* bug#47429: Re 47429:allocating JIT code buffer failed: Permission denied
2021-03-26 22:48 bug#47429: allocating JIT code buffer failed: Permission denied Jeffrey Walton
@ 2021-03-26 23:27 ` Jeffrey Walton
2021-03-26 23:34 ` Jeffrey Walton
` (2 subsequent siblings)
3 siblings, 0 replies; 5+ messages in thread
From: Jeffrey Walton @ 2021-03-26 23:27 UTC (permalink / raw)
To: 47429
I'm looking at OS X's man page on mmap. The EACCES does not seem to
fit one of the stated reasons in the man page. Also see
https://developer.apple.com/library/archive/documentation/System/Conceptual/ManPages_iPhoneOS/man2/mmap.2.html.
I have two theories, both of which are guesses. Neither seems to be
very good. First, this may be related to W^X pages on OS X.
Second, the size is 0 but the wrong error code is returned. For the
second case, size=0 should result in EINVAL. I don't know what size is
so this is probably a bad guess.
I'm not sure how to get this under a debugger.
Here's a test of the first theory:
% cat mmap-test.c
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <sys/mman.h>
int main(int argc, char* argv[])
{
size_t len = 10;
int prot = PROT_EXEC | PROT_READ | PROT_WRITE;
int flags = MAP_PRIVATE | MAP_ANONYMOUS;
void *p = mmap (NULL, len, prot, flags, -1, 0);
int err = errno;
if (p) {
printf("p is good\n");
munmap(p, len);
}
else {
printf("p is bad (%d)\n", err);
}
return 0;
}
% clang -Wall mmap-test.c -o mmap-test.exe
% ./mmap-test.exe
p is good
Here's a test of the second theory with len = 0:
% ./mmap-test.exe
p is good
I'm out of ideas...
=========================
This is the relevant part of libguile/jit.c:
1330 static struct code_arena *
1331 allocate_code_arena (size_t size, struct code_arena *prev)
1332 {
1333 struct code_arena *ret = malloc (sizeof (struct code_arena));
1334
1335 if (!ret) return NULL;
1336
1337 memset (ret, 0, sizeof (*ret));
1338 ret->used = 0;
1339 ret->size = size;
1340 ret->prev = prev;
1341 ret->base = mmap (NULL, ret->size,
1342 PROT_EXEC | PROT_READ | PROT_WRITE,
1343 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
1344
1345 if (ret->base == MAP_FAILED)
1346 {
1347 perror ("allocating JIT code buffer failed");
1348 free (ret);
1349 return NULL;
1350 }
1351
1352 INFO ("allocated code arena, %p-%p\n", ret->base, ret->base
+ ret->size);
1353
1354 return ret;
1355 }
^ permalink raw reply [flat|nested] 5+ messages in thread* bug#47429: Re 47429:allocating JIT code buffer failed: Permission denied
2021-03-26 22:48 bug#47429: allocating JIT code buffer failed: Permission denied Jeffrey Walton
2021-03-26 23:27 ` bug#47429: Re 47429:allocating " Jeffrey Walton
@ 2021-03-26 23:34 ` Jeffrey Walton
2021-03-27 1:29 ` Jeffrey Walton
2021-03-27 2:39 ` Jeffrey Walton
3 siblings, 0 replies; 5+ messages in thread
From: Jeffrey Walton @ 2021-03-26 23:34 UTC (permalink / raw)
To: 47429
Dammit, check for the proper return value...
% cat mmap-test.c
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <sys/mman.h>
int main(int argc, char* argv[])
{
size_t len = 10;
int prot = PROT_EXEC | PROT_READ | PROT_WRITE;
int flags = MAP_PRIVATE | MAP_ANONYMOUS;
void *p = mmap (NULL, len, prot, flags, -1, 0);
int err = errno;
if (p != (void*) -1) {
printf("p is good\n");
munmap(p, len);
}
else {
printf("p is bad (%d)\n", err);
}
return 0;
}
% clang -Wall mmap-test.c -o mmap-test.exe
% ./mmap-test.exe
p is bad (13)
That is the permission denied.
Next, avoid W+X:
% cat mmap-test.c
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <sys/mman.h>
int main(int argc, char* argv[])
{
size_t len = 10;
int prot = /*PROT_EXEC |*/ PROT_READ | PROT_WRITE;
int flags = MAP_PRIVATE | MAP_ANONYMOUS;
void *p = mmap (NULL, len, prot, flags, -1, 0);
int err = errno;
if (p != (void*) -1) {
printf("p is good\n");
munmap(p, len);
}
else {
printf("p is bad (%d)\n", err);
}
return 0;
}
It looks like W^X is the culprit.
Jeff
^ permalink raw reply [flat|nested] 5+ messages in thread* bug#47429: Re 47429:allocating JIT code buffer failed: Permission denied
2021-03-26 22:48 bug#47429: allocating JIT code buffer failed: Permission denied Jeffrey Walton
2021-03-26 23:27 ` bug#47429: Re 47429:allocating " Jeffrey Walton
2021-03-26 23:34 ` Jeffrey Walton
@ 2021-03-27 1:29 ` Jeffrey Walton
2021-03-27 2:39 ` Jeffrey Walton
3 siblings, 0 replies; 5+ messages in thread
From: Jeffrey Walton @ 2021-03-27 1:29 UTC (permalink / raw)
To: 47429
This looks like the Apple article of interest:
https://developer.apple.com/documentation/apple-silicon/porting-just-in-time-compilers-to-apple-silicon.
According to the article, the page should be mapped with MAP_JIT.
Before instructions are written, pthread_jit_write_protect_np should
be called. After writing the jitted code, call
pthread_jit_write_protect_np again and call sys_icache_invalidate.
It looks like the OpenJDK folks are also experiencing similar issues.
Jeff
^ permalink raw reply [flat|nested] 5+ messages in thread
* bug#47429: Re 47429:allocating JIT code buffer failed: Permission denied
2021-03-26 22:48 bug#47429: allocating JIT code buffer failed: Permission denied Jeffrey Walton
` (2 preceding siblings ...)
2021-03-27 1:29 ` Jeffrey Walton
@ 2021-03-27 2:39 ` Jeffrey Walton
3 siblings, 0 replies; 5+ messages in thread
From: Jeffrey Walton @ 2021-03-27 2:39 UTC (permalink / raw)
To: 47429
The GUILE_ENABLE_JIT m4 macro should _not_enable JIT on the Apple
silicon. That's going to take a real port.
This may help in detecting the M1 as long as build==host.
apple_silicon=`sysctl machdep.cpu.brand_string 2>/dev/null | grep
-i -c "Apple M1"`
If a cross-compile is happening, then TARGET_OSX and MAP_JIT macros may help.
Jeff
^ permalink raw reply [flat|nested] 5+ messages in thread