Possible Bug in mincore or mmap
When testing an installation with tests from the Linux Test Project, my kernels fail one instance of the mincore01 tests:
mincoremincore01 1 PASS : expected failure: errno = 22 (Invalid
argument)
mincore01 2 PASS : expected failure: errno = 14 (Bad address) mincore01 3 FAIL : call succeeded unexpectedly mincore01 4 PASS : expected failure: errno = 12 (Cannot allocate memory)01 1 PASS : expected failure: errno = 22 (Invalid argument) mincore01 2 PASS : expected failure: errno = 14 (Bad address) mincore01 3 FAIL : call succeeded unexpectedly mincore01 4 PASS : expected failure: errno = 12 (Cannot allocate
memory)
I pared down the test to the attached program. The result is supposed to fail as it is asking for memory information 5 times what should be allocated.
Upon experimenting, I found the test works properly if a printf is executed before the mmap call. I have tested on locally built, but unmodified, 2.4.25, 2.6.12.5, and a 2.6.20.3 kernels and get the same behavior. The tests fail on IA32 architecture, but not 64-bit kernels.
The test always works properly on FC6 and RHEL3.
I've checked the archives for this issue and could not find anything appropriate.
Could this be a potential security issue as memory that is not supposed to be accessible seems to be available to the user? Is it expected behavior?
Thanks.
-- Bruce
#include <sys/mman.h>
#include <stdlib.h>
#include <string.h>
#include <stdio.h>
static int PAGESIZE;
static char file_name[] = "fooXXXXXX";
static void* global_pointer = NULL;
static int global_len = 0; static int file_desc = 0;
int main(int argc, char **argv)
{
int i;
int result;
char* buf;
unsigned char vect[20] = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
PAGESIZE = getpagesize();
/* global_pointer will point to a mmapped area of global_len bytes */
global_len = PAGESIZE*2;
buf = (char*)malloc(global_len);
memset(buf, 42, global_len); // Asterisks
/* create a temporary file */
file_desc = mkstemp(file_name);
/* fill the temporary file with two pages of data */
write(file_desc, buf, global_len);
free(buf);
// Will work properly as long as print is before mmap function.
if ( argc > 1 ) printf("argc=%d\n", argc);
/* map the file in memory */
global_pointer = mmap( NULL, global_len,
PROT_READ|PROT_WRITE|PROT_EXEC, MAP_SHARED, file_desc, 0);
// Result should be -1 as the request is 5 times actual mapping
result = mincore(global_pointer, (size_t)(global_len*5), vect);
// Print some data
printf("PAGESIZE=%d\n", PAGESIZE);
printf("global_len=%d\n", global_len);
printf("global_pointer=0x%x\n", (unsigned int)global_pointer);
printf("alloc=%d\n", (global_len+PAGESIZE-1) / PAGESIZE );
printf("Result=%d\n", result);
printf("vect: ");
for ( i=0; i<20; i++) printf("%02x ", vect[i]);
printf("\n");
// Clean up
munmap(global_pointer, (size_t)global_len);
close(file_desc);
unlink(file_name);
}Quoted 24 lines · Bruce Dubbs
When testing an installation with tests from the Linux Test Project, my kernels fail one instance of the mincore01 tests:
mincoremincore01 1 PASS : expected failure: errno = 22 (Invalid
argument)mincore01 2 PASS : expected failure: errno = 14 (Bad address) mincore01 3 FAIL : call succeeded unexpectedly mincore01 4 PASS : expected failure: errno = 12 (Cannot allocate memory)01 1 PASS : expected failure: errno = 22 (Invalid argument) mincore01 2 PASS : expected failure: errno = 14 (Bad address) mincore01 3 FAIL : call succeeded unexpectedly mincore01 4 PASS : expected failure: errno = 12 (Cannot allocatememory)
I pared down the test to the attached program. The result is supposed to fail as it is asking for memory information 5 times what should be allocated.
Upon experimenting, I found the test works properly if a printf is executed before the mmap call. I have tested on locally built, but unmodified, 2.4.25, 2.6.12.5, and a 2.6.20.3 kernels and get the same behavior. The tests fail on IA32 architecture, but not 64-bit kernels.
The test always works properly on FC6 and RHEL3.
Quoted 6 lines
I've checked the archives for this issue and could not find anything appropriate.
Could this be a potential security issue as memory that is not supposed to be accessible seems to be available to the user? Is it expected behavior?
It shouldn't be a security problem if mincore doesn't actually return the data.
The other thing is, that test may not be valid, because it doesn't guarantee that you have nothing mapped immediately after the global_pointer region. Maybe a difference in address space layout is causing it to "correctly" fail on x86-64.
Quoted 68 lines
Thanks.
