Skip to contentSkip to conversation

lib/vdso: Improve do_hres() and update vdso data unconditionally

PATCH In review Huacai Chen 2 participants 5 messages Oct 21, 2019 – Oct 23, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 110/110 lib/vdso: Improve do_hres() and update vdso data unconditionally

In do_hres(), we currently use whether the return value of __arch_get_ hw_counter() is negative to indicate fallback, but this is not a good idea because:

1, ARM64 returns ULL_MAX but MIPS returns 0 when clock_mode is invalid; 2, For a 64bit counter, a "negative" value of counter is actually valid.

It is sure that MIPS has a bug when clock_mode is invalid and should return ULL_MAX as ARM64 does (Vincenzo has already submitted a patch). But do_hres() can still be improved so we use U64_MAX as the only "invalid" return value -- this is still not fully correct, but it is the simplest fix and has no problem in most cases (we can hardly see a 64bit counter overflow).

By the way, currently update_vdso_data() and update_vsyscall_tz() rely on __arch_use_vsyscall(), which causes __cvdso_clock_getres() and some other functions get wrong results when clock_mode is invalid. So, we update vdso data unconditionally.

Fixes 00b26474c2f1613d7ab894c5 ("lib/vdso: Provide generic VDSO implementation")
Fixes 44f57d788e7deecb50484353 ("timekeeping: Provide a generic update_vsyscall() implementation")
Cc Arnd Bergmann <[email protected]>
Cc Paul Burton <[email protected]>
---
 kernel/time/vsyscall.c  | 9 +++------
 lib/vdso/gettimeofday.c | 2 +-
 2 files changed, 4 insertions(+), 7 deletions(-)
kernel/time/vsyscall.c +3 -6
index 4bc37ac..5ee0f77 100644
@@ -110,8 +110,7 @@ void update_vsyscall(struct timekeeper *tk)
 	nsec		= nsec + tk->wall_to_monotonic.tv_nsec;
 	vdso_ts->sec	+= __iter_div_u64_rem(nsec, NSEC_PER_SEC, &vdso_ts->nsec);
 
-	if (__arch_use_vsyscall(vdata))
-		update_vdso_data(vdata, tk);
+	update_vdso_data(vdata, tk);
 
 	__arch_update_vsyscall(vdata, tk);
 
@@ -124,10 +123,8 @@ void update_vsyscall_tz(void)
 {
 	struct vdso_data *vdata = __arch_get_k_vdso_data();
 
-	if (__arch_use_vsyscall(vdata)) {
-		vdata[CS_HRES_COARSE].tz_minuteswest = sys_tz.tz_minuteswest;
-		vdata[CS_HRES_COARSE].tz_dsttime = sys_tz.tz_dsttime;
-	}
+	vdata[CS_HRES_COARSE].tz_minuteswest = sys_tz.tz_minuteswest;
+	vdata[CS_HRES_COARSE].tz_dsttime = sys_tz.tz_dsttime;
 
 	__arch_sync_vdso_data(vdata);
 }
lib/vdso/gettimeofday.c +1 -1
index e630e7f..5a31643 100644
@@ -50,7 +50,7 @@ static int do_hres(const struct vdso_data *vd, clockid_t clk,
 		cycles = __arch_get_hw_counter(vd->clock_mode);
 		ns = vdso_ts->nsec;
 		last = vd->cycle_last;
-		if (unlikely((s64)cycles < 0))
+		if (unlikely(cycles == U64_MAX))
 			return -1;
 
 		ns += vdso_calc_delta(cycles, last, vd->mask, vd->mult);
Signature · 1 line
2.7.0
#1 lore raw

lib/vdso/gettimeofday.c static int do_hres(const struct vdso_data *vd, clockid_t clk, · line 54

Quoted 7 lines · Huacai Chen
lib/vdso/gettimeofday.c +1 -1
@@ -50,7 +50,7 @@ static int do_hres(const struct vdso_data *vd, clockid_t clk,
 		cycles = __arch_get_hw_counter(vd->clock_mode);
 		ns = vdso_ts->nsec;
 		last = vd->cycle_last;
-		if (unlikely((s64)cycles < 0))
+		if (unlikely(cycles == U64_MAX))
 			return -1;

That used to create worse code than the weird (s64) type cast which has the same effect. Did you double check that there is no change?

Thanks,

tglx

#2 lore raw
Quoted 11 lines · Thomas Gleixner

lib/vdso/gettimeofday.c static int do_hres(const struct vdso_data *vd, clockid_t clk, · line 54

Quoted 7 lines · Huacai Chen
lib/vdso/gettimeofday.c +1 -1
@@ -50,7 +50,7 @@ static int do_hres(const struct vdso_data *vd, clockid_t clk,
 		cycles = __arch_get_hw_counter(vd->clock_mode);
 		ns = vdso_ts->nsec;
 		last = vd->cycle_last;
-		if (unlikely((s64)cycles < 0))
+		if (unlikely(cycles == U64_MAX))
 			return -1;

That used to create worse code than the weird (s64) type cast which has the same effect. Did you double check that there is no change?

It still does for 32bit.

#3 lore raw

Hi, Thomas,

If we use (s64)cycles < 0, then how to solve the problem that a 64bit counter become negative?

Maybe we can change the "invalid" value from U64_MAX to 0? I think the performance of "cycles == 0" is better than "cycles == U64_MAX".

Huacai

Quoted 16 lines · Thomas Gleixner
Quoted 11 lines · Thomas Gleixner

lib/vdso/gettimeofday.c static int do_hres(const struct vdso_data *vd, clockid_t clk, · line 54

Quoted 7 lines · Huacai Chen
lib/vdso/gettimeofday.c +1 -1
@@ -50,7 +50,7 @@ static int do_hres(const struct vdso_data *vd, clockid_t clk,
            cycles = __arch_get_hw_counter(vd->clock_mode);
            ns = vdso_ts->nsec;
            last = vd->cycle_last;
-           if (unlikely((s64)cycles < 0))
+           if (unlikely(cycles == U64_MAX))
                    return -1;

That used to create worse code than the weird (s64) type cast which has the same effect. Did you double check that there is no change?

It still does for 32bit.

#4 lore raw

On Tue, 22 Oct 2019, Huacai Chen wrote:

https://people.kernel.org/tglx/notes-about-netiquette

Look for Toppost

Quoted 2 lines

If we use (s64)cycles < 0, then how to solve the problem that a 64bit counter become negative?

I doubt that you will be able to observe that. A 64bit value becomes negative after 1<<63 cycles, i.e. at 1GHz thats 292 years of uptime.

What's your problem?

Thanks,

tglx