cyberax / mdtrim Goto Github PK

Memory use gets really high when trimming RAID with two SSD drives and a lot of free space. Python also crashed with some internal double free errors when trying to trim using a file as large as 40 GB. I have successfully tested the program up to 30 GB temporary file sizes and it needed about 4 GB of RAM.

I was able to work around this by removing temporary file deletion from the code and using several ~10 GB files sequentially until I run out of disk space. Then I just removed all the trimmed temporary files at the end. This seemed to restore my write speeds to original performance level.

Additional benefit of using several smaller files is that the disk is almost full for shorter amount of time as checking of the last piece takes less time now.

it seems that, on 32bit systems, libc.fallocate takes 32bit arguments

Hi Cyberax,

I'm having an issue with the script, I doubt it requires any changes to it, but this is really the only way I could find to contact you.

I just purchased two of the crucial M4 ssds, and set each partition up in a in a raid 1 with MDADM. As far as I can tell, I did everything right, everything is working, except trim. When I use the hdparm fibmap & read-sector test, I do not see zeros, as I did with my old vertex2.

Looking around online, it appears that this is caused by no support for discard in the raid1 module. There's the thread on LWN about a patch, which you have probably read, but it's not committed yet, and it's probably going to be some time before it is.

I bet if I failed the raid and just operated on each drive one at a time, discard would work. I was going to go about figuring out how to do that, but then I found your script, which looks like a perfectly fine work around until the kernel supports it. I did this with wiper.sh for a few years.. I tried the script, but I get stuck at:

print "Checking misalignment signatures"
for sl in slaves.values():
    slave_offset = sl["offset"]*sector_size
    for rng in trim_ranges:
        for test in rng["test_data"].values():
# snip...
            sector_data = data + ' '*padding + data

            what_we_read = string_at(read_buf+read_buf_offset, sector_size)
            if what_we_read != sector_data:
                print "Data written into the file and data on the physical HD do not match!"
                sys.exit(2)

what_we_read is: empty string
sector_data is something like:

Zb45aa627634f4682950f49b4e9e9f8a0


 b45aa627634f4682950f49b4e9e9f8a0A

Any ideas why that would happen and what I can try next to fix it? Is it a hole in the script, or is something horribly wrong with my setup, or even a hardware problem?

Thanks