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 <stdlib.h>
#if defined(_WIN32)
#include <windows.h>
#endif
-#if defined(i386_TARGET_ARCH)
-static rtsBool execPage (void* addr, int writable);
+#if defined(openbsd_TARGET_OS)
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/mman.h>
+
+/* 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)
}
#endif
+#if defined(powerpc_TARGET_ARCH) || defined(powerpc64_TARGET_ARCH)
+#if !(defined(powerpc_TARGET_ARCH) && defined(linux_TARGET_OS))
+
+/* !!! !!! WARNING: !!! !!!
+ * This structure is accessed from AdjustorAsm.s
+ * Any changes here have to be mirrored in the offsets there.
+ */
+
+typedef struct AdjustorStub {
+#if defined(powerpc_TARGET_ARCH) && defined(darwin_TARGET_OS)
+ unsigned lis;
+ unsigned ori;
+ unsigned lwz;
+ unsigned mtctr;
+ unsigned bctr;
+ StgFunPtr code;
+#elif defined(powerpc64_TARGET_ARCH) && defined(darwin_TARGET_OS)
+ /* powerpc64-darwin: just guessing that it won't use fundescs. */
+ unsigned lis;
+ unsigned ori;
+ unsigned rldimi;
+ unsigned oris;
+ unsigned ori2;
+ unsigned lwz;
+ unsigned mtctr;
+ unsigned bctr;
+ StgFunPtr code;
+#else
+ /* fundesc-based ABIs */
+#define FUNDESCS
+ StgFunPtr code;
+ struct AdjustorStub
+ *toc;
+ void *env;
+#endif
+ StgStablePtr hptr;
+ StgFunPtr wptr;
+ StgInt negative_framesize;
+ StgInt extrawords_plus_one;
+} AdjustorStub;
+
+#endif
+#endif
+
void*
-createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr)
+createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString)
{
void *adjustor = NULL;
<c>: 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 */
adj_code[0x0c] = (unsigned char)0xff; /* jmp %eax */
adj_code[0x0d] = (unsigned char)0xe0;
-
-#if 0
- /* not yet */
- execPage(adjustor,rtsTrue);
-#endif
}
#endif
break;
<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
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.
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 ) */
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;
-
-#if 0
- /* not yet */
- execPage(adjustor,rtsTrue);
-#endif
}
#elif defined(sparc_TARGET_ARCH)
/* Magic constant computed by inspecting the code length of the following
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 */
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;
/* Ensure that instruction cache is consistent with our new code */
__asm__ volatile("call_pal %0" : : "i" (PAL_imb));
}
-#elif defined(powerpc_TARGET_ARCH)
+#elif defined(powerpc_TARGET_ARCH) && defined(linux_TARGET_OS)
/*
- For PowerPC, the following code is used:
+ For PowerPC Linux, the following code is used:
mr r10,r8
mr r9,r7
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
__asm__ volatile ("sync\n\tisync");
}
}
+
+#elif defined(powerpc_TARGET_ARCH) || defined(powerpc64_TARGET_ARCH)
+
+#define OP_LO(op,lo) ((((unsigned)(op)) << 16) | (((unsigned)(lo)) & 0xFFFF))
+#define OP_HI(op,hi) ((((unsigned)(op)) << 16) | (((unsigned)(hi)) >> 16))
+ {
+ AdjustorStub *adjustorStub;
+ int sz = 0, extra_sz, total_sz;
+
+ // from AdjustorAsm.s
+ // not declared as a function so that AIX-style
+ // fundescs can never get in the way.
+ extern void *adjustorCode;
+
+#ifdef FUNDESCS
+ adjustorStub = stgMallocBytes(sizeof(AdjustorStub), "createAdjustor");
+#else
+ adjustorStub = mallocBytesRWX(sizeof(AdjustorStub));
+#endif
+ adjustor = adjustorStub;
+
+ adjustorStub->code = (void*) &adjustorCode;
+
+#ifdef FUNDESCS
+ // function descriptors are a cool idea.
+ // We don't need to generate any code at runtime.
+ adjustorStub->toc = adjustorStub;
+#else
+
+ // no function descriptors :-(
+ // We need to do things "by hand".
