Blame SOURCES/0108-Make-pmtimer-tsc-calibration-not-take-51-seconds-to-.patch

8631a2
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
8631a2
From: Peter Jones <pjones@redhat.com>
8631a2
Date: Tue, 7 Nov 2017 17:12:17 -0500
8631a2
Subject: [PATCH] Make pmtimer tsc calibration not take 51 seconds to fail.
8631a2
8631a2
On my laptop running at 2.4GHz, if I run a VM where tsc calibration
8631a2
using pmtimer will fail presuming a broken pmtimer, it takes ~51 seconds
8631a2
to do so (as measured with the stopwatch on my phone), with a tsc delta
8631a2
of 0x1cd1c85300, or around 125 billion cycles.
8631a2
8631a2
If instead of trying to wait for 5-200ms to show up on the pmtimer, we try
8631a2
to wait for 5-200us, it decides it's broken in ~0x2626aa0 TSCs, aka ~2.4
8631a2
million cycles, or more or less instantly.
8631a2
8631a2
Additionally, this reading the pmtimer was returning 0xffffffff anyway,
8631a2
and that's obviously an invalid return.  I've added a check for that and
8631a2
0 so we don't bother waiting for the test if what we're seeing is dead
8631a2
pins with no response at all.
8631a2
8631a2
If "debug" is includes "pmtimer", you will see one of the following
8631a2
three outcomes.  If pmtimer gives all 0 or all 1 bits, you will see:
8631a2
8631a2
kern/i386/tsc_pmtimer.c:77: pmtimer: 0xffffff bad_reads: 1
8631a2
kern/i386/tsc_pmtimer.c:77: pmtimer: 0xffffff bad_reads: 2
8631a2
kern/i386/tsc_pmtimer.c:77: pmtimer: 0xffffff bad_reads: 3
8631a2
kern/i386/tsc_pmtimer.c:77: pmtimer: 0xffffff bad_reads: 4
8631a2
kern/i386/tsc_pmtimer.c:77: pmtimer: 0xffffff bad_reads: 5
8631a2
kern/i386/tsc_pmtimer.c:77: pmtimer: 0xffffff bad_reads: 6
8631a2
kern/i386/tsc_pmtimer.c:77: pmtimer: 0xffffff bad_reads: 7
8631a2
kern/i386/tsc_pmtimer.c:77: pmtimer: 0xffffff bad_reads: 8
8631a2
kern/i386/tsc_pmtimer.c:77: pmtimer: 0xffffff bad_reads: 9
8631a2
kern/i386/tsc_pmtimer.c:77: pmtimer: 0xffffff bad_reads: 10
8631a2
kern/i386/tsc_pmtimer.c:78: timer is broken; giving up.
8631a2
8631a2
This outcome was tested using qemu+kvm with UEFI (OVMF) firmware and
8631a2
these options: -machine pc-q35-2.10 -cpu Broadwell-noTSX
8631a2
8631a2
If pmtimer gives any other bit patterns but is not actually marching
8631a2
forward fast enough to use for clock calibration, you will see:
8631a2
8631a2
kern/i386/tsc_pmtimer.c:121: pmtimer delta is 0x0 (1904 iterations)
8631a2
kern/i386/tsc_pmtimer.c:124: tsc delta is implausible: 0x2626aa0
8631a2
8631a2
This outcome was tested using grub compiled with GRUB_PMTIMER_IGNORE_BAD_READS
8631a2
defined (so as not to trip the bad read test) using qemu+kvm with UEFI
8631a2
(OVMF) firmware, and these options: -machine pc-q35-2.10 -cpu Broadwell-noTSX
8631a2
8631a2
If pmtimer actually works, you'll see something like:
8631a2
8631a2
kern/i386/tsc_pmtimer.c:121: pmtimer delta is 0x0 (1904 iterations)
8631a2
kern/i386/tsc_pmtimer.c:124: tsc delta is implausible: 0x2626aa0
8631a2
8631a2
This outcome was tested using qemu+kvm with UEFI (OVMF) firmware, and
8631a2
these options: -machine pc-i440fx-2.4 -cpu Broadwell-noTSX
8631a2
8631a2
I've also tested this outcome on a real Intel Xeon E3-1275v3 on an Intel
8631a2
Server Board S1200V3RPS using the SDV.RP.B8 "Release" build here:
8631a2
https://firmware.intel.com/sites/default/files/UEFIDevKit_S1200RP_vB8.zip
8631a2
8631a2
Signed-off-by: Peter Jones <pjones@redhat.com>
8631a2
---
8631a2
 grub-core/kern/i386/tsc_pmtimer.c | 109 +++++++++++++++++++++++++++++++-------
8631a2
 1 file changed, 89 insertions(+), 20 deletions(-)
8631a2
8631a2
diff --git a/grub-core/kern/i386/tsc_pmtimer.c b/grub-core/kern/i386/tsc_pmtimer.c
f6e916
index c9c361699..ca15c3aac 100644
8631a2
--- a/grub-core/kern/i386/tsc_pmtimer.c
8631a2
+++ b/grub-core/kern/i386/tsc_pmtimer.c
8631a2
@@ -28,40 +28,101 @@
8631a2
 #include <grub/acpi.h>
8631a2
 #include <grub/cpu/io.h>
8631a2
 
