#!/bin/sh

# Copyright (c) 2010 The Chromium OS Authors. All rights reserved.
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.

# This script is responsible for suspending and resuming the system.  It is run
# as root by powerd_setuid_helper, which is run by powerd.

. /usr/share/misc/shflags

DEFINE_integer wakeup_count -1 \
  "wakeup event count from start of suspend. -1 to disable" w
DEFINE_boolean suspend_to_idle false \
  "indicates that the system should suspend to idle (freeze)" i

# Exit codes returned by this script and reported via the
# Power.SuspendResult histogram:

# The system successfully suspended and resumed.
readonly RESULT_SUCCESS=0

# A suspend failure occurred in the kernel after writing to /sys/power/state
# (i.e. the write to /sys/power/state failed with a non-EBUSY reason).
readonly RESULT_FAILURE=1

# The suspend attempt was canceled as the result of a wakeup event occurring
# between powerd's read from /sys/power/wakeup_count and this script's write to
# it (i.e. the write to wakeup_count failed).
readonly RESULT_CANCELED_EARLY=2

# The suspend attempt was canceled as the result of a wakeup event occurring
# between the write to /sys/power/wakeup_count and the write to
# /sys/power/state (i.e. the write to /sys/power/state failed with EBUSY).
readonly RESULT_CANCELED_LATE=3

# Histogram-only max value.
readonly RESULT_MAX=4

# Directory where this script and send_metrics_on_resume (running as root)
# write temporary state files.
readonly ROOT_RUN_DIR=/run/power_manager/root

# For testing: write RTC resume timestamp to this file if it exists
readonly TIMESTAMP_FILE="${ROOT_RUN_DIR}"/hwclock-on-resume

# File used to record the power status (AC vs. battery) before suspending.  Read
# by send_metrics_on_resume (but only when run after resuming; this doesn't
# need to persist across reboots).
readonly POWER_STATUS_ON_SUSPEND_FILE="${ROOT_RUN_DIR}"/power-status-on-suspend

# File containing resume-related timing information.  Written by
# send_metrics_on_resume and read by autotests; doesn't need to persist
# across reboots.
readonly LAST_RESUME_TIMINGS_FILE="${ROOT_RUN_DIR}"/last_resume_timings

# Directory where this script (running as root) writes files that must
# persist across reboots.
readonly ROOT_SPOOL_DIR=/var/spool/power_manager/root

# File containing the time at which we started suspending.  Read by
# send_metrics_on_resume.
readonly SUSPEND_TO_RAM_TIME_FILE="${ROOT_SPOOL_DIR}"/suspend-to-ram-time

# File containing suspend-related information to log on failure.
readonly SUSPEND_STATS_FILE=/sys/kernel/debug/suspend_stats

# Compact if order 3 page count is less than this.
readonly COMPACT_ORDER3_THRESHOLD=500

# For testing: do not compact before suspending if this file exists.
readonly NO_COMPACT_FILE="${ROOT_RUN_DIR}"/no-compact

log_msg() {
  logger -t "powerd_suspend" --id="$$" -- "$@"
}

log_msg_from_file() {
  logger -t "powerd_suspend" --id="$$" -f "$1"
}

log_wakeup_count() {
  log_msg "wakeup_count is $(cat /sys/power/wakeup_count)"
}

log_suspend_stats() {
  log_msg "--- begin ${SUSPEND_STATS_FILE} ---"
  local tempfile="$(mktemp --tmpdir powerd_suspend_stats.XXXXXXXXXX)"
  # Omit blank lines and replace tabs with spaces so syslog won't escape them
  # to "#011".
  grep -e '\S' "${SUSPEND_STATS_FILE}" | tr '\t' ' ' >"${tempfile}"
  log_msg_from_file "${tempfile}"
  log_msg "--- end ${SUSPEND_STATS_FILE} ---"
  rm "${tempfile}"
}

is_kip() {
  crossystem hwid | grep -q '^KIP '
}

igb_devices=""

