Trim with LVM and dm-crypt

lvm ssd troubleshooting dm-crypt trim

11,650

Solution 1

I suggest using a different testing method. hdparm is a bit weird as it gives device addresses rather than filesystem addresses, and it doesn't say which device those addresses relate to (e.g. it resolves partitions, but not devicemapper targets, etc.). Much easier to use something that sticks with filesystem addresses, that way it's consistent (maybe except for non-traditional filesystems like zfs/btrfs).

Create a test file: (not random on purpose)

# yes | dd iflag=fullblock bs=1M count=1 of=trim.test

Get the address, length and blocksize: (exact command depends on filefrag version)

# filefrag -s -v trim.test
File size of trim.test is 1048576 (256 blocks, blocksize 4096)
 ext logical physical expected length flags
   0       0    34048             256 eof
trim.test: 1 extent found

Get the device and mountpoint:

# df trim.test
/dev/mapper/something  32896880 11722824  20838512   37% /mount/point

With this set up, you have a file trim.test filled with yes-pattern on /dev/mapper/something at address 34048 with length of 256 blocks of 4096 bytes.

Reading that from the device directly should produce the yes-pattern:

# dd bs=4096 skip=34048 count=256 if=/dev/mapper/something | hexdump -C
00000000  79 0a 79 0a 79 0a 79 0a  79 0a 79 0a 79 0a 79 0a  |y.y.y.y.y.y.y.y.|
*
00100000

If TRIM is enabled, this pattern should change when you delete the file. Note that caches need to be dropped also, otherwise dd will not re-read the data from disk.

# rm trim.test
# sync
# fstrim -v /mount/point/ # when not using 'discard' mount option
# echo 1 > /proc/sys/vm/drop_caches
# dd bs=4096 skip=34048 count=256 if=/dev/mapper/something | hexdump -C

On most SSD that would result in a zero pattern:

00000000  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |................|
*
00100000

If encryption is involved, you will see a random pattern instead:

00000000  1f c9 55 7d 07 15 00 d1  4a 1c 41 1a 43 84 15 c0  |..U}....J.A.C...|
00000010  24 35 37 fe 05 f7 43 93  1e f4 3c cc d8 83 44 ad  |$57...C...<...D.|
00000020  46 80 c2 26 13 06 dc 20  7e 22 e4 94 21 7c 8b 2c  |F..&... ~"..!|.,|

That's because physically trimmed, the crypto layer reads zeroes and decrypts those zeroes to "random" data.

If the yes-pattern persists, most likely no trimming has been done.

Solution 2

This is just a script I would like to share if some lazy person come here. It was made out of the accepted answer from frostschutz.


#!/bin/bash
#
# This script is provided "as is" without warranty of any kind, either expressed or implied, including, but not limited to, the implied warranties of merchantability, fitness for a particular purpose, or non-infringement.
#
# License GPL2
#
# by desgua 2014/04/29

function CLEAN {
cd "$pasta"
[ -f test-trim-by-desgua ] && rm test-trim-by-desgua && echo "Temp file removed"
echo "Goodbye"
exit 0
}

trap 'echo ; echo "Aborted." ; CLEAN; echo ; exit 0' INT HUP

if [[ "$(echo $USER)" != "root" ]]; then

read -n 1 -p 'Become root? [Y/n]' a
    if [[ $a == "Y" || $a == "y" || $a == "" ]]; then
        sudo $0 $1
        exit 0
    else
        echo "
        This script needs root privilege.
        "
        exit 1

    fi

fi


name=$(echo $0 | sed 's/.*\///')
if [ $# -ne 1 ]; then

echo "
Usage: $name /folder/to/test/

"
exit 1
fi

pasta=$1

read -n 1 -p 'Use fstrim? [y/N]' a
if [[ $a == "Y" || $a == "y" ]]; then
    fs=1
fi

method=
while [[ "$method" != "1" && "$method" != "2" ]]; do
read -n 1 -s -p 'Choose a method:
[1] hdparm (will fail in LUKS on LVM)
[2] filefrag (warning: you may have to force quit - close the terminal - in some cases of success trim if you see an output that never ends) 
' method
done

function SDATEST {
disk=$(fdisk -l | grep /dev/sda)
if [ "$disk" == "" ]; then
echo "
fdisk did not found /dev/sda 
"
exit 1
fi
}

