Merge pull request #1128 from hjelmn/v2_0_0_patcher

v2.0.0 patcher fixes

Author: jsquyres

opal/mca/memory/patcher/configure.m4

@@ -84,6 +84,8 @@ AC_DEFUN([MCA_opal_memory_patcher_CONFIG],[
     AC_DEFINE_UNQUOTED([OPAL_MEMORY_PATCHER_HAVE___SYSCALL], [$memory_patcher_have___syscall],
                        [Whether the internal __syscall call exists])
 
+    AC_CHECK_HEADERS([linux/mman.h])
+
     [$1]
 
     OPAL_VAR_SCOPE_POP

opal/mca/memory/patcher/memory_patcher_component.c

@@ -43,6 +43,10 @@
 #include <sys/time.h>
 #include <sys/syscall.h>
 
+#if defined(HAVE_LINUX_MMAN_H)
+#include <linux/mman.h>
+#endif
+
 #include "memory_patcher.h"
 #undef opal_memory_changed
 
@@ -162,11 +166,20 @@ static int intercept_munmap(void *start, size_t length)
 
 #if defined (SYS_mremap)
 
+#if defined(__linux__)
 /* on linux this function has an optional extra argument but ... can not be used here because it
  * causes issues when intercepting a 4-argument mremap call */
 static void *(*original_mremap) (void *, size_t, size_t, int, void *);
+#else
+/* mremap has a different signature on BSD systems */
+static void *(*original_mremap) (void *, size_t, void *, size_t, int);
+#endif
 
+#if defined(__linux__)
 static void *intercept_mremap (void *start, size_t oldlen, size_t newlen, int flags, void *new_address)
+#else
+static void *intercept_mremap (void *start, size_t oldlen, void *new_address, size_t newlen, int flags)
+#endif
 {
     OPAL_PATCHER_BEGIN;
     void *result = MAP_FAILED;
@@ -175,15 +188,25 @@ static void *intercept_mremap (void *start, size_t oldlen, size_t newlen, int fl
         opal_mem_hooks_release_hook (start, oldlen, true);
     }
 
+#if defined(MREMAP_FIXED)
     if (!(flags & MREMAP_FIXED)) {
         new_address = NULL;
     }
+#endif
 
+#if defined(__linux__)
     if (!original_mremap) {
         result = (void *)(intptr_t) memory_patcher_syscall (SYS_mremap, start, oldlen, newlen, flags, new_address);
     } else {
         result = original_mremap (start, oldlen, newlen, flags, new_address);
     }
+#else
+    if (!original_mremap) {
+        result = (void *)(intptr_t) memory_patcher_syscall (SYS_mremap, start, oldlen, new_address, newlen, flags);
+    } else {
+        result = original_mremap (start, oldlen, new_address, newlen, flags);
+    }
+#endif
 
     OPAL_PATCHER_END;
     return result;

opal/mca/patcher/base/base.h

@@ -68,7 +68,7 @@ OPAL_DECLSPEC int mca_patcher_base_patch_hook (mca_patcher_base_module_t *module
 OPAL_DECLSPEC void mca_base_patcher_patch_apply_binary (mca_patcher_base_patch_t *patch);
 
 static inline uintptr_t mca_patcher_base_addr_text (uintptr_t addr) {
-#if (defined(__PPC64__) || defined(__powerpc64__) || defined(__PPC__)) && _CALL_ELF != 2
+#if (OPAL_ASSEMBLY_ARCH == OPAL_POWERPC64) && (!defined (_CALL_ELF) || (_CALL_ELF != 2))
     struct odp_t {
         uintptr_t text;
         uintptr_t toc;

opal/mca/patcher/base/patcher_base_patch.c

@@ -75,20 +75,37 @@ static int PatchLoadImm (uintptr_t addr, unsigned int reg, size_t value)
 