# The Intel gigabit driver (igb) prevents Intel platforms (the only platforms
# that use this driver currently) from going into suspend-to-idle i.e. S0ix
# (see b/189345571). So in order to enter S0ix, need to unbind its devices while
# suspending, and rebind again at resume.
unbind_igb_devices() {
  /sbin/lsmod | grep -qw "^igb"
  if [ $? -ne 0 ]; then
    return
  fi

  for pcidev in `readlink /sys/class/net/*/device`; do
    pcidev="$(basename ${pcidev})"
    pcidrvr="$(readlink /sys/bus/pci/devices/${pcidev}/driver)"
    pcidrvr="$(basename ${pcidrvr})"

    if [ "${pcidrvr}" = "igb" ]; then
      log_msg "Temporarily unbinding igb driver from ${pcidev} for suspend"
      echo "${pcidev}" > /sys/bus/pci/drivers/igb/unbind
      igb_devices="${igb_devices} ${pcidev}"
    fi
  done
}

bind_igb_devices() {
  for pcidev in ${igb_devices}; do
    log_msg "Re-binding igb driver to ${pcidev} after resume"
    echo "${pcidev}" > /sys/bus/pci/drivers/igb/bind
  done
}

# Checks if the system has a Huawei Technologies Mobile Broadband Module.
has_huawei_modem() {
  lsusb -d 12d1:15bb
}

# Powers "on" or "off" the built-in modem via EC.
ec_power_modem() {
  local op="$1"

  case "${op}" in
    on)
      ectool wireless 0x4 0x4
      ;;
    off)
      ectool wireless 0x0 0x4
      ;;
    *)
      log_msg "Invalid operation 'ec_power_modem ${op}'"
      return 1
      ;;
  esac

  if [ $? -eq 0 ]; then
    log_msg "Power ${op} built-in modem"
    return 0
  else
    log_msg "Failed to power ${op} built-in modem"
    return 1
  fi
}

to_hex() {
  echo "obase=16; $(($1))" | bc
}


# Exit Sniff mode for all ACL connections.
#
# This workaround is to make sure BlueZ stack and the controller are aligned
# on connections state after resume, since the 'sniff mode change' event could
# have been masked out while the system was is suspend mode
bt_exit_sniff() {
  local file
  for file in /sys/class/bluetooth/hci[0-9]; do
    local hcidev="$(basename "${file}")"
    # The command of "hcitool con" shows something like
    # Connections:
    #    > ACL 04:52:C7:83:95:10 handle 256 state 1 lm MASTER AUTH ENCRYPT
    #    > ACL 04:52:C7:C3:65:B5 handle 512 state 1 lm MASTER AUTH ENCRYPT

    handles="$(echo "$(hcitool con)" | awk '/ACL/ {print $5}')"
    for handle in $handles; do
      lo_byte="$(to_hex $((handle % 256)))"
      hi_byte="$(to_hex $((handle / 256)))"
      log_msg "Exiting Sniff mode for connection handle ${lo_byte} ${hi_byte}"
      timeout -k 2 2 hcitool -i "${hcidev}" cmd 02 04 "${lo_byte}" \
          "${hi_byte}"
     done
   done
}

# Writes the wakeup count that was passed to this script to
# /sys/power/wakeup_count.  Returns success if the write fails, indicating
# that one or more wakeup events occurred in the meantime.
saw_wakeup_event() {
  if [ "${FLAGS_wakeup_count}" -eq -1 ] ||
    echo "${FLAGS_wakeup_count}" > /sys/power/wakeup_count; then
    return 1
  else
    log_msg "Aborting suspend, wake event received"
    log_wakeup_count
    return 0
  fi
}


# Sends suspend result (one of the above RESULT_* values) to UMA.
send_uma_result() {
  metrics_client -e Power.SuspendResult "$1" "${RESULT_MAX}" &
}

