package/kernel/linux/up-next-acpi-usleep_range.patch

   1 From: "Rafael J. Wysocki" <rjw@rjwysocki.net>
   2 To: Linux ACPI <linux-acpi@vger.kernel.org>
   3 Cc: LKML <linux-kernel@vger.kernel.org>, Linux PM <linux-pm@vger.kernel.org>,
   4  Len Brown <len.brown@intel.com>, Arjan van de Ven <arjan@linux.intel.com>,
   5  Pierre Gondois <pierre.gondois@arm.com>,
   6  Dietmar Eggemann <dietmar.eggemann@arm.com>,
   7  Hans de Goede <hdegoede@redhat.com>,
   8  Mario Limonciello <mario.limonciello@amd.com>,
   9  "Gautham R. Shenoy" <gautham.shenoy@amd.com>
  10 Subject: [PATCH v1] ACPI: OSL: Use usleep_range() in acpi_os_sleep()
  11 Date: Thu, 05 Dec 2024 12:24:35 +0100
  12 Message-ID: <5857066.DvuYhMxLoT@rjwysocki.net>
  13 Precedence: bulk
  14 X-Mailing-List: linux-kernel@vger.kernel.org
  15 List-Id: <linux-kernel.vger.kernel.org>
  16 List-Subscribe: <mailto:linux-kernel+subscribe@vger.kernel.org>
  17 List-Unsubscribe: <mailto:linux-kernel+unsubscribe@vger.kernel.org>
  18 MIME-Version: 1.0
  19 Content-Transfer-Encoding: 7Bit
  20 Content-Type: text/plain; charset="UTF-8"
  21
  22 From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
  23 Subject: [PATCH v0.1] ACPI: OSL: Use usleep_range() in acpi_os_sleep()
  24
  25 As stated by Len in [1], the extra delay added by msleep() to the
  26 sleep time value passed to it can be significant, roughly between
  27 1.5 ns on systems with HZ = 1000 and as much as 15 ms on systems with
  28 HZ = 100, which is hardly acceptable, at least for small sleep time
  29 values.
  30
  31 msleep(5) on the default HZ = 250 in Ubuntu on a modern PC takes about
  32 12 ms.  This results in over 800 ms of spurious system resume delay on
  33 systems such as the Dell XPS-13-9300, which use ASL Sleep(5ms) in a
  34 tight loop.
  35
  36 Address this by using usleep_range() in acpi_os_sleep() instead of
  37 msleep().  For short sleep times this is a no brainer, but even for
  38 long sleeps usleep_range() should be preferred because timer wheel
  39 timers are optimized for cancelation before they expire and this
  40 particular timer is not going to be canceled.
  41
  42 Add at least 50 us on top of the requested sleep time in case the
  43 timer can be subject to coalescing, which is consistent with what's
  44 done in user space in this context [2], but for sleeps longer than 5 ms
  45 use 1% of the requested sleep time for this purpose.
  46
  47 The rationale here is that longer sleeps don't need that much of a timer
  48 precision as a rule and making the timer a more likely candidate for
  49 coalescing in these cases is generally desirable.  It starts at 5 ms so
  50 that the delta between the requested sleep time and the effective
  51 deadline is a contiuous function of the former.
  52
  53 Link: https://lore.kernel.org/linux-pm/c7db7e804c453629c116d508558eaf46477a2d73.1731708405.git.len.brown@intel.com/ [1]
  54 Link: https://lore.kernel.org/linux-pm/
  55 CAJvTdK=Q1kwWA6Wxn8Zcf0OicDEk6cHYFAvQVizgA47mXu63+g@mail.gmail.com/ [2]
  56 Closes: https://bugzilla.kernel.org/show_bug.cgi?id=216263
  57 Reported-by: Len Brown <lenb@kernel.org>
  58 Reviewed-by: Hans de Goede <hdegoede@redhat.com>
  59 Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
  60 Tested-by: Mario Limonciello <mario.limonciello@amd.com>
  61 Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
  62 ---
  63
  64 The previous RFC version of this patch is here:
  65
  66 https://lore.kernel.org/linux-pm/5839859.DvuYhMxLoT@rjwysocki.net/
  67
  68 The difference between this version and the RFC is a changelog update
  69 suggested by Len, the Closes: tag and the tags from Hans and Mario.
  70
  71 ---
  72  drivers/acpi/osl.c |   22 +++++++++++++++++++++-
  73  1 file changed, 21 insertions(+), 1 deletion(-)
  74
  75 Index: linux-pm/drivers/acpi/osl.c
  76 ===================================================================
  77 --- linux-pm.orig/drivers/acpi/osl.c
  78 +++ linux-pm/drivers/acpi/osl.c
  79 @@ -607,7 +607,27 @@ acpi_status acpi_os_remove_interrupt_han
  80
  81  void acpi_os_sleep(u64 ms)
  82  {
  83 -       msleep(ms);
  84 +       u64 usec = ms * USEC_PER_MSEC, delta_us = 50;
  85 +
  86 +       /*
  87 +        * Use a hrtimer because the timer wheel timers are optimized for
  88 +        * cancelation before they expire and this timer is not going to be
  89 +        * canceled.
  90 +        *
  91 +        * Set the delta between the requested sleep time and the effective
  92 +        * deadline to at least 50 us in case there is an opportunity for timer
  93 +        * coalescing.
  94 +        *
  95 +        * Moreover, longer sleeps can be assumed to need somewhat less timer
  96 +        * precision, so sacrifice some of it for making the timer a more likely
  97 +        * candidate for coalescing by setting the delta to 1% of the sleep time
  98 +        * if it is above 5 ms (this value is chosen so that the delta is a
  99 +        * continuous function of the sleep time).
 100 +        */
 101 +       if (ms > 5)
 102 +               delta_us = (USEC_PER_MSEC / 100) * ms;
 103 +
 104 +       usleep_range(usec, usec + delta_us);
 105  }
 106
 107  void acpi_os_stall(u32 us)
 108
 109
 110
 111