init: Add extra_free_kbytes.sh script to adjust watermark_scale_factor am: 642048d9

                "init.rc",

                "ueventd.rc",

                "e2fsdroid",

                "extra_free_kbytes.sh",

                "make_f2fs",

                "mke2fs",

                "sload_f2fs",

        },

    },

}

sh_binary {

    name: "extra_free_kbytes.sh",

    src: "extra_free_kbytes.sh",

}

#!/bin/sh

# Script implements watermark_scale calculation which results in the same low

# watermark as if extra_free_kbytes tunable were to be used.

#

# Usage: extra_free_kbytes.sh <extra_free_kbytes value>

#

# extra_free_kbytes is distributed between zones based on

# zone.managed_pages/vm_total_pages ratio, where vm_total_pages is the sum of

# zone.managed_pages for all zones (zone.high used in this calculation is 0

# when this is calculated). Therefore for each zone its share is calculated as:

#

# extra_free_pages = extra_free_kbytes / page_size

# extra_share = extra_free_pages * managed_pages / vm_total_pages

#

# This extra_share is added to the low and high watermarks:

#

# low = min + max(min / 4, managed_pages * (watermark_scale / 10000)) + extra_share

# high = min + 2 * max(min / 4, managed_pages * (watermark_scale / 10000)) + extra_share

#

# Because Android uses extra_free_kbytes to adjust the low watermark, we ignore

# the difference in how watermark_scale and extra_free_kbytes affect the high

# watermark and will match the low watermark only.

#

# To eliminate extra_share and compansate the difference with watermark_scale,

# a new watermark_scale_new is calculated as:

#

# (1) max(min / 4, managed_pages * (watermark_scale / 10000)) + extra_share =

#   max(min / 4, managed_pages * (watermark_scale_new / 10000))

#

# Two cases to consider:

# A. managed_pages * (watermark_scale / 10000) > min / 4

# The formula (1) becomes:

#

# managed_pages * (watermark_scale / 10000) + extra_share =

#   managed_pages * (watermark_scale_new / 10000)

#

# after simplifying and substituting extra_share formula becomes:

#

# (2) watermark_scale_new = watermark_scale + extra_free_pages / vm_total_pages * 10000

#

# B. managed_pages * (watermark_scale / 10000) < min / 4

# The formula (1) becomes:

#

# min / 4 + extra_share = max(min / 4, managed_pages * (watermark_scale_new / 10000))

#

# after calculating watermark_scale_new, if (managed_pages * (watermark_scale_new / 10000))

# is still smaller than min / 4 then we can't compensate extra_share with

# watermark_scale anyway. Therefore calculation becomes:

#

# watermark_scale_new = (min / 4 + extra_share) / managed_pages * 10000

#

# after simplifying and substituting extra_share formula becomes:

#

# (3) watermark_scale_new = (min / 4) * 10000 / managed_pages + extra_free_pages / vm_total_pages * 10000

#

# After defining watermark_delta = extra_free_pages / vm_total_pages * 10000:

#

# if (managed_pages * (watermark_scale / 10000) > min / 4)

#     watermark_scale_new = watermark_scale + watermark_delta

# else

#     watermark_scale_new = (min / 4) * 10000 / managed_pages + watermark_delta

#

if [ "$#" -ne 1 ]

then

    echo "Usage: $0 <extra_free_kbytes value>"

    exit

fi

extra_free_kbytes=$1

# if extra_free_kbytes knob exists, use it and exit

if [ -e /proc/sys/vm/extra_free_kbytes ]

then

    echo $extra_free_kbytes > /proc/sys/vm/extra_free_kbytes

    exit

fi

watermark_scale=`cat /proc/sys/vm/watermark_scale_factor`

# convert extra_free_kbytes to pages

page_size=$(getconf PAGESIZE)

page_size_kb=$((page_size/1024))

extra_free_pg=$((extra_free_kbytes/page_size_kb))

managed=($(grep managed /proc/zoneinfo | awk '{print $2}'))

length=${#managed[@]}

min=($(grep "min" /proc/zoneinfo | awk '{print $2}'))

# calculate vm_total_pages.

# WARNING: if the final low watermark differs from the original, the source of

# the error is likely vm_total_pages which is impossible to get exact from the

# userspace. Grep for "Total pages" in the kernel logs to see the actual

# vm_total_pages and plug it in the calculation to confirm the source of the

# error. Error caused by this inaccuracy is normally within 1% range.

vm_total_pages=0

i=0

while [ $i -lt $length ]

do

    vm_total_pages=$((vm_total_pages + managed[i]))

    i=$((i+1))

done

# calculate watermark_scale_new for each zone and choose the max

max_watermark_scale=0

i=0

while [ $i -lt $length ]

do

    # skip unmanaged zones

    if [ ${managed[i]} -eq 0 ]

    then

        i=$((i+1))

        continue

    fi

    base_margin=$((min[i] / 4))

    calc_margin=$(echo "${managed[i]} * $watermark_scale / 10000" | bc)

    # round the value by adding 0.5 and truncating the decimal part

    watermark_delta=$(echo "x=($extra_free_pg / ($vm_total_pages / 10000) + 0.5); scale = 0; x/1" | bc -l)

    if [ $calc_margin -gt $base_margin ]

    then

        watermark_scale_new=$(echo "$watermark_scale + $watermark_delta" | bc)

    else

        watermark_scale_new=$(echo "$base_margin / (${managed[i]} / 10000) + $watermark_delta" | bc)

    fi

    if [ $max_watermark_scale -lt $watermark_scale_new ]

    then

        max_watermark_scale=$watermark_scale_new

    fi

    i=$((i+1))

done

echo $max_watermark_scale > /proc/sys/vm/watermark_scale_factor