Provide timespec to guests rather than jiffies clock.

A non-periodic clock_event_device and the "jiffies" clock don't mix well:
tick_handle_periodic() can go into an infinite loop.

Currently lguest guests use the jiffies clock when the TSC is
unusable.  Instead, make the Host write the current time into the lguest
page on every interrupt.  This doesn't cost much but is more precise
and at least as accurate as the jiffies clock.  It also gets rid of
the GET_WALLCLOCK hypercall.

Also, delay setting sched_clock until our clock is set up, otherwise
the early printk timestamps can go backwards (not harmful, just ugly).

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

4 files changed, 57 insertions, 24 deletions

diff --git a/drivers/lguest/hypercalls.c b/drivers/lguest/hypercalls.c
index 7a5299f9679..db6caace3b9 100644
--- a/drivers/lguest/hypercalls.c
+++ b/drivers/lguest/hypercalls.c
@@ -64,14 +64,6 @@ static void do_hcall(struct lguest *lg, struct lguest_regs *regs)
 		else
 			guest_pagetable_flush_user(lg);
 		break;
-	case LHCALL_GET_WALLCLOCK: {
-		/* The Guest wants to know the real time in seconds since 1970,
-		 * in good Unix tradition. */
-		struct timespec ts;
-		ktime_get_real_ts(&ts);
-		regs->eax = ts.tv_sec;
-		break;
-	}
 	case LHCALL_BIND_DMA:
 		/* BIND_DMA really wants four arguments, but it's the only call
 		 * which does.  So the Guest packs the number of buffers and
@@ -235,6 +227,9 @@ static void initialize(struct lguest *lg)
 	    || put_user(lg->guestid, &lg->lguest_data->guestid))
 		kill_guest(lg, "bad guest page %p", lg->lguest_data);
 
+	/* We write the current time into the Guest's data page once now. */
+	write_timestamp(lg);
+
 	/* This is the one case where the above accesses might have been the
 	 * first write to a Guest page.  This may have caused a copy-on-write
 	 * fault, but the Guest might be referring to the old (read-only)
@@ -293,3 +288,13 @@ void do_hypercalls(struct lguest *lg)
 		clear_hcall(lg);
 	}
 }
+
+/* This routine supplies the Guest with time: it's used for wallclock time at
+ * initial boot and as a rough time source if the TSC isn't available. */
+void write_timestamp(struct lguest *lg)
+{
+	struct timespec now;
+	ktime_get_real_ts(&now);
+	if (put_user(now, &lg->lguest_data->time))
+		kill_guest(lg, "Writing timestamp");
+}
diff --git a/drivers/lguest/interrupts_and_traps.c b/drivers/lguest/interrupts_and_traps.c
index bd0091bf79e..49787e964a0 100644
--- a/drivers/lguest/interrupts_and_traps.c
+++ b/drivers/lguest/interrupts_and_traps.c
@@ -175,6 +175,13 @@ void maybe_do_interrupt(struct lguest *lg)
 		 * the stack as well: virtual interrupts never do. */
 		set_guest_interrupt(lg, idt->a, idt->b, 0);
 	}
+
+	/* Every time we deliver an interrupt, we update the timestamp in the
+	 * Guest's lguest_data struct.  It would be better for the Guest if we
+	 * did this more often, but it can actually be quite slow: doing it
+	 * here is a compromise which means at least it gets updated every
+	 * timer interrupt. */
+	write_timestamp(lg);
 }
 
 /*H:220 Now we've got the routines to deliver interrupts, delivering traps
diff --git a/drivers/lguest/lg.h b/drivers/lguest/lg.h
index 269116eee85..64f0abed317 100644
--- a/drivers/lguest/lg.h
+++ b/drivers/lguest/lg.h
@@ -256,6 +256,7 @@ unsigned long get_dma_buffer(struct lguest *lg, unsigned long key,
 
 /* hypercalls.c: */
 void do_hypercalls(struct lguest *lg);
+void write_timestamp(struct lguest *lg);
 
 /*L:035
  * Let's step aside for the moment, to study one important routine that's used
diff --git a/drivers/lguest/lguest.c b/drivers/lguest/lguest.c
index 3386b0e7690..1bc1546c7fd 100644
--- a/drivers/lguest/lguest.c
+++ b/drivers/lguest/lguest.c
@@ -643,21 +643,42 @@ static void __init lguest_init_IRQ(void)
  * Time.
  *
  * It would be far better for everyone if the Guest had its own clock, but
- * until then it must ask the Host for the time.
+ * until then the Host gives us the time on every interrupt.
  */
 static unsigned long lguest_get_wallclock(void)
 {
-	return hcall(LHCALL_GET_WALLCLOCK, 0, 0, 0);
+	return lguest_data.time.tv_sec;
 }
 