# Gets the allocatable order 3 page count. Using order 3 threshold because
# there are many order 3 allocation by network modules when running low memory
# suspending stress test. /proc/pagetypeinfo provides available high order
# pages count, example:
# Free pages count per migrate type at order    0    1    2    3   4 ... 10
# ...
# Node    0, zone    DMA32, type    Unmovable   1    9   29    5   2 ...  0
# Node    0, zone    DMA32, type  Reclaimable 160  204   17    8   0 ...  0
# Node    0, zone    DMA32, type      Movable   0 1163 1530 1236 298 ...  0
# Node    0, zone    DMA32, type      Reserve   0    0    0    0   0 ...  2
# Movable order 3 field 1236: there are 1236 order 3 pages ((2 ^ 3) * 4KB
# consecutive memory).
order3_available() {
  awk '(NF == 17) &&
    ($6 == "Unmovable" || $6 == "Movable" || $6 == "Reclaimable") {
    size = 1
    for (i = 10; i <= 17; i++) {
      sum += $i * size
      size *= 2
    }
  }
  END { print sum }' /proc/pagetypeinfo
}

# When system is suspending, memory compaction and swap are disabled. If
# memory is fragmented at this time, high order allocation would fail.
# When memory fragmentation is detected (too few allocatable order 3 pages),
# compact memory before suspending. For 4.4 or earlier kernel without
# kcompactd.
# TODO(crbug.com/827053): Check if compaction before suspending is necessary
# on newer kernel.
try_compact_memory() {
  local order3="$(order3_available)"
  log_msg "Available order 3 pages: ${order3}"
  if uname -r | grep -q "^3\.\|^4\.4\." && [ ! -f ${NO_COMPACT_FILE} ] && \
    [ "${order3}" -le "${COMPACT_ORDER3_THRESHOLD}" ]; then
    log_msg "Compacting memory before suspend"
    echo 1 > /proc/sys/vm/compact_memory
    log_msg "Available order 3 pages after compaction: $(order3_available)"
  fi
}

# Encourage people to use powerd_dbus_suspend instead of running this
# script directly so that e.g. suspend delays will happen.
if [ -z "${POWERD_SETUID_HELPER}" ]; then
  echo "This script is called by powerd. Please run powerd_dbus_suspend" 1>&2
  echo "to manually exercise the full suspend path." 1>&2
  exit 1
fi

# Note: Don't change or remove this line without also updating
# send_metrics_on_resume.
log_msg "Going to suspend-to-RAM state: args=$@"

# Parse command line.
FLAGS "$@" || exit 1
eval set -- "${FLAGS_ARGV}"

# The metrics library requires a max value of 2 rather than 1
# (http://crbug.com/338015).
metrics_client -e Power.SuspendAttempt 0 2 &

# Log the time before going to suspend (no-op if no RTC).
cp /sys/class/rtc/rtc0/since_epoch "${SUSPEND_TO_RAM_TIME_FILE}" \
  2> /dev/null || true

# Remove last_resume_timings to ensure the file is fresh on resume.
rm -f "${LAST_RESUME_TIMINGS_FILE}"

# Store the current power status.
power_supply_info 2> /dev/null \
  | grep -Eq '^[[:space:]]+online:[[:space:]]+no$'
if [ $? -eq 0 ]; then
  echo OnBattery > "${POWER_STATUS_ON_SUSPEND_FILE}"
else
  echo OnAC > "${POWER_STATUS_ON_SUSPEND_FILE}"
fi

# Is this a Kip system with Huawei modem?
kip_with_huawei_modem=0
if is_kip && has_huawei_modem; then
  log_msg "Kip with Huawei modem"
  kip_with_huawei_modem=1
fi

result="${RESULT_FAILURE}"

if saw_wakeup_event; then
  # Note: The client/cros/power_suspend.py module in autotest depends on
  # this message.
  log_msg "Cancel suspend at kernel"

  # This file is usually removed by send_metrics_on_resume, but that script
  # isn't run if the suspend attempt is unsuccessful.
  rm -f "${SUSPEND_TO_RAM_TIME_FILE}"

  result="${RESULT_CANCELED_EARLY}"
else
  # Before suspending, power down problematic modems if necessary.
  # `modem_shutdown` (along with `modem_startup`) exists only if board-specific
  # overlays provide one. Future project should re-use the script name when
  # powering down modems before suspending is required.
  # See https://crrev.com/c/2963943 as an example.
  if [ -x "/usr/sbin/modem_shutdown" ]; then
    /usr/sbin/modem_shutdown
  fi
  if [ "${kip_with_huawei_modem}" -eq 1 ]; then
    ec_power_modem off
  fi

