Make the loader runs on 16k page size kernel with SCons (2)

[illwieckz]

Replay of:

• Make the loader runs on 16k page size kernel with SCons #11

[slipher]

I think it is breaking the bootstrap loader memory mapping somehow based on nacl_bootstrap_prereservation_test. I'm testing just on a 4k page system. The test also doesn't work on master because mmap is buggy and refuses to allocate the page starting right at vm.mmap_min_addr but the test can be made to work by modifying IsPageMappable to accept EINVAL with any address. So with that change, the mappings from master are like

00002000-00010000 ---p 00000000 00:00 0 
00010000-00012000 r-xp 00000000 b3:02 137384     /home/me/elsewhere/scons-out/opt-linux-arm/staging/nacl_helper_bootstrap
00012000-00014000 rw-p 00000000 00:00 0 
00014000-40002000 ---p 00000000 b3:02 137384     /home/me/elsewhere/scons-out/opt-linux-arm/staging/nacl_helper_bootstrap
40002000-40003000 rw-p 00002000 b3:02 137384     /home/me/elsewhere/scons-out/opt-linux-arm/staging/nacl_helper_bootstrap
4112a000-4114b000 rw-p 00000000 00:00 0          [heap]

The test, which checks that the first 1 GiB is reserved, passes. With this branch, the patched test fails and shows there was no additional memory reserved:

00002000-00010000 ---p 00000000 00:00 0 
00010000-00012000 r-xp 00000000 b3:02 146288     /home/me/elsewhere/scons-out/opt-linux-arm/staging/nacl_helper_bootstrap
00020000-00022000 rw-p 00000000 00:00 0 

The patch to the test:

diff --git a/src/trusted/service_runtime/linux/nacl_bootstrap_prereservation_test.c b/src/trusted/service_runtime/linux/nacl_bootstrap_prereservation_test.c
index 9f0a888e2..10812bcfa 100644
--- a/src/trusted/service_runtime/linux/nacl_bootstrap_prereservation_test.c
+++ b/src/trusted/service_runtime/linux/nacl_bootstrap_prereservation_test.c
@@ -28,7 +28,7 @@ int ParseMapsLine(FILE *fp, uintptr_t *start, uintptr_t *end) {
   if (!fgets(buf, sizeof(buf), fp)) {
     return 0;
   }
-
+ printf("%s", buf);
   if (sscanf(buf, "%"SCNxPTR"-%"SCNxPTR" %*4[-rwxp] %*8x %*2x:%*2x %*16x",
              start, end) != 2) {
     return 0;
@@ -62,9 +62,10 @@ int IsPageMappable(uintptr_t addr) {
                       MAP_ANONYMOUS | MAP_NORESERVE,
                       -1, 0);
   if (retval != (void *) addr)   {
+printf("addr=%p retval=%p errno=%d\n", (void*)addr, retval, errno);
     CHECK(MAP_FAILED == retval &&
           (EPERM == errno || EACCES == errno ||
-           (EINVAL == errno && 0 == addr)));
+           (EINVAL == errno /*&& 0 == addr*/)));
     return 0;
   }
   return 1;

[slipher]

#36 has a proper fix for the aforementioned test.

I haven't investigated deeply yet why it fails with this PR, but probably we need to change the code to look up the page size at runtime for the mapping to work correctly.

Is there a strong reason to increase the supported page size to 64k rather than just 16k? That may require additional changes due to the ARM kernel's rule that the first two pages can't be mapped. We would also have to increase the loader's base address to at least 128k for it to be loadable. Though we were thinking of increasing it anyway due to the Slackware vm.min_mmap_addr issue.

[illwieckz]

Is there a strong reason to increase the supported page size to 64k rather than just 16k?

No more reason other than “the more can do the less”. I discovered that all other binaries were aligned on 64K, and since 64k can run on 16K that can run on 4k, I just thought it would be good practice to standardize the alignment.

Now if 64K is harder to achieve than 16k, we can do 16k.