[Server-devel] [XSCE] sdcard for /opt and /library on xo 1.5

[James Cameron]

On Wed, Jul 15, 2015 at 12:47:10AM -0400, Samuel Greenfeld wrote:
> Would periodically erasing SD cards and eMMC storage help to extend
> their useful life?

No, I don't think so.  Nor would it hurt their life.  It is just a few
writes to metadata mapping tables kept in a different physical area of
the device.  (The tables relate virtual block numbers used by the host
to physical pages used by the device.)

> Or is this potentially dangerous for chips known to be used which
> implement these instructions incorrectly?

Doesn't seem likely, but good point; there may exist a set of devices
that implement the command incorrectly.

Although the same could be said for other corner cases, like writing a
block in the last position of a page after the page has triggered the
read disturb rewrite handler.

Hopefully the test suites will have caught both scenarios.

> A few years ago, XO firmware would erase the SD card prior to
> running fs-update.  But if I recall correctly this was disabled
> because it caused certain SD cards to hang.

If I recall correctly it was because certain SD cards would refuse the
command; they didn't really hang, they just gave an error.

While investigating we found that the erasure was only covering a tiny
part of the device block numbers (our mistake), so it wasn't being
effective anyway.

We didn't want to increase the time taken to fs-update, so we removed
it.

We also switched from linear to sparse input files.

> It might also be possible to enable the eMMC equivalent of TRIM in
> XO software builds provided XO eMMC's don't accidentally discard the
> wrong block.

Yes.  TRIM is like ERASE except the device is not obligated to return
a block of zeros.

Both become a write to the mapping table.

Acting on the wrong block seems unlikely though.

A secure ERASE is different; in addition to a mapping table write they
also COPY the page less any unerased blocks and then do an erase page.

I guess writing zero to all blocks might not erase any page; it might
instead just write to the mapping table to map each block to a zero
page.

-- 
James Cameron
http://quozl.linux.org.au/