On 10/28/21 1:33 PM, James Bottomley wrote:
On Thu, 2021-10-28 at 13:21 -0700, Bart Van Assche wrote:
[...]
Hi James,
The help with trying to find a solution is appreciated.
One of the software developers who is familiar with HPB explained to
me that READ BUFFER and WRITE BUFFER commands may be received in an
arbitrary order by UFS devices. The UFS HPB spec requires UFS devices
to be able to stash up to 128 such pairs. I'm concerned that leaving
out WRITE BUFFER commands only will break the HPB protocol in a
subtle way.
Based on the publicly available information (the hotstorage paper) I
don't belive it can. The Samsung guys also appear to confirm that the
use of WRITE BUFFER is simply an optimzation for large requests:
https://lore.kernel.org/all/20211025051654epcms2p36b259d237eb2b8b885210148118c5d3f@epcms2p3/
As did the excerpt from the spec you posted. It will cause slowdowns
for reads of > 32kb, because they have to go through the native FTL
lookup now, but there shouldn't be any functional change. Unless
there's anything else in the proprietary spec that contradicts this?
Are the UFS standards really proprietary? On https://www.t10.org/pubs.htm
I found the following: "Approved American National Standards and Technical
Reports may be purchased from: [ ... ]". And on
https://www.jedec.org/document_search?search_api_views_fulltext=hpb I found
the following: "Available for purchase: $80.00". It seems to me that SCSI
and JEDEC standards are available under similar conditions?
Regarding the question about the impact of leaving out WRITE BUFFER
commands on the HPB protocol, I hope that one of the HPB experts will be so
kind to answer that question.
Bart.