 #endif
 
-#if defined(__i386__) || defined(__x86_64__) || defined(__ia64__)
-
-static void flush_and_invalidate_cache (unsigned long  a)
+static void flush_and_invalidate_cache (unsigned long a)
 {
-#if defined(__i386__)
-    /* does not work with AMD processors */
-    __asm__ volatile("mfence;clflush %0;mfence" : :"m" (*(char*)a));
-#elif defined(__x86_64__)
+#if OPAL_ASSEMBLY_ARCH == OPAL_IA32
+    static int have_clflush = -1;
+
+    if (OPAL_UNLIKELY(-1 == have_clflush)) {
+        int32_t cpuid1, cpuid2, tmp;
+        const int32_t level = 1;
+
+        /* cpuid clobbers ebx but it must be restored for -fPIC so save
+         * then restore ebx */
+        __asm__ volatile ("xchgl %%ebx, %2\n"
+                          "cpuid\n"
+                          "xchgl %%ebx, %2\n":
+                          "=a" (cpuid1), "=d" (cpuid2), "=r" (tmp) :
+                          "a" (level) :
+                          "ecx");
+        /* clflush is in edx bit 19 */
+        have_clflush = !!(cpuid2 & (1 << 19));
+    }
+
+    if (have_clflush) {
+        /* does not work with AMD processors */
+        __asm__ volatile("mfence;clflush %0;mfence" : :"m" (*(char*)a));
+    }
+#elif OPAL_ASSEMBLY_ARCH == OPAL_AMD64
     __asm__ volatile("mfence;clflush %0;mfence" : :"m" (*(char*)a));
-#elif defined(__ia64__)
+#elif OPAL_ASSEMBLY_ARCH == OPAL_IA64
     __asm__ volatile ("fc %0;; sync.i;; srlz.i;;" : : "r"(a) : "memory");
 #endif
 }
-#endif
 
 // modify protection of memory range
 static void ModifyMemoryProtection (uintptr_t addr, size_t length, int prot)
@@ -105,7 +122,7 @@ static void ModifyMemoryProtection (uintptr_t addr, size_t length, int prot)
         if (mprotect((void *)base, page_size, prot))
             perror("MemHook: mprotect failed");
         base += page_size;
-    } while (base < addr + length);
+    } while (base < bound);
 #else
     if (mprotect((void *) base, length, prot)) {
             perror("MemHook: mprotect failed");
@@ -117,11 +134,9 @@ static inline void apply_patch (unsigned char *patch_data, uintptr_t address, si
 {
     ModifyMemoryProtection (address, data_size, PROT_EXEC|PROT_READ|PROT_WRITE);
     memcpy ((void *) address, patch_data, data_size);
-#if defined(__i386__) || defined(__x86_64__) || defined(__ia64__)
     for (size_t i = 0 ; i < data_size ; i += 16) {
         flush_and_invalidate_cache (address + i);
     }
-#endif
 
     ModifyMemoryProtection (address, data_size, PROT_EXEC|PROT_READ);
 }
@@ -141,26 +156,26 @@ void mca_base_patcher_patch_apply_binary (mca_patcher_base_patch_t *patch)
 
 int mca_patcher_base_patch_hook (mca_patcher_base_module_t *module, uintptr_t hook_addr)
 {
-#if defined(__PPC64__) || defined(__powerpc64__) || defined(__PPC__)
+#if (OPAL_ASSEMBLY_ARCH == OPAL_POWERPC64)
     mca_patcher_base_patch_t *hook_patch;
     const unsigned int nop = 0x60000000;
-    unsigned int *nop_addr;
 
     hook_patch = OBJ_NEW(mca_patcher_base_patch_t);
     if (OPAL_UNLIKELY(NULL == hook_patch)) {
         return OPAL_ERR_OUT_OF_RESOURCE;
     }
 
     // locate reserved code space in hook function
-    for (nop_addr = (unsigned int *)hook_addr ; ; nop_addr++) {
+    for (unsigned int *nop_addr = (unsigned int *)hook_addr ; ; nop_addr++) {
         if (nop_addr[0] == nop && nop_addr[1] == nop && nop_addr[2] == nop
                 && nop_addr[3] == nop && nop_addr[4] == nop) {
+            hook_patch->patch_orig = (uintptr_t) nop_addr;
             break;
         }
     }
+
     // generate code to restore TOC
     register unsigned long toc asm("r2");
-    hook_patch->patch_orig = (uintptr_t) nop_addr;
     hook_patch->patch_data_size = PatchLoadImm((uintptr_t)hook_patch->patch_data, 2, toc);
 
     /* put the hook patch on the patch list so it will be undone on finalize */