_code[0x02] = (unsigned char)0x04; + + obscure_ccall_ret_code[0x03] = (unsigned char)0xc3; /* ret */ +#endif +}