[RFC] arm64: Enforce gettimeofday vdso structure read ordering
From: Brent DeGraaf
Date: Fri Aug 19 2016 - 16:03:21 EST
Introduce explicit control-flow logic immediately prior to virtual
counter register read in all cases so that the mrs read will
always be accessed after all vdso data elements are read and
sequence count is verified. Ensure data elements under the
protection provided by the sequence counter are read only within
the protection logic loop. Read the virtual counter as soon as
possible after the data elements are confirmed correctly read,
rather than after several other operations which can affect time.
Reduce full barriers required in register read code in favor of
the lighter-weight one-way barrier supplied by a load-acquire
wherever possible.
Signed-off-by: Brent DeGraaf <bdegraaf@xxxxxxxxxxxxxx>
---
arch/arm64/include/asm/vdso_datapage.h | 4 +-
arch/arm64/kernel/vdso/gettimeofday.S | 107 +++++++++++++++------------------
2 files changed, 50 insertions(+), 61 deletions(-)
diff --git a/arch/arm64/include/asm/vdso_datapage.h b/arch/arm64/include/asm/vdso_datapage.h
index 2b9a637..49a0a51 100644
--- a/arch/arm64/include/asm/vdso_datapage.h
+++ b/arch/arm64/include/asm/vdso_datapage.h
@@ -21,6 +21,8 @@
#ifndef __ASSEMBLY__
struct vdso_data {
+ __u32 tb_seq_count; /* Timebase sequence counter */
+ __u32 use_syscall;
__u64 cs_cycle_last; /* Timebase at clocksource init */
__u64 raw_time_sec; /* Raw time */
__u64 raw_time_nsec;
@@ -30,14 +32,12 @@ struct vdso_data {
__u64 xtime_coarse_nsec;
__u64 wtm_clock_sec; /* Wall to monotonic time */
__u64 wtm_clock_nsec;
- __u32 tb_seq_count; /* Timebase sequence counter */
/* cs_* members must be adjacent and in this order (ldp accesses) */
__u32 cs_mono_mult; /* NTP-adjusted clocksource multiplier */
__u32 cs_shift; /* Clocksource shift (mono = raw) */
__u32 cs_raw_mult; /* Raw clocksource multiplier */
__u32 tz_minuteswest; /* Whacky timezone stuff */
__u32 tz_dsttime;
- __u32 use_syscall;
};
#endif /* !__ASSEMBLY__ */
diff --git a/arch/arm64/kernel/vdso/gettimeofday.S b/arch/arm64/kernel/vdso/gettimeofday.S
index e00b467..131ac6b 100644
--- a/arch/arm64/kernel/vdso/gettimeofday.S
+++ b/arch/arm64/kernel/vdso/gettimeofday.S
@@ -25,6 +25,10 @@
#define NSEC_PER_SEC_LO16 0xca00
#define NSEC_PER_SEC_HI16 0x3b9a
+#if VDSO_TB_SEQ_COUNT
+#error tb_seq_count MUST be first element of vdso_data
+#endif
+
vdso_data .req x6
seqcnt .req w7
w_tmp .req w8
@@ -36,22 +40,24 @@ x_tmp .req x8
* - All other arguments are read-only, unless otherwise specified.
*/
- .macro seqcnt_acquire
-9999: ldr seqcnt, [vdso_data, #VDSO_TB_SEQ_COUNT]
- tbnz seqcnt, #0, 9999b
- dmb ishld
- .endm
-
- .macro seqcnt_check fail
- dmb ishld
- ldr w_tmp, [vdso_data, #VDSO_TB_SEQ_COUNT]
- cmp w_tmp, seqcnt
- b.ne \fail
- .endm
-
- .macro syscall_check fail
+ .macro seqdata_acquire fallback, tzonly=NO_TZ, skipvcnt=0, getdata
+9999: ldar seqcnt, [vdso_data, #VDSO_TB_SEQ_COUNT]
+8888: tbnz seqcnt, #0, 9999b
ldr w_tmp, [vdso_data, #VDSO_USE_SYSCALL]
- cbnz w_tmp, \fail
+ cbnz w_tmp, \fallback
+ \getdata
+ dmb ishld /* No loads from vdso_data after this point */
+ mov w9, seqcnt
+ ldar seqcnt, [vdso_data, #VDSO_TB_SEQ_COUNT]
+ cmp w9, seqcnt
+ bne 8888b /* Do not needlessly repeat ldar and its implicit barrier */
+ .if (\tzonly) != NO_TZ
+ cbz x0, \tzonly
+ .endif
+ .if (\skipvcnt) == 0
+ isb
+ mrs x_tmp, cntvct_el0
+ .endif
.endm
.macro get_nsec_per_sec res
@@ -64,9 +70,6 @@ x_tmp .req x8
* shift.