function TEST {
echo "Entrying /" ; echo
cd $pasta
echo "Creating the file test-trim-by-desgua at $pasta" ; echo
dd if=/dev/urandom of=test-trim-by-desgua count=10 bs=512k
echo "Syncing and sleeping 2 seconds." ; echo
sync
sleep 2

hdparm --fibmap test-trim-by-desgua
lbab=$(hdparm --fibmap test-trim-by-desgua | tail -n1 | awk '{ print $2 }')

echo "As you can see, the file was created and its LBA begins at $lbab" ; echo

echo "Syncing and sleeping 2 seconds." ; echo
sync
sleep 2

echo "Removing file test-trim-by-desgua" ; echo 
rm test-trim-by-desgua

trap 'echo ; echo ; echo "Aborted." ; echo ; exit 0' INT
echo "Syncing and sleeping 2 seconds." ; echo
sync
sleep 2

if [[ "$fs" == "1" ]]; then 
    echo "fstrim $pasta && sleep 2" ; echo
    fstrim $pasta
    sleep 2
fi

echo "This is readed from sector $lbab: "
hdparm --read-sector $lbab /dev/sda

pass=$(hdparm --read-sector $lbab /dev/sda | grep "0000 0000 0000 0000")

if [[ $pass == "" ]]; then
    echo "
Trim failed... 
You should see only 0000 0000 0000 0000 ...
"
else
    echo "Success!!!"
fi
exit 0

}

function LUKSTEST {
# Reference: https://unix.stackexchange.com/questions/85865/trim-with-lvm-and-dm-crypt#
echo 1 > /proc/sys/vm/drop_caches
cd $pasta
echo "Creating a \"yes\" file."
yes | dd iflag=fullblock bs=1M count=1 of=test-trim-by-desgua

#position=`filefrag -s -v test-trim-by-desgua | grep "eof" | awk '{ print $3 }'`
position=`filefrag -s -v test-trim-by-desgua | grep "eof" | sed 's| ||g ; s|.*255:|| ; s|\.\..*||'`
[[ "$position" == "" ]] && echo "Could not find the position of the file. Are you on a LUKS on LVM?" && CLEAN;

device=`df test-trim-by-desgua | grep "dev/" | awk '{ print $1 }'`

yes=`dd bs=4096 skip=$position count=256 if=$device | hexdump -C`

echo "In the next line you should see a pattern like: 
00000000  79 0a 79 0a 79 0a 79 0a  79 0a 79 0a 79 0a 79 0a  |y.y.y.y.y.y.y.y.|
$yes
"

if [[ "`echo "$yes" | grep "y.y.y"`" == "" ]]; then 
    echo "The pattern could not be checked. Something went wrong. Exiting."
    CLEAN;
else
    echo "Pattern confirmed."
fi

echo "Removing the temp file." 
rm test-trim-by-desgua

echo "Syncing."
sync
sleep 1

if [[ "$fs" == "1" ]]; then 
    echo "fstrim -v $pasta && sleep 2" ; echo
    fstrim -v $pasta
    sleep 2
fi

# Drop cache
echo 1 > /proc/sys/vm/drop_caches

echo "In the next line you should **NOT** see a yes pattern like: 
00000000  79 0a 79 0a 79 0a 79 0a  79 0a 79 0a 79 0a 79 0a  |y.y.y.y.y.y.y.y.| 
If you see, then trim is not working:
`dd bs=4096 skip=$position count=256 if=$device | hexdump -C`"

yes=`dd bs=4096 skip=$position count=256 if=$device | hexdump -C`
if [[ "`echo "$yes" | grep "y.y.y"`" != "" ]]; then 
    echo "TRIM not working."
else
    echo "TRIM is working!"
fi
CLEAN;
}

if [[ "$method" == "1" ]]; then
    SDATEST;
    TEST;
elif [[ "$method" == "2" ]]; then
    LUKSTEST;
fi
exit 0

Solution 3

Your test routine is wrong—you're getting sector numbers relative to the block device which the filesystem sits on—which in this case, is a logical volume. The logical volume, of course, does not start at the first sector of the physical volume (and may not even be contiguous).

Even if the logical volume started at sector 0 of the physical volume (which it doesn't), then the physical volume is actually another device-mapper target, this one for encryption. And probably there is a LUKS header in front, so the sector numbers don't match there, either.