-/* If the Host tells us we can trust the TSC, we use that, otherwise we simply
- * use the imprecise but reliable "jiffies" counter. */
 static cycle_t lguest_clock_read(void)
 {
+	unsigned long sec, nsec;
+
+	/* If the Host tells the TSC speed, we can trust that. */
 	if (lguest_data.tsc_khz)
 		return native_read_tsc();
-	else
-		return jiffies;
+
+	/* If we can't use the TSC, we read the time value written by the Host.
+	 * Since it's in two parts (seconds and nanoseconds), we risk reading
+	 * it just as it's changing from 99 & 0.999999999 to 100 and 0, and
+	 * getting 99 and 0.  As Linux tends to come apart under the stress of
+	 * time travel, we must be careful: */
+	do {
+		/* First we read the seconds part. */
+		sec = lguest_data.time.tv_sec;
+		/* This read memory barrier tells the compiler and the CPU that
+		 * this can't be reordered: we have to complete the above
+		 * before going on. */
+		rmb();
+		/* Now we read the nanoseconds part. */
+		nsec = lguest_data.time.tv_nsec;
+		/* Make sure we've done that. */
+		rmb();
+		/* Now if the seconds part has changed, try again. */
+	} while (unlikely(lguest_data.time.tv_sec != sec));
+
+	/* Our non-TSC clock is in real nanoseconds. */
+	return sec*1000000000ULL + nsec;
 }
 
 /* This is what we tell the kernel is our clocksource.  */
@@ -665,8 +686,11 @@ static struct clocksource lguest_clock = {
 	.name		= "lguest",
 	.rating		= 400,
 	.read		= lguest_clock_read,
+	.mask		= CLOCKSOURCE_MASK(64),
+	.mult		= 1,
 };
 
+/* The "scheduler clock" is just our real clock, adjusted to start at zero */
 static unsigned long long lguest_sched_clock(void)
 {
 	return cyc2ns(&lguest_clock, lguest_clock_read() - clock_base);
@@ -742,24 +766,21 @@ static void lguest_time_init(void)
 	set_irq_handler(0, lguest_time_irq);
 
 	/* Our clock structure look like arch/i386/kernel/tsc.c if we can use
-	 * the TSC, otherwise it looks like kernel/time/jiffies.c.  Either way,
-	 * the "rating" is initialized so high that it's always chosen over any
-	 * other clocksource. */
+	 * the TSC, otherwise it's a dumb nanosecond-resolution clock.  Either
+	 * way, the "rating" is initialized so high that it's always chosen
+	 * over any other clocksource. */
 	if (lguest_data.tsc_khz) {
 		lguest_clock.shift = 22;
 		lguest_clock.mult = clocksource_khz2mult(lguest_data.tsc_khz,
 							 lguest_clock.shift);
-		lguest_clock.mask = CLOCKSOURCE_MASK(64);
 		lguest_clock.flags = CLOCK_SOURCE_IS_CONTINUOUS;
-	} else {
-		/* To understand this, start at kernel/time/jiffies.c... */
-		lguest_clock.shift = 8;
-		lguest_clock.mult = (((u64)NSEC_PER_SEC<<8)/ACTHZ) << 8;
-		lguest_clock.mask = CLOCKSOURCE_MASK(32);
 	}
 	clock_base = lguest_clock_read();
 	clocksource_register(&lguest_clock);
 
+	/* Now we've set up our clock, we can use it as the scheduler clock */
+	paravirt_ops.sched_clock = lguest_sched_clock;
+
 	/* We can't set cpumask in the initializer: damn C limitations!  Set it
 	 * here and register our timer device. */
 	lguest_clockevent.cpumask = cpumask_of_cpu(0);
@@ -996,7 +1017,6 @@ __init void lguest_init(void *boot)
 	paravirt_ops.time_init = lguest_time_init;
 	paravirt_ops.set_lazy_mode = lguest_lazy_mode;
 	paravirt_ops.wbinvd = lguest_wbinvd;
-	paravirt_ops.sched_clock = lguest_sched_clock;
 	/* Now is a good time to look at the implementations of these functions
 	 * before returning to the rest of lguest_init(). */