Re: Re: BUG in kernel: Wrong Handling of USB HDD’s in scsiglue(slave_configure) and scsi/sd(sd_read_cache_type)

[Alan Stern]

On Fri, 30 Sep 2011, Amit Sahrawat wrote:

> Adding linux-usb - to get more insight's into the problem.
> 
> On Fri, Sep 30, 2011 at 11:23 PM, Amit Sahrawat
> <amit.sahrawat83@xxxxxxxxx> wrote:
> > On Fri, Sep 30, 2011 at 5:48 PM, James Bottomley
> > <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> wrote:
> >> On Fri, 2011-09-30 at 12:26 +0530, Amit Sahrawat wrote:
> >>> Now, for the USB HDD which do have write cache - sginfo is showing
> >>> them to Write Cache Enabled as false.
> >>> Why do the result of hdparm identification and sginfo varies- (I know
> >>> they have different interface to work with and hdparm takes care of
> >>> that by using SG_IO interface from it's code)? hdparm showed me
> >>> correct results - that lead me to digging in the kernel code and
> >>> checking the performance for USB HDD with Write cache enabled/disabled
> >>> - which also showed that QUEUE ordering chosen for USB HDD is not
> >>> correct.
> >>
> >> Well, what all this means is the SATL in the USB device is implemented
> >> wrongly.  Since USB devices only preset SCSI interfaces, that's what we
> >> have to believe.
> >>
> >> hdparm when used correctly sends an ATA inquiry command wrapped in an
> >> ATA_12 or ATA_16 SCSI command.  A large number of legacy SATLs are known
> >> to crash on these commands.
> >>
> >> Are you sure the ATA command is reporting correctly?  A write back cache
> >> is a remarkably silly thing to enable for a USB device because they're
> >> highly likely to be surprise ejected which powers the device down.

A usbmon trace would be very helpful for understanding this.  
Instructions can be found in the kernel source file 
Documentation/usb/usbmon.txt.  Post a trace showing what happens when 
the drive is first plugged in.

> > In addition to the problem reported - there is one more thing I have
> > noticed with USB HDD - they should be shown as 'removable' but the
> > removable is marked only for USB PEN Drives. This seems to be a bit of
> > confusing, any mass storage media connected on USB port should be
> > recognized as removable.

That is not true.  The word "removable" refers to the storage media.  
Thus, a cdrom drive is removable because you can remove the disc from
the drive.  The same for a card reader, because you can remove the card
from the reader.  But a flash drive is not removable, because you can't
take the flash memory chip out of the device.

You are confusing "removable" with "hot-unpluggable".  All USB drives
are hot-unpluggable, but relatively few of them are removable.  Vendors 
often get this wrong, however.

> > So, for handling the issue, I would consider adding the handling in
> > slave_configure()(usb/storage/scsiglue) which marks the HDD/pen drives
> > as removable also signifying them to be USB based.

This does not happen in usb-storage at all.  The SCSI core is
responsible for determining whether or not a device is removable, and
it does this by looking at second byte of the data returned by the SCSI
INQUIRY command.

> > Then, as part of sd_revalidation – how about adding the ATA_IDENTIFY
> > command part if the device is USB HDD? As far as the result of
> > ATA_IDENTIFY is concerned – they return proper ‘256’ bytes - response
> > and the Words – 82, 85 used for feature supported and enabled/disabled
> > returns proper values for the USB HDD’s I have seen. In case of USB
> > pen drives – they return failure – I did not see any crash – maybe I
> > don’t have one of the legacy SATL based disk.
> > Since, I am new to this – I will check more on this to get a viable
> > solution. Please add your opinion. Can you share the name of the
> > device which causes crash with these ATA commands, If I am able to get
> > one I can try on that also.
> >
> >>> I have a large number of USB HDD's - with different vendors, and for
> >>> all of them - it is showing Write Cache Enabled as false.

This indicates those drives do not implement the SCSI MODE SENSE 
command correctly.  usbmon will help track down the problem.

> >>> The code works only for the Pen Drives or the USB HDD which do not
> >>> have internal cache.
> >>>
> >>> Also, for journalling filesystem being used on USB HDD - it does
> >>> becomes a cause of concern.
> >>>
> >>> Please share your opinion, I guess we need a change for mode sensing
> >>> in the kernel code for USB HDD.
> >>
> >> Well that's a nastily complex problem.  It really needs to be
> >> whitelisted in the USB stack, but if every drive is doing it, that's
> >> quite a task.
> >>
> >> The question becomes how do we detect in a SCSI fashion that the device
> >> has a write back cache if none of the standard SCSI mechanisms reports
> >> it?
> > As far as detecting in SCSI fashion – I wonder using that I would have
> > never reached the conclusion that it is the Write Cache of USB HDD
> > which is causing problem instead I would have been focusing on
> > particular file system (XFS in my case –which in itself is complex) –
> > there BARRIER support and also the Queue handling in the elevator with
> > I/O scheduler.
> > None of the sg utils is showing anything related with the Write Cache
> > in USB HDD – which provide any hint that the Cache is enabled – this
> > is a bit surprising because most of the high end USB mass storages
> > device in the market have Write Cache in them.

Alan Stern

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/