Skip to contentSkip to conversation

USB storage & write barrier support?

DISCUSSION Ric Wheeler 2 participants 5 messages Nov 23, 2009 VFS

lore
Share & tools
lore
#0 lore raw

Hi Alan,

One quick question - how robust is our support for write barriers over USB?

Specifically, I am looking to get together some testing with both external USB and e-sata enclosure for various file systems and would be very interested in helping make sure that this is well handled...

Best regards,

Ric

#1 lore raw
Quoted 7 lines · Ric Wheeler

Hi Alan,

One quick question - how robust is our support for write barriers over USB?

Specifically, I am looking to get together some testing with both external USB and e-sata enclosure for various file systems and would be very interested in helping make sure that this is well handled...

It's extremely robust -- the USB mass-storage driver has a maximum command-queue length of 1!

With the coming of USB 3.0 and the UASP (USB-Attached SCSI Protocol) specification this will change. I presume barriers will then be implemented as the need arises.

Alan

#2 lore raw
Quoted 10 lines · Alan Stern
Quoted 7 lines · Ric Wheeler

Hi Alan,

One quick question - how robust is our support for write barriers over USB?

Specifically, I am looking to get together some testing with both external USB and e-sata enclosure for various file systems and would be very interested in helping make sure that this is well handled...

Quoted 9 lines

It's extremely robust -- the USB mass-storage driver has a maximum command-queue length of 1!

With the coming of USB 3.0 and the UASP (USB-Attached SCSI Protocol) specification this will change. I presume barriers will then be implemented as the need arises.

Alan

What we need is to pass down cache flush commands (ATA_CACHE_FLUSH_EXT is what flushed the cache for ATA/S-ATA devices). Even with a command-queue length of 1, the write cache a USB connected s-ata drive would still loose data on power off without this kind of support.

Ric

#3 lore raw
Quoted 4 lines · Ric Wheeler

What we need is to pass down cache flush commands (ATA_CACHE_FLUSH_EXT is what flushed the cache for ATA/S-ATA devices). Even with a command-queue length of 1, the write cache a USB connected s-ata drive would still loose data on power off without this kind of support.

You can't send ATA commands directly over a USB connection without using a passthru protocol of some sort. Some of the USB interfaces in various drive enclosures support such a protocol but a lot of them don't.

Of course, you _can_ send SCSI's SYNCHRONIZE CACHE command. What ATA/SATA command (if any) the drive's USB interface will translate it into is known only to the interface's manufacturer. Nevertheless, that's what we've been depending on because that's what the sd driver uses.

Alan

#4 lore raw
Quoted 12 lines · Alan Stern
Quoted 4 lines · Ric Wheeler

What we need is to pass down cache flush commands (ATA_CACHE_FLUSH_EXT is what flushed the cache for ATA/S-ATA devices). Even with a command-queue length of 1, the write cache a USB connected s-ata drive would still loose data on power off without this kind of support.

You can't send ATA commands directly over a USB connection without using a passthru protocol of some sort. Some of the USB interfaces in various drive enclosures support such a protocol but a lot of them don't.

Quoted 8 lines

Of course, you _can_ send SCSI's SYNCHRONIZE CACHE command. What ATA/SATA command (if any) the drive's USB interface will translate it into is known only to the interface's manufacturer. Nevertheless, that's what we've been depending on because that's what the sd driver uses.

Alan

That makes sense - so the usb storage stack does the right thing, what happens is up to the SCSI -> ATA translation in the target device.

If I still had occasional access to analyzers, we could always test & verify that some of the common enclosures work correctly but I have unfortunately lost my toys :-)

Thanks!

Ric