ARM: gcc5 causes undefined behaviour in mmu_init()
Enrico Scholz
enrico.scholz at sigma-chemnitz.de
Wed Sep 16 10:58:38 PDT 2015
Hi,
a barebox built with gcc5 hangs in tlb_invalidate() very early. This
seems to be caused by an integer overflow:
--- arch/arm/cpu/mmu.c:
static int arm_mmu_remap_sdram(struct memory_bank *bank)
{
unsigned long num_ptes = bank->size >> 10;
int i;
for (i = 0; i < num_ptes; i++) {
ptes[i] = (phys + i * PAGE_SIZE) | PTE_TYPE_SMALL |
pte_flags_cached;
}
For 1GiB RAM, 'num_ptes' is 1MiB and due to integer promotion, the 'i *
PAGE_SIZE' overflows and causes undefined behavior.
A trivial fix is to make 'i' an unsigned int or long.
But I wonder whether calculation of 'num_ptes' is really correct. Does
barebox really use a 1KiB pagesize for PTEs or should the '>> 10' be a
'>> 12'? When it is really 1KiB, the mapping for memory sizes > 1GiB
seem to be ambiguous then.
For reference; in broken case ('int i'), gcc5 generates:
| 4badc: f7b4 fe7e bl 7dc <pr_print>
| 4bae0: f8d9 3000 ldr.w r3, [r9]
| 4bae4: f1aa 0104 sub.w r1, sl, #4
| 4bae8: f043 0202 orr.w r2, r3, #2
| 4baec: 9b03 ldr r3, [sp, #12]
| 4baee: eb04 3303 add.w r3, r4, r3, lsl #12
|
| 4baf2: 42a3 cmp r3, r4
| 4baf4: d006 beq.n 4bb04 <mmu_init+0x1c0>
| 4baf6: ea42 0004 orr.w r0, r2, r4
| 4bafa: f504 5480 add.w r4, r4, #4096 ; 0x1000
| 4bafe: f841 0f04 str.w r0, [r1, #4]!
| 4bb02: e7f6 b.n 4baf2 <mmu_init+0x1ae>
Building with 'unsigned long i' generates:
| 4bad8: f7b4 fe80 bl 7dc <pr_print>
| 4badc: 683b ldr r3, [r7, #0]
| 4bade: f043 0202 orr.w r2, r3, #2
| 4bae2: 9b05 ldr r3, [sp, #20]
|
| 4bae4: 9904 ldr r1, [sp, #16]
| 4bae6: 4299 cmp r1, r3
| 4bae8: d006 beq.n 4baf8 <mmu_init+0x1b4>
| 4baea: eb0a 3103 add.w r1, sl, r3, lsl #12
| 4baee: 4311 orrs r1, r2
| 4baf0: f849 1023 str.w r1, [r9, r3, lsl #2]
| 4baf4: 3301 adds r3, #1
| 4baf6: e7f5 b.n 4bae4 <mmu_init+0x1a0>
Enrico
--
SIGMA Chemnitz GmbH Registergericht: Amtsgericht Chemnitz HRB 1750
Am Erlenwald 13 Geschaeftsfuehrer: Grit Freitag, Frank Pyritz
09128 Chemnitz
More information about the barebox
mailing list