If you want to work through mapping the sector number to the underlying disk, dmsetup tables will give you the info you need. If you paste it here, make sure yours is a version that does not show the key in the output (it should show all 0's instead)! (There is no recovery from disclosing the key—it can not be changed—it is far worse than disclosing the password).

I suggest that to debug (once you get the sector mapping worked out) you start at the lowest level, and confirm it works there. TRIM a filesystem directly on /dev/sdaX and make sure that works (its quite possible that the device lies, and trim doesn't read back zeros). Then dm-crypt on top of that, and trim a filesystem on that, and make sure it works. Finally, put LVM on top, and check that works.

11,650

student

Updated on September 18, 2022

Comments

student over 1 year

I tried to setup TRIM with LVM and dm-crypt on ubuntu 13.04 following this tutorial:

http://blog.neutrino.es/2013/howto-properly-activate-trim-for-your-ssd-on-linux-fstrim-lvm-and-dmcrypt/

See the notes about my configuration and my testing procedure below.

Questions

Is there a reliable test if TRIM works properly?
Is my test routine wrong or is my TRIM not working?
If it's not working: what is wrong with my setup?
How can I debug TRIM for my setup and make TRIM work?

Configuration

Here ist my configuration:

cat /etc/crypttab

sda3_crypt UUID=[...] none luks,discard

and

cat /etc/lvm/lvm.conf

# [...]
devices  {
      # [ ... ]
      issue_discards = 1
      # [ ... ]
   }
# [...]

The SSD is a Samsung 840 Pro.

Here is my test-procedure

To test the setup I just did sudo fstrim -v / which resulted in

/: [...] bytes were trimmed

Doing this again resulted in /: 0 bytes were trimmed which seems to make sense and indicated that TRIM seems to work.

However then I did this test:

dd if=/dev/urandom of=tempfile count=100 bs=512k oflag=direct

sudo hdparm --fibmap tempfile                                 

tempfile:
 filesystem blocksize 4096, begins at LBA 0; assuming 512 byte sectors.
 byte_offset  begin_LBA    end_LBA    sectors
           0    5520384    5521407       1024
      524288    5528576    5529599       1024
     1048576    5523456    5525503       2048
     2097152    5607424    5619711      12288
     8388608    5570560    5603327      32768
    25165824    5963776    5980159      16384
    33554432    6012928    6029311      16384
    41943040    6275072    6291455      16384
    50331648    6635520    6639615       4096

sync

sudo hdparm --read-sector 5520384 /dev/sda                    

/dev/sda:
reading sector 5520384: succeeded
7746 4e11 bf42 0c93 25d3 2825 19fd 8eda
bd93 8ec6 9942 bb98 ed55 87eb 53e1 01d5
c61a 3f52 19a1 0ae5 0798 c6e2 39d9 771a
b89f 3fc5 e786 9b1d 3452 d5d7 9479 a80d
114a 7528 a79f f475 57dc aeaf 25f4 998c
3dd5 b44d 23bf 77f3 0ad9 8688 6518 28ee
81db 1473 08b5 befe 8f2e 5b86 c84e c7d2
1bdd 1065 6a23 fd0f 2951 d879 e823 021b
fa84 b9c1 eadd 9154 c9f4 2ebe cd70 64ec
75a8 4d93 c8fa 3174 7277 1ffb e858 5eca
7586 8b2e 9dbc ab12 40ab eb17 8187 e67d
5e0d 0005 5867 b924 5cfd 6723 9e4a 6f5f
99a4 a3b0 eeac 454a 83b6 c528 1106 6682
ca77 4edf 2180 bf0c b175 fabb 3d4b 37e2
b834 9e3e 82f2 2fdd 2c6a c6ca 873f e71e
f979 160f 5778 356f 2aea 6176 46b6 72b9
f76e ee51 979c 326b 1436 7cfe f677 bfcd
4c3c 9e11 4747 45c1 4bb2 4137 03a1 e4c8
e9dd 43b4 a3b4 ce1b d218 4161 bf64 727b
75d8 dcc2 e14c ebec 2126 25da 0300 12bd
6b1a 28b3 824f 3911 c960 527d 97cd de1b
9f08 9a8e dcdc e65f 1875 58ca be65 82bf
e844 50b8 cc1b 7466 58b8 e708 bd3d c01f
64fb 9317 a77a e43b 671f e1fb e328 93a9
c9c7 291c 56e0 c6c1 f011 b94d 9dc7 71e6
c8b1 5720 b8c9 b1a6 14f1 7299 9122 912b
312a 0f2f a31a 8bf9 9f8c 54e6 96f3 60b8
04a7 7dc9 3caa db0a a837 e5d7 2752 b477
c22d 7598 44e1 84e9 25d4 5db5 9f19 f73b
85a0 c656 373a ec34 55fb e1fc 124e 4674
1ba8 1a84 6aa4 7cb5 455e f416 adc6 a125
c4d4 8323 4eee 2493 2920 4e38 524c 1981