*/
.macro get_clock_shifted_nsec res, cycle_last, mult
- /* Read the virtual counter. */
- isb
- mrs x_tmp, cntvct_el0
/* Calculate cycle delta and convert to ns. */
sub \res, x_tmp, \cycle_last
/* We can only guarantee 56 bits of precision. */
@@ -137,17 +140,12 @@ x_tmp .req x8
ENTRY(__kernel_gettimeofday)
.cfi_startproc
adr vdso_data, _vdso_data
- /* If tv is NULL, skip to the timezone code. */
- cbz x0, 2f
-
- /* Compute the time of day. */
-1: seqcnt_acquire
- syscall_check fail=4f
- ldr x10, [vdso_data, #VDSO_CS_CYCLE_LAST]
- /* w11 = cs_mono_mult, w12 = cs_shift */
- ldp w11, w12, [vdso_data, #VDSO_CS_MONO_MULT]
- ldp x13, x14, [vdso_data, #VDSO_XTIME_CLK_SEC]
- seqcnt_check fail=1b
+ seqdata_acquire fallback=4f tzonly=2f getdata=__stringify(\
+ ldr x10, [vdso_data, #VDSO_CS_CYCLE_LAST];\
+ /* w11 = cs_mono_mult, w12 = cs_shift */;\
+ ldp w11, w12, [vdso_data, #VDSO_CS_MONO_MULT];\
+ ldp x13, x14, [vdso_data, #VDSO_XTIME_CLK_SEC];\
+ ldp w4, w5, [vdso_data, #VDSO_TZ_MINWEST])
get_nsec_per_sec res=x9
lsl x9, x9, x12
@@ -164,7 +162,6 @@ ENTRY(__kernel_gettimeofday)
2:
/* If tz is NULL, return 0. */
cbz x1, 3f
- ldp w4, w5, [vdso_data, #VDSO_TZ_MINWEST]
stp w4, w5, [x1, #TZ_MINWEST]
3:
mov x0, xzr
@@ -205,13 +202,11 @@ jumptable:
ALIGN
realtime:
- seqcnt_acquire
- syscall_check fail=syscall
- ldr x10, [vdso_data, #VDSO_CS_CYCLE_LAST]
- /* w11 = cs_mono_mult, w12 = cs_shift */
- ldp w11, w12, [vdso_data, #VDSO_CS_MONO_MULT]
- ldp x13, x14, [vdso_data, #VDSO_XTIME_CLK_SEC]
- seqcnt_check fail=realtime
+ seqdata_acquire fallback=syscall getdata=__stringify(\
+ ldr x10, [vdso_data, #VDSO_CS_CYCLE_LAST];\
+ /* w11 = cs_mono_mult, w12 = cs_shift */;\
+ ldp w11, w12, [vdso_data, #VDSO_CS_MONO_MULT];\
+ ldp x13, x14, [vdso_data, #VDSO_XTIME_CLK_SEC])
/* All computations are done with left-shifted nsecs. */
get_nsec_per_sec res=x9
@@ -224,14 +219,12 @@ realtime:
ALIGN
monotonic:
- seqcnt_acquire
- syscall_check fail=syscall
- ldr x10, [vdso_data, #VDSO_CS_CYCLE_LAST]
- /* w11 = cs_mono_mult, w12 = cs_shift */
- ldp w11, w12, [vdso_data, #VDSO_CS_MONO_MULT]
- ldp x13, x14, [vdso_data, #VDSO_XTIME_CLK_SEC]
- ldp x3, x4, [vdso_data, #VDSO_WTM_CLK_SEC]
- seqcnt_check fail=monotonic
+ seqdata_acquire fallback=syscall getdata=__stringify(\
+ ldr x10, [vdso_data, #VDSO_CS_CYCLE_LAST];\
+ /* w11 = cs_mono_mult, w12 = cs_shift */;\
+ ldp w11, w12, [vdso_data, #VDSO_CS_MONO_MULT];\
+ ldp x13, x14, [vdso_data, #VDSO_XTIME_CLK_SEC];\
+ ldp x3, x4, [vdso_data, #VDSO_WTM_CLK_SEC])
/* All computations are done with left-shifted nsecs. */
lsl x4, x4, x12
@@ -247,13 +240,11 @@ monotonic:
ALIGN
monotonic_raw:
- seqcnt_acquire
- syscall_check fail=syscall
- ldr x10, [vdso_data, #VDSO_CS_CYCLE_LAST]
- /* w11 = cs_raw_mult, w12 = cs_shift */
- ldp w12, w11, [vdso_data, #VDSO_CS_SHIFT]
- ldp x13, x14, [vdso_data, #VDSO_RAW_TIME_SEC]
- seqcnt_check fail=monotonic_raw
+ seqdata_acquire fallback=syscall getdata=__stringify(\
+ ldr x10, [vdso_data, #VDSO_CS_CYCLE_LAST];\
+ /* w11 = cs_raw_mult, w12 = cs_shift */;\
+ ldp w12, w11, [vdso_data, #VDSO_CS_SHIFT];\
+ ldp x13, x14, [vdso_data, #VDSO_RAW_TIME_SEC])
/* All computations are done with left-shifted nsecs. */
lsl x14, x14, x12
@@ -269,17 +260,15 @@ monotonic_raw:
ALIGN
realtime_coarse:
- seqcnt_acquire
- ldp x10, x11, [vdso_data, #VDSO_XTIME_CRS_SEC]
- seqcnt_check fail=realtime_coarse
+ seqdata_acquire fallback=syscall skipvcnt=1 getdata=__stringify(\
+ ldp x10, x11, [vdso_data, #VDSO_XTIME_CRS_SEC])
clock_gettime_return
ALIGN
monotonic_coarse:
- seqcnt_acquire
- ldp x10, x11, [vdso_data, #VDSO_XTIME_CRS_SEC]
- ldp x13, x14, [vdso_data, #VDSO_WTM_CLK_SEC]
- seqcnt_check fail=monotonic_coarse
+ seqdata_acquire fallback=syscall skipvcnt=1 getdata=__stringify(\
+ ldp x10, x11, [vdso_data, #VDSO_XTIME_CRS_SEC];\
+ ldp x13, x14, [vdso_data, #VDSO_WTM_CLK_SEC])
/* Computations are done in (non-shifted) nsecs. */
get_nsec_per_sec res=x9
--
Qualcomm Datacenter Technologies, Inc. as an affiliate of Qualcomm
Technologies, Inc. Qualcomm Technologies, Inc. is a member of the Code Aurora
Forum, a Linux Foundation Collaborative Project.