8631a2
+/*
8631a2
+ * Define GRUB_PMTIMER_IGNORE_BAD_READS if you're trying to test a timer that's
8631a2
+ * present but doesn't keep time well.
8631a2
+ */
8631a2
+// #define GRUB_PMTIMER_IGNORE_BAD_READS
8631a2
+
8631a2
 grub_uint64_t
8631a2
 grub_pmtimer_wait_count_tsc (grub_port_t pmtimer,
8631a2
 			     grub_uint16_t num_pm_ticks)
8631a2
 {
8631a2
   grub_uint32_t start;
8631a2
-  grub_uint32_t last;
8631a2
-  grub_uint32_t cur, end;
8631a2
+  grub_uint64_t cur, end;
8631a2
   grub_uint64_t start_tsc;
8631a2
   grub_uint64_t end_tsc;
8631a2
-  int num_iter = 0;
8631a2
+  unsigned int num_iter = 0;
8631a2
+#ifndef GRUB_PMTIMER_IGNORE_BAD_READS
8631a2
+  int bad_reads = 0;
8631a2
+#endif
8631a2
 
8631a2
-  start = grub_inl (pmtimer) & 0xffffff;
8631a2
-  last = start;
8631a2
+  /*
8631a2
+   * Some timers are 24-bit and some are 32-bit, but it doesn't make much
8631a2
+   * difference to us.  Caring which one we have isn't really worth it since
8631a2
+   * the low-order digits will give us enough data to calibrate TSC.  So just
8631a2
+   * mask the top-order byte off.
8631a2
+   */
8631a2
+  cur = start = grub_inl (pmtimer) & 0xffffffUL;
8631a2
   end = start + num_pm_ticks;
8631a2
   start_tsc = grub_get_tsc ();
8631a2
   while (1)
8631a2
     {
8631a2
-      cur = grub_inl (pmtimer) & 0xffffff;
8631a2
-      if (cur < last)
8631a2
-	cur |= 0x1000000;
8631a2
-      num_iter++;
8631a2
+      cur &= 0xffffffffff000000ULL;
8631a2
+      cur |= grub_inl (pmtimer) & 0xffffffUL;
8631a2
+
8631a2
+      end_tsc = grub_get_tsc();
8631a2
+
8631a2
+#ifndef GRUB_PMTIMER_IGNORE_BAD_READS
8631a2
+      /*
8631a2
+       * If we get 10 reads in a row that are obviously dead pins, there's no
8631a2
+       * reason to do this thousands of times.
8631a2
+       */
8631a2
+      if (cur == 0xffffffUL || cur == 0)
8631a2
+	{
8631a2
+	  bad_reads++;
8631a2
+	  grub_dprintf ("pmtimer",
8631a2
+			"pmtimer: 0x%"PRIxGRUB_UINT64_T" bad_reads: %d\n",
8631a2
+			cur, bad_reads);
8631a2
+	  grub_dprintf ("pmtimer", "timer is broken; giving up.\n");
8631a2
+
8631a2
+	  if (bad_reads == 10)
8631a2
+	    return 0;
8631a2
+	}
8631a2
+#endif
8631a2
+
8631a2
+      if (cur < start)
8631a2
+	cur += 0x1000000;
8631a2
+
8631a2
       if (cur >= end)
8631a2
 	{
8631a2
-	  end_tsc = grub_get_tsc ();
8631a2
+	  grub_dprintf ("pmtimer", "pmtimer delta is 0x%"PRIxGRUB_UINT64_T"\n",
8631a2
+			cur - start);
8631a2
+	  grub_dprintf ("pmtimer", "tsc delta is 0x%"PRIxGRUB_UINT64_T"\n",
8631a2
+			end_tsc - start_tsc);
8631a2
 	  return end_tsc - start_tsc;
8631a2
 	}
8631a2
-      /* Check for broken PM timer.
8631a2
-	 50000000 TSCs is between 5 ms (10GHz) and 200 ms (250 MHz)
8631a2
