Re: [PATCH v2 0/3] arm64: perf_event: Fix time offset prior to epoch

[Peter Zijlstra]

On Tue, May 12, 2020 at 11:19:18AM +0200, Peter Zijlstra wrote:

> Now, your arm64 counter is 56 bits, so wrapping is rare, but still, we
> should probably fix that. And that probably needs an ABI extention
> *sigh*.

Something like so, on top of the other (2).

---

--- a/arch/arm64/kernel/perf_event.c
+++ b/arch/arm64/kernel/perf_event.c
@@ -1174,6 +1174,7 @@ void arch_perf_update_userpage(struct pe
 
 	userpg->cap_user_time = 0;
 	userpg->cap_user_time_zero = 0;
+	userpg->cap_user_time_short = 0;
 
 	do {
 		rd = sched_clock_read_begin(&seq);
@@ -1184,13 +1185,16 @@ void arch_perf_update_userpage(struct pe
 		userpg->time_mult = rd->mult;
 		userpg->time_shift = rd->shift;
 		userpg->time_zero = rd->epoch_ns;
+		userpg->time_cycle = rd->epoch_cyc;
+		userpg->time_mask = rd->sched_clock_mask;
 
 		/*
 		 * This isn't strictly correct, the ARM64 counter can be
-		 * 'short' and then we get funnies when it wraps. The correct
-		 * thing would be to extend the perf ABI with a cycle and mask
-		 * value, but because wrapping on ARM64 is very rare in
-		 * practise this 'works'.
+		 * 'short' and then we get funnies when it wraps. But this
+		 * 'works' with the cap_user_time ABI.
+		 *
+		 * When userspace knows about cap_user_time_short, it
+		 * can do the correct thing.
 		 */
 		userpg->time_zero -= (rd->epoch_cyc * rd->mult) >> rd->shift;
 
@@ -1215,4 +1219,5 @@ void arch_perf_update_userpage(struct pe
 	 */
 	userpg->cap_user_time = 1;
 	userpg->cap_user_time_zero = 1;
+	userpg->cap_user_time_short = 1;
 }
--- a/include/uapi/linux/perf_event.h
+++ b/include/uapi/linux/perf_event.h
@@ -532,9 +532,10 @@ struct perf_event_mmap_page {
 				cap_bit0_is_deprecated	: 1, /* Always 1, signals that bit 0 is zero */
 
 				cap_user_rdpmc		: 1, /* The RDPMC instruction can be used to read counts */
-				cap_user_time		: 1, /* The time_* fields are used */
+				cap_user_time		: 1, /* The time_{shift,mult,offset} fields are used */
 				cap_user_time_zero	: 1, /* The time_zero field is used */
-				cap_____res		: 59;
+				cap_user_time_short	: 1, /* the time_{cycle,mask} fields are used */
+				cap_____res		: 58;
 		};
 	};
 
@@ -593,13 +594,29 @@ struct perf_event_mmap_page {
 	 *               ((rem * time_mult) >> time_shift);
 	 */
 	__u64	time_zero;
+
 	__u32	size;			/* Header size up to __reserved[] fields. */
+	__u32	__reserved_1;
+
+	/*
+	 * If cap_usr_time_short, the hardware clock is less than 64bit wide
+	 * and we must compute the 'cyc' value, as used by cap_usr_time, as:
+	 *
+	 *   cyc = time_cycles + ((cyc - time_cycles) & time_mask)
+	 *
+	 * NOTE: this form is explicitly chosen such that cap_usr_time_short
+	 *       is an extention on top of cap_usr_time, and code that doesn't
+	 *       know about cap_usr_time_short still works under the assumption
+	 *       the counter doesn't wrap.
+	 */
+	__u64	time_cycles;
+	__u64	time_mask;
 
 		/*
 		 * Hole for extension of the self monitor capabilities
 		 */
 
-	__u8	__reserved[118*8+4];	/* align to 1k. */
+	__u8	__reserved[116*8];	/* align to 1k. */
 
 	/*
 	 * Control data for the mmap() data buffer.