sudo rm tempfile

sync

sudo fstrim /

sync

sudo hdparm --read-sector 5520384 /dev/sda

/dev/sda:
reading sector 5520384: succeeded
7746 4e11 bf42 0c93 25d3 2825 19fd 8eda
bd93 8ec6 9942 bb98 ed55 87eb 53e1 01d5
c61a 3f52 19a1 0ae5 0798 c6e2 39d9 771a
b89f 3fc5 e786 9b1d 3452 d5d7 9479 a80d
114a 7528 a79f f475 57dc aeaf 25f4 998c
3dd5 b44d 23bf 77f3 0ad9 8688 6518 28ee
81db 1473 08b5 befe 8f2e 5b86 c84e c7d2
1bdd 1065 6a23 fd0f 2951 d879 e823 021b
fa84 b9c1 eadd 9154 c9f4 2ebe cd70 64ec
75a8 4d93 c8fa 3174 7277 1ffb e858 5eca
7586 8b2e 9dbc ab12 40ab eb17 8187 e67d
5e0d 0005 5867 b924 5cfd 6723 9e4a 6f5f
99a4 a3b0 eeac 454a 83b6 c528 1106 6682
ca77 4edf 2180 bf0c b175 fabb 3d4b 37e2
b834 9e3e 82f2 2fdd 2c6a c6ca 873f e71e
f979 160f 5778 356f 2aea 6176 46b6 72b9
f76e ee51 979c 326b 1436 7cfe f677 bfcd
4c3c 9e11 4747 45c1 4bb2 4137 03a1 e4c8
e9dd 43b4 a3b4 ce1b d218 4161 bf64 727b
75d8 dcc2 e14c ebec 2126 25da 0300 12bd
6b1a 28b3 824f 3911 c960 527d 97cd de1b
9f08 9a8e dcdc e65f 1875 58ca be65 82bf
e844 50b8 cc1b 7466 58b8 e708 bd3d c01f
64fb 9317 a77a e43b 671f e1fb e328 93a9
c9c7 291c 56e0 c6c1 f011 b94d 9dc7 71e6
c8b1 5720 b8c9 b1a6 14f1 7299 9122 912b
312a 0f2f a31a 8bf9 9f8c 54e6 96f3 60b8
04a7 7dc9 3caa db0a a837 e5d7 2752 b477
c22d 7598 44e1 84e9 25d4 5db5 9f19 f73b
85a0 c656 373a ec34 55fb e1fc 124e 4674
1ba8 1a84 6aa4 7cb5 455e f416 adc6 a125
c4d4 8323 4eee 2493 2920 4e38 524c 1981

This seems to indicate that TRIM doesn't work. Since

sudo hdparm -I /dev/sda | grep -i TRIM                        
       *    Data Set Management TRIM supported (limit 8 blocks)
       *    Deterministic read ZEROs after TRIM

Edit

Here is the output of sudo dmsetup table

lubuntu--vg-root: 0 465903616 linear 252:0 2048
lubuntu--vg-swap_1: 0 33308672 linear 252:0 465905664
sda3_crypt: 0 499222528 crypt aes-xts-plain64 00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 0 8:3 4096 1 allow_discards

Here is my /etc/fstab:

# <file system> <mount point>   <type>  <options>       <dump>  <pass>
/dev/mapper/lubuntu--vg-root /               ext4    errors=remount-ro 0       1
# /boot was on /dev/sda2 during installation
UUID=f700d855-96d0-495e-a480-81f52b965bda /boot           ext2    defaults        0       2
# /boot/efi was on /dev/sda1 during installation
UUID=2296-2E49  /boot/efi       vfat    defaults        0       1
/dev/mapper/lubuntu--vg-swap_1 none            swap    sw              0       0
# tmp
tmpfs /tmp tmpfs nodev,nosuid,noexec,mode=1777          0       0

Edit:

I finally reported it as a bug in https://bugs.launchpad.net/ubuntu/+source/lvm2/+bug/1213631

Hope somebody will find a solution there or at least test the setup and verify the bug.