-	 if after this time we still don't have 1 ms on pmtimer, then
8631a2
-	 pmtimer is broken.
8631a2
+
8631a2
+      /*
8631a2
+       * Check for broken PM timer.  1ms at 10GHz should be 1E+7 TSCs; at
8631a2
+       * 250MHz it should be 2.5E6.  So if after 4E+7 TSCs on a 10GHz machine,
8631a2
+       * we should have seen pmtimer show 4ms of change (i.e. cur =~
8631a2
+       * start+14320); on a 250MHz machine that should be 16ms (start+57280).
8631a2
+       * If after this a time we still don't have 1ms on pmtimer, then pmtimer
8631a2
+       * is broken.
8631a2
+       *
8631a2
+       * Likewise, if our code is perfectly efficient and introduces no delays
8631a2
+       * whatsoever, on a 10GHz system we should see a TSC delta of 3580 in
8631a2
+       * ~3580 iterations.  On a 250MHz machine that should be ~900 iterations.
8631a2
+       *
8631a2
+       * With those factors in mind, there are two limits here.  There's a hard
8631a2
+       * limit here at 8x our desired pm timer delta, picked as an arbitrarily
8631a2
+       * large value that's still not a lot of time to humans, because if we
8631a2
+       * get that far this is either an implausibly fast machine or the pmtimer
8631a2
+       * is not running.  And there's another limit on 4x our 10GHz tsc delta
8631a2
+       * without seeing cur converge on our target value.
8631a2
        */
8631a2
-      if ((num_iter & 0xffffff) == 0 && grub_get_tsc () - start_tsc > 5000000) {
8631a2
-	return 0;
8631a2
-      }
8631a2
+      if ((++num_iter > (grub_uint32_t)num_pm_ticks << 3UL) ||
8631a2
+	  end_tsc - start_tsc > 40000000)
8631a2
+	{
8631a2
+	  grub_dprintf ("pmtimer",
8631a2
+			"pmtimer delta is 0x%"PRIxGRUB_UINT64_T" (%u iterations)\n",
8631a2
+			cur - start, num_iter);
8631a2
+	  grub_dprintf ("pmtimer",
8631a2
+			"tsc delta is implausible: 0x%"PRIxGRUB_UINT64_T"\n",
8631a2
+			end_tsc - start_tsc);
8631a2
+	  return 0;
8631a2
+	}
8631a2
     }
8631a2
 }
8631a2
 
8631a2
@@ -74,12 +135,20 @@ grub_tsc_calibrate_from_pmtimer (void)
8631a2
 
8631a2
   fadt = grub_acpi_find_fadt ();
8631a2
   if (!fadt)
8631a2
-    return 0;
8631a2
+    {
8631a2
+      grub_dprintf ("pmtimer", "No FADT found; not using pmtimer.\n");
8631a2
+      return 0;
8631a2
+    }
8631a2
   pmtimer = fadt->pmtimer;
8631a2
   if (!pmtimer)
8631a2
-    return 0;
8631a2
+    {
8631a2
+      grub_dprintf ("pmtimer", "FADT does not specify pmtimer; skipping.\n");
8631a2
+      return 0;
8631a2
+    }
8631a2
 
8631a2
-  /* It's 3.579545 MHz clock. Wait 1 ms.  */
8631a2
+  /*
8631a2
+   * It's 3.579545 MHz clock. Wait 1 ms.
8631a2
+   */
8631a2
   tsc_diff = grub_pmtimer_wait_count_tsc (pmtimer, 3580);
8631a2
   if (tsc_diff == 0)
8631a2
     return 0;