]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/commitdiff
arm64: arch_timer: Ensure counter register reads occur with seqlock held
authorWill Deacon <will.deacon@arm.com>
Mon, 29 Apr 2019 16:26:22 +0000 (17:26 +0100)
committerKleber Sacilotto de Souza <kleber.souza@canonical.com>
Wed, 14 Aug 2019 09:18:49 +0000 (11:18 +0200)
BugLink: https://bugs.launchpad.net/bugs/1838576
commit 75a19a0202db21638a1c2b424afb867e1f9a2376 upstream.

When executing clock_gettime(), either in the vDSO or via a system call,
we need to ensure that the read of the counter register occurs within
the seqlock reader critical section. This ensures that updates to the
clocksource parameters (e.g. the multiplier) are consistent with the
counter value and therefore avoids the situation where time appears to
go backwards across multiple reads.

Extend the vDSO logic so that the seqlock critical section covers the
read of the counter register as well as accesses to the data page. Since
reads of the counter system registers are not ordered by memory barrier
instructions, introduce dependency ordering from the counter read to a
subsequent memory access so that the seqlock memory barriers apply to
the counter access in both the vDSO and the system call paths.

Cc: <stable@vger.kernel.org>
Cc: Marc Zyngier <marc.zyngier@arm.com>
Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com>
Link: https://lore.kernel.org/linux-arm-kernel/alpine.DEB.2.21.1902081950260.1662@nanos.tec.linutronix.de/
Reported-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Kamal Mostafa <kamal@canonical.com>
Signed-off-by: Khalid Elmously <khalid.elmously@canonical.com>
arch/arm64/include/asm/arch_timer.h
arch/arm64/kernel/vdso/gettimeofday.S

index f2a234d6516cf5b537b80134b137729dd9abb798..93e07512b4b61cad3c9fc1832a2c973ab0b6b1ad 100644 (file)
@@ -148,18 +148,47 @@ static inline void arch_timer_set_cntkctl(u32 cntkctl)
        isb();
 }
 
+/*
+ * Ensure that reads of the counter are treated the same as memory reads
+ * for the purposes of ordering by subsequent memory barriers.
+ *
+ * This insanity brought to you by speculative system register reads,
+ * out-of-order memory accesses, sequence locks and Thomas Gleixner.
+ *
+ * http://lists.infradead.org/pipermail/linux-arm-kernel/2019-February/631195.html
+ */
+#define arch_counter_enforce_ordering(val) do {                                \
+       u64 tmp, _val = (val);                                          \
+                                                                       \
+       asm volatile(                                                   \
+       "       eor     %0, %1, %1\n"                                   \
+       "       add     %0, sp, %0\n"                                   \
+       "       ldr     xzr, [%0]"                                      \
+       : "=r" (tmp) : "r" (_val));                                     \
+} while (0)
+
 static inline u64 arch_counter_get_cntpct(void)
 {
+       u64 cnt;
+
        isb();
-       return arch_timer_reg_read_stable(cntpct_el0);
+       cnt = arch_timer_reg_read_stable(cntpct_el0);
+       arch_counter_enforce_ordering(cnt);
+       return cnt;
 }
 
 static inline u64 arch_counter_get_cntvct(void)
 {
+       u64 cnt;
+
        isb();
-       return arch_timer_reg_read_stable(cntvct_el0);
+       cnt = arch_timer_reg_read_stable(cntvct_el0);
+       arch_counter_enforce_ordering(cnt);
+       return cnt;
 }
 
+#undef arch_counter_enforce_ordering
+
 static inline int arch_timer_arch_init(void)
 {
        return 0;
index c39872a7b03c3e152315781c753f3e6b186524ed..e8f60112818fcdb57af23ad02c7cccb901446aea 100644 (file)
@@ -73,6 +73,13 @@ x_tmp                .req    x8
        movn    x_tmp, #0xff00, lsl #48
        and     \res, x_tmp, \res
        mul     \res, \res, \mult
+       /*
+        * Fake address dependency from the value computed from the counter
+        * register to subsequent data page accesses so that the sequence
+        * locking also orders the read of the counter.
+        */
+       and     x_tmp, \res, xzr
+       add     vdso_data, vdso_data, x_tmp
        .endm
 
        /*
@@ -147,12 +154,12 @@ ENTRY(__kernel_gettimeofday)
        /* 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
 
        get_nsec_per_sec res=x9
        lsl     x9, x9, x12
 
        get_clock_shifted_nsec res=x15, cycle_last=x10, mult=x11
+       seqcnt_check fail=1b
        get_ts_realtime res_sec=x10, res_nsec=x11, \
                clock_nsec=x15, xtime_sec=x13, xtime_nsec=x14, nsec_to_sec=x9
 
@@ -211,13 +218,13 @@ realtime:
        /* 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
 
        /* All computations are done with left-shifted nsecs. */
        get_nsec_per_sec res=x9
        lsl     x9, x9, x12
 
        get_clock_shifted_nsec res=x15, cycle_last=x10, mult=x11
+       seqcnt_check fail=realtime
        get_ts_realtime res_sec=x10, res_nsec=x11, \
                clock_nsec=x15, xtime_sec=x13, xtime_nsec=x14, nsec_to_sec=x9
        clock_gettime_return, shift=1
@@ -231,7 +238,6 @@ monotonic:
        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
 
        /* All computations are done with left-shifted nsecs. */
        lsl     x4, x4, x12
@@ -239,6 +245,7 @@ monotonic:
        lsl     x9, x9, x12
 
        get_clock_shifted_nsec res=x15, cycle_last=x10, mult=x11
+       seqcnt_check fail=monotonic
        get_ts_realtime res_sec=x10, res_nsec=x11, \
                clock_nsec=x15, xtime_sec=x13, xtime_nsec=x14, nsec_to_sec=x9
 
@@ -253,13 +260,13 @@ monotonic_raw:
        /* 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
 
        /* All computations are done with left-shifted nsecs. */
        get_nsec_per_sec res=x9
        lsl     x9, x9, x12
 
        get_clock_shifted_nsec res=x15, cycle_last=x10, mult=x11
+       seqcnt_check fail=monotonic_raw
        get_ts_clock_raw res_sec=x10, res_nsec=x11, \
                clock_nsec=x15, nsec_to_sec=x9