  # Compact memory if it is fragmented.
  try_compact_memory

  log_msg "Finalizing suspend"
  # Make sure that hwclock is cached for when we wake up, this way we don't
  # wait for io and get a more accurate time for wakeup
  # -V prints a little version information and exits, loading .so but not
  # making actual RTC accesses. Do *NOT* change this to a normal execution...
  # it will trigger an RTC interrupt that may count as a wakeup reason, abort
  # your suspend and cause you hours of pain and confusion!
  if [ -f "${TIMESTAMP_FILE}" ]; then
    /sbin/hwclock -V > /dev/null
  fi

  # Echo freeze to /sys/power/state on platforms that support suspend to idle.
  action='mem'
  if [ "${FLAGS_suspend_to_idle}" -eq 0 ]; then
    action='freeze'
    unbind_igb_devices
  fi

  # Suspend to RAM. This is piped through cat since we want to determine
  # the causes of failures -- dash's "echo" command appears to not print
  # anything in response to write errors.
  error=$( (printf '%s' "${action}" | cat > /sys/power/state) 2>&1)
  if [ $? -eq 0 ]; then
    # On resume:
    result="${RESULT_SUCCESS}"

    if [ -f "${TIMESTAMP_FILE}" ]; then
      # Record the hwclock time at resume for the power_Resume test.
      if /sbin/hwclock -r --utc > "${TIMESTAMP_FILE}" 2>&1; then
        log_msg "Recorded hwclock time for test"
      else
        log_msg "/sbin/hwclock failed: $(cat "${TIMESTAMP_FILE}")"
        log_msg "debug: $(/sbin/hwclock -r --verbose --utc 2>&1)"
      fi
    fi

    # TODO (b/193451073): Move this to powerd to reduce the delay of blocking
    # display turning on.
    if [ -e /sys/firmware/log ]; then
      # Grep /sys/firmware/log for WAK for x86 platforms.
      wakeup_source="$(grep "PM1_STS: WAK" /sys/firmware/log | tail -n 1)"
      # Try another pattern for MTK (ARM) platforms.
      if [ -z "${wakeup_source}" ]; then
        wakeup_source="$(grep "wake up by" /sys/firmware/log | tail -n 1)"
      fi
      # Found nothing in the end, the wakeup source is unknown.
      if [ -z "${wakeup_source}" ]; then
        wakeup_source='unknown'
      fi
      log_msg "wake source: ${wakeup_source}"
    fi

    # Send UMA metrics.
    send_metrics_on_resume &
  else
    # The write will fail with EBUSY if additional wakeup events have
    # occurred since the earlier write to /sys/power/wakeup_count. Since this
    # can happen if e.g. a device fails to suspend, we check the wakeup count
    # again to make sure it has actually gone up.
    current_wakeup_count="$(cat /sys/power/wakeup_count)"
    if echo "${error}" | grep -q 'Device or resource busy' && \
      [ "${FLAGS_wakeup_count}" -ne -1 ] && \
      [ "${FLAGS_wakeup_count}" -ne "${current_wakeup_count}" ]; then
      log_msg "Warning: Device or resource busy on write to /sys/power/state"
      log_suspend_stats
      log_wakeup_count
      result="${RESULT_CANCELED_LATE}"
    else
      # Note: The client/cros/power_suspend.py module in autotest depends on
      # this message.
      log_msg "Error writing to /sys/power/state: ${error}"
      log_suspend_stats
      result="${RESULT_FAILURE}"
    fi
  fi

  bind_igb_devices

  # Power back on modems if they were powered off before suspend.
  if [ "${kip_with_huawei_modem}" -eq 1 ]; then
    ec_power_modem on
  fi
  # `modem_startup` (along with `modem_shutdown`) exists only if board-specific
  # overlays provide one. Future project should re-use the script name when
  # powering back on modems after suspending is required.
  # See https://crrev.com/c/2963943 as an example.
  if [ -x "/usr/sbin/modem_startup" ]; then
    /usr/sbin/modem_startup
  fi

  bt_exit_sniff
fi

send_uma_result "${result}"

# Refresh mosys eventlog to help feedback reports pick up the latest snapshot.
/usr/share/userfeedback/scripts/eventlog &

log_msg "Resume finished"

return "${result}"