Re: Update on TRIM support

[Tim Bisson <bissont@xxxxxxx>]

Here's another update on Trim:

First, the code:
https://github.com/bissont/DragonFlyBSD


Updates:
* Support for slices (UFS/hammer):
** "newfs -E /dev/da0s1b" trims the right sectors

* ability to trim swap
** swapon -e /dev/da0s1b
** adjusted rcorder scripts so swapon starts after savecore
*** allows us to use the same device for savecore and swap since "swapon 
-e" will trim the device only after savecore has run.
***https://github.com/bissont/DragonFlyBSD/commit/dd76a0f84303304566f5c907e86fedf634ca8aa9
***http://pastie.org/2086097

* sysctl to control trim
** kern.cam.da.X.trim_enabled
** sysctl appears only if the device supports trim
** sysctl needs to be set in order to trim a new fs and/or enable online 
FS trim

* enable online trim in ufs with mount option "-o trim"


For example, one could use a device in the following manner:
$ sysctl kern.cam.da.0.trim_enabled=1    #enables trim for /dev/da0

$ newfs_hammer -L da0s1a -E /dev/da0s1a  #trim device slice a
$ mount_hammer /dev/da0s1a /mnt/test     #mount hammer slice

$ newfs -E /dev/da0s1b                   #trims the device slice b
$ mount -o trim /dev/da0s1b /mnt/test    #enable online trim for UFS
$ umount /mnt/test
$ mount  /dev/da0s1b /mnt/test           #online trim for UFS off

Feedback is always welcome.

Thanks,
Tim

On 6/1/11 11:12 PM, Tim Bisson wrote:
> Hi All,
>
> Here's an update on TRIM support based on feedback from Matt, Alex,
> Venkatesh, and others. Thanks guys.
>
> First, the code:
> https://github.com/bissont/DragonFlyBSD
>
> If you clone the code and do a:
> git difftool upstream/master..HEAD
> you can get a diff against the current master (as of today).
>
> Changes:
> * Moved the underlying TRIM code from the IDE driver to the AHCI/SCSI
> driver.
> * Made UFS block frees (ffs_blkfree()) execute TRIM command on blocks
> asynchronously instead of synchronously. Once the TRIM command is
> completed, we free the actual blocks in the freemap in the callback.
> This makes delete performance much faster as we don't have to wait for
> the TRIM command to finish before returning from the ffs_blkfree().
> * Added an option, -E to newfs_hammer, that will run TRIM on the devices
> making up the hammer file system before the hammer file system is created.
> * Modified the IOCTL to take in bytes rather than sectors. In the IOCTL,
> we  issue a BUF_CMD_FREEBLKS command using a struct buf, which has a
> size limit of 2^31 (b_bcount) bytes. Therefore, we potentially have to
> break up the IOCTL TRIM command into multiple requests if the original
> command is for more than 2^31 bytes in size.
>
>   From "man newfs":
> -R      Turn on the TRIM enable flag.  If enabled, and if the underlying
>                device supports the BIO_DELETE command, the file system
> will send
>                a delete request to the underlying device for each freed
> block.
>                The trim enable flag is typically set when the underlying
> device
>                uses flash-memory as the device can use the delete command to
>                pre-zero or at least avoid copying blocks that have been
> deleted.
>
> -E      Use TRIM to erase the file system before creating it.  If used
>                and the underlying device supports TRIM, the device's LBAs
> will
>                be erased using the TRIM command before the file system is
> cre-
>                ated.
>
>   From "man newfs_hammer"
> -E      TRIM Device(s) before FS creation. Should only be called if all
>                devices are SSDs that support TRIM.
>
>
> Any suggestions, ideas, or feedback is welcome.
>
> Thanks,
> Tim