On 07/12/15 11:08, Will Deacon wrote:
[...]
+#define ARM_LPAE_GRANULE(d) (1UL << (d)->pg_shift)
+
#define ARM_LPAE_PAGES_PER_PGD(d) \
- DIV_ROUND_UP((d)->pgd_size, 1UL << (d)->pg_shift)
+ DIV_ROUND_UP((d)->pgd_size, ARM_LPAE_GRANULE(d))
/*
* Calculate the index at level l used to map virtual address a using the
@@ -169,7 +171,7 @@
/* IOPTE accessors */
#define iopte_deref(pte,d) \
(__va((pte) & ((1ULL << ARM_LPAE_MAX_ADDR_BITS) - 1) \
- & ~((1ULL << (d)->pg_shift) - 1)))
+ & ~(ARM_LPAE_GRANULE(d) - 1)))
Do we run the risk of truncating the VA on 32-bit ARM here?
Indeed, in all honesty I'd missed that, but since __va is going to
truncate it to a 32-bit void * anyway, doing it earlier in the
expression actually seems to result in better code - with iopte_deref
wrapped like so to make it easier to pick out:
arm_lpae_iopte *__iopte_deref(arm_lpae_iopte pte, struct
arm_lpae_io_pgtable *d) {
return iopte_deref(pte, d);
}
the result goes from this:
1568: e92d4070 push {r4, r5, r6, lr}
156c: e3a0e001 mov lr, #1
1570: e592c060 ldr ip, [r2, #96] ; 0x60
1574: e3e04000 mvn r4, #0
1578: e30f5fff movw r5, #65535 ; 0xffff
157c: e26c6020 rsb r6, ip, #32
1580: e1a02c1e lsl r2, lr, ip
1584: e2722000 rsbs r2, r2, #0
1588: e0000002 and r0, r0, r2
158c: e0000004 and r0, r0, r4
1590: e2400481 sub r0, r0, #-2130706432 ; 0x81000000
1594: e8bd8070 pop {r4, r5, r6, pc}
to this:
151c: e5922060 ldr r2, [r2, #96] ; 0x60
1520: e3e03000 mvn r3, #0
1524: e1a03213 lsl r3, r3, r2
1528: e0000003 and r0, r0, r3
152c: e2400481 sub r0, r0, #-2130706432 ; 0x81000000
1530: e12fff1e bx lr
which, given how mind-bogglingly silly some of the former code looks,
seems like an inadvertent win.
Robin.
Will
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