Update

Now it works, see accepted answer.

student almost 11 years

Sorry, this was a typo. I have issue_discards = 1 in my config file.
Razzlero almost 11 years

If I were you I would try to use an iSCSI target and test this via tcpdump/wireshark to see if the setup works, though I do not know if the Linux iSCSI target supports trim or not. I do believe that dm-crypt should not blank the blocks on the physical disk because that makes it easier to ignore the free space on the device when trying to brute-force it (I don't know if it does that or not, though). In addition, SSDs are not required to return zeros after blanking, since the wear-leveling can redirect the read to a different block than the one blanked.
frostschutz over 10 years

According to bugzilla.redhat.com/show_bug.cgi?id=958096 I misunderstood the issue_discards = 1.

derobert almost 11 years

@student OK, that's the wrong sector then (the first two paragraphs of my answer). I'll edit my answer to remove that sentence about sector 6575104, as it isn't relevant any more.
student almost 11 years

Thanks! Instead of rm trim.test; sync I did `rm trim.test; sync; sudo fstrim -v .; However the pattern y Pattern persists. So TRIM seems not to work. Do you think my config is correct?
student almost 11 years

I am not sure which device I should take for dmsetup. I just did: sudo dmsetup table /dev/mapper/lubuntu--vg-root which gives 0 465903616 linear 252:0 2048
derobert almost 11 years

@student That means sector 0 is at sector 2048 on device 252:0. You'll have to figure out what 252:0 is, I'd guess its your dm crypto device (that's the major and minor number, will show up in /dev for example). And you'll need to look at the table for that device, to continue chasing it down to a block on an underlying device.
frostschutz almost 11 years

You've successfully tested that TRIM doesn't work, then... ;) as for your setup, you haven't explained it in great detail (filesystem, mount options, does dmsetup table show allow_discards for both LVM and crypt, anything else in the mix like raid...?).
student almost 11 years

See my edit. I do not use raid. Do you need more info?
frostschutz over 10 years

@student: I feel bad for not noticing this earlier, edited the answer to drop caches before hexdump.
student over 10 years

Thanks, that was the missing point. Now it seems to work!
desgua about 10 years

@frostschutz Thank you for this great solution. I made a script to automate the process if some lazy person come here.
Marc.2377 almost 7 years

Newcomers, do note that the TRIM command won't always "zero-fill" the blocks right away. See here, here and here. Although it should, in OP's case, since his hdparm -I result indicates "Deterministic read ZEROs after TRIM".
frostschutz almost 7 years

@Marc.2377 Maybe true, but it worked for every SSD I tested so far, which makes it difficult for me to update or change the answer. Some people got old data because they didn't drop caches. If it doesn't work for you, maybe try a larger file (like 128MiB instead of 1MiB), but that's just guessing, I can't verify what would need to be done instead for SSD models that somehow insist on returning discarded data.
Marc.2377 almost 7 years

@frostschutz Thank you for considering updating your answer, I wasn't expecting that really. Well, my ssd in particular (the "SV300S37A/120G") doesn't support discard_zeroes_data. I tried with a 250MiB file and even restarted the computer. I don't even have LVM set up, only dm-crypt. It's an old model, SATA rev.3.0, TRIM works but does not zero the blocks.
frostschutz almost 7 years

@Marc.2377 dm-crypt with allow_discards? Check with dmsetup table should say luks: ... crypt aes-xts-plain64 ... ... 1 allow_discards otherwise it can't work. I don't have this model but another that doesn't support zeroes data either but yet this test still works, original data is gone after trim.
frostschutz almost 7 years

@Marc.2377 can't reproduce with any SSD I own - would be nice to get a similar gist for a model that doesn't work. Guess I got lucky, not sure if I want a model that returns original data after TRIM. I don't like losing data to TRIM either but that's a different problem (too much trim in linux in general).
Marc.2377 almost 7 years

@frostschutz I don't like it a lot either, but... it's encrypted so no reason to bother much. allow_discards is definitely enabled (result from dmsetup table is like yours, sans the preceeding luks part for some reason. I'll generate a gist soon and update you then.