-- Bruce
------------------------------------------------------------------------
#include <
sys/mman.h>
#include <stdlib.h>
#include <string.h>
#include <stdio.h>static int PAGESIZE;
static char file_name[] = "fooXXXXXX";
static void*global_pointer= NULL;static int global_len = 0; static int file_desc = 0;int
main(int argc, char **argv){ int i; int result; char* buf; unsigned char vect[20] = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0}; PAGESIZE = getpagesize(); /* global_pointer will point to a mmapped area of global_len bytes */ global_len = PAGESIZE*2; buf = (char*)malloc(global_len); memset(buf, 42, global_len); // Asterisks /* create a temporary file */ file_desc = mkstemp(file_name); /* fill the temporary file with two pages of data */ write(file_desc, buf, global_len); free(buf); // Will work properly as long as print is before mmap function. if ( argc > 1 ) printf("argc=%d\n", argc); /* map the file in memory */ global_pointer = mmap( NULL, global_len, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_SHARED, file_desc, 0); // Result should be -1 as the request is 5 times actual mapping result = mincore(global_pointer, (size_t)(global_len*5), vect); // Print some data printf("PAGESIZE=%d\n", PAGESIZE); printf("global_len=%d\n", global_len); printf("global_pointer=0x%x\n", (unsigned int)global_pointer); printf("alloc=%d\n", (global_len+PAGESIZE-1) / PAGESIZE ); printf("Result=%d\n", result); printf("vect: "); for ( i=0; i<20; i++) printf("%02x ", vect[i]); printf("\n"); // Clean up munmap(global_pointer, (size_t)global_len); close(file_desc); unlink(file_name); }
Signature · 2 lines
SUSE Labs, Novell Inc. Send instant messages to your online friends http://au.messenger.yahoo.com
Quoted 28 lines · Nick Piggin
Quoted 27 lines · Bruce Dubbs
When testing an installation with tests from the Linux Test Project, my kernels fail one instance of the mincore01 tests:
mincoremincore01 1 PASS : expected failure: errno = 22 (Invalid
argument)mincore01 2 PASS : expected failure: errno = 14 (Bad address) mincore01 3 FAIL : call succeeded unexpectedly mincore01 4 PASS : expected failure: errno = 12 (Cannot allocate memory)01 1 PASS : expected failure: errno = 22 (Invalid argument) mincore01 2 PASS : expected failure: errno = 14 (Bad address) mincore01 3 FAIL : call succeeded unexpectedly mincore01 4 PASS : expected failure: errno = 12 (Cannot allocatememory)
I pared down the test to the attached program. The result is supposed to fail as it is asking for memory information 5 times what should be allocated.
Upon experimenting, I found the test works properly if a printf is executed before the mmap call. I have tested on locally built, but unmodified, 2.4.25, 2.6.12.5, and a 2.6.20.3 kernels and get the same behavior. The tests fail on IA32 architecture, but not 64-bit kernels.
The test always works properly on FC6 and RHEL3.I've checked the archives for this issue and could not find anything appropriate.
Quoted 6 lines
Quoted 3 lines
Could this be a potential security issue as memory that is not supposed to be accessible seems to be available to the user? Is it expected behavior?
It shouldn't be a security problem if mincore doesn't actually return the data.
Thanks for the response. It may be interesting to note that adding:
buf = (char*)global_pointer + 2 * global_len;
i = *buf;
after the mincore call does not fail. Changing the 2nd line above to *buf = 1; gives a segmentation fault as you would expect.
As a minimum, it appears the mmap function is allowing read access beyond its allocated address space in some circumstances.
Upon thinking about it, it may be that the test is invalid. I don't believe there is anything tying the mincore query to the memory region allocated by mmap. If memory mapping occurs beyond the mmap requested memory size to anticipate another memory request, mincore wouldn't fail.
Does this make any sense?
Quoted 58 lines
Quoted 56 lines
------------------------------------------------------------------------
#include <
sys/mman.h>
#include <stdlib.h>
#include <string.h>
#include <stdio.h>static int PAGESIZE;
static char file_name[] = "fooXXXXXX";
static void*global_pointer= NULL;static int global_len = 0; static int file_desc = 0;int
main(int argc, char **argv){ int i; int result; char* buf; unsigned char vect[20] = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0}; PAGESIZE = getpagesize(); /* global_pointer will point to a mmapped area of global_len bytes */ global_len = PAGESIZE*2; buf = (char*)malloc(global_len); memset(buf, 42, global_len); // Asterisks /* create a temporary file */ file_desc = mkstemp(file_name); /* fill the temporary file with two pages of data */ write(file_desc, buf, global_len); free(buf); // Will work properly as long as print is before mmap function. if ( argc > 1 ) printf("argc=%d\n", argc); /* map the file in memory */ global_pointer = mmap( NULL, global_len, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_SHARED, file_desc, 0); // Result should be -1 as the request is 5 times actual mapping result = mincore(global_pointer, (size_t)(global_len*5), vect); // Print some data printf("PAGESIZE=%d\n", PAGESIZE); printf("global_len=%d\n", global_len); printf("global_pointer=0x%x\n", (unsigned int)global_pointer); printf("alloc=%d\n", (global_len+PAGESIZE-1) / PAGESIZE ); printf("Result=%d\n", result); printf("vect: "); for ( i=0; i<20; i++) printf("%02x ", vect[i]); printf("\n"); // Clean up munmap(global_pointer, (size_t)global_len); close(file_desc); unlink(file_name); }