+#if defined(powerpc_TARGET_ARCH)
+ // lis r2, hi(adjustorStub)
+ adjustorStub->lis = OP_HI(0x3c40, adjustorStub);
+ // ori r2, r2, lo(adjustorStub)
+ adjustorStub->ori = OP_LO(0x6042, adjustorStub);
+ // lwz r0, code(r2)
+ adjustorStub->lwz = OP_LO(0x8002, (char*)(&adjustorStub->code)
+ - (char*)adjustorStub);
+ // mtctr r0
+ adjustorStub->mtctr = 0x7c0903a6;
+ // bctr
+ adjustorStub->bctr = 0x4e800420;
+#else
+ barf("adjustor creation not supported on this platform");
+#endif
+
+ // Flush the Instruction cache:
+ {
+ int n = sizeof(AdjustorStub)/sizeof(unsigned);
+ unsigned *p = (unsigned*)adjustor;
+ while(n--)
+ {
+ __asm__ volatile ("dcbf 0,%0\n\tsync\n\ticbi 0,%0"
+ : : "r" (p));
+ p++;
+ }
+ __asm__ volatile ("sync\n\tisync");
+ }
+#endif
+
+ printf("createAdjustor: %s\n", typeString);
+ while(*typeString)
+ {
+ char t = *typeString++;
+
+ switch(t)
+ {
+#if defined(powerpc64_TARGET_ARCH)
+ case 'd': sz += 1; break;
+ case 'l': sz += 1; break;
+#else
+ case 'd': sz += 2; break;
+ case 'l': sz += 2; break;
+#endif
+ case 'f': sz += 1; break;
+ case 'i': sz += 1; break;
+ }
+ }
+ extra_sz = sz - 8;
+ if(extra_sz < 0)
+ extra_sz = 0;
+ total_sz = (6 /* linkage area */
+ + 8 /* minimum parameter area */
+ + 2 /* two extra arguments */
+ + extra_sz)*sizeof(StgWord);
+
+ // align to 16 bytes.
+ // AIX only requires 8 bytes, but who cares?
+ total_sz = (total_sz+15) & ~0xF;
+
+ adjustorStub->hptr = hptr;
+ adjustorStub->wptr = wptr;
+ adjustorStub->negative_framesize = -total_sz;
+ adjustorStub->extrawords_plus_one = extra_sz + 1;
+ }
+
#elif defined(ia64_TARGET_ARCH)
/*
Up to 8 inputs are passed in registers. We flush the last two inputs to
#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;
}
}
#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;
}
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;
}
/* Free the stable pointer first..*/
freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x10)));
-#elif defined(powerpc_TARGET_ARCH)
+#elif defined(powerpc_TARGET_ARCH) && defined(linux_TARGET_OS)
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)));
+#elif defined(powerpc_TARGET_ARCH) || defined(powerpc64_TARGET_ARCH)
+ extern void* adjustorCode;
+ if ( ((AdjustorStub*)ptr)->code != (StgFunPtr) &adjustorCode ) {
+ errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
+ return;
+ }
+ freeStablePtr(((AdjustorStub*)ptr)->hptr);
#elif defined(ia64_TARGET_ARCH)
IA64FunDesc *fdesc = (IA64FunDesc *)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]);
stgFree(ptr);
}
-#if defined(i386_TARGET_ARCH)
-/*
- * Function: execPage()
- *
- * Set the executable bit on page containin
- */
-static
-rtsBool
-execPage (void* addr, int writable)
-{
-#if defined(_WIN32)
- SYSTEM_INFO sInfo;
- DWORD dwOldProtect = 0;
-
- /* doesn't return a result, so presumably it can't fail... */
- GetSystemInfo(&sInfo);
-
- if ( VirtualProtect ( (void*)((unsigned long)addr & (sInfo.dwPageSize - 1)),
- sInfo.dwPageSize,
- ( writable ? PAGE_EXECUTE_READWRITE : PAGE_EXECUTE_READ),
- &dwOldProtect) == 0 ) {
-# if 1
- DWORD rc = GetLastError();
- fprintf(stderr, "execPage: failed to protect 0x%p; error=%lu; old protection: %lu\n", addr, rc, dwOldProtect);
-# endif
- return rtsFalse;
- }
- return rtsTrue;
-#else
- return rtsTrue;
-#endif
-}
-#endif
/*
* Function: initAdjustor()
*
* Perform initialisation of adjustor thunk layer (if needed.)
*/
-rtsBool
+void
initAdjustor(void)
{
-#if defined(i386_TARGET_ARCH) && defined(_WIN32)
- return execPage(__obscure_ccall_ret_code, rtsFalse);
-#else
- return rtsTrue;
+#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
}