James,
Am 10.11.2019 um 09:12 schrieb James Bottomley:
On Sat, 2019-11-09 at 20:14 +0100, Kars de Jong wrote:
When using this driver on a Blizzard 1260, there were failures
whenever DMA transfers from the SCSI bus to memory of 65535 bytes
were followed by a DMA transfer of 1 byte. This caused the byte at
offset 65535 to be overwritten with 0xff. The Blizzard hardware can't
handle single byte DMA transfers.
Besides this issue, limiting the DMA length to something that is not
a multiple of the page size is very inefficient on most file systems.
It seems this limit was chosen because the DMA transfer counter of
the ESP by default is 16 bits wide, thus limiting the length to 65535
bytes. However, the value 0 means 65536 bytes, which is handled by
the ESP and the Blizzard just fine. It is also the default maximum
used by esp_scsi when drivers don't provide their own
dma_length_limit() function.
Have you tested this on any other hardware? the reason most legacy
hardware would have a setting like this is because they have a two byte
transfer length register and zero doesn't mean 65536. If this is the
The data book for the NCR53FC94/NCR53FC96 (the 'fast' SCSI controller
used on the board Kars tried) states that with the features enable bit
clear (no 24 bit DMA counts used), zero does mean 64k indeed. The
features enable bit is never set by esp_scsi.c. I chose the incorrect
length limit without realizing this special case for the transfer count.
and before we found out that 1-byte DMA just won't work.
I need to confirm this from a data book of the older (pre-fast)
revisions of this chip family. but since as Kars also states, the core
driver default for the 16 bit transfer size is 64k as well, I very much
suspect the same behaviour for the older revisions.
All of the old board-specific drivers used a max transfer length of
0x1000000, only the fastlane driver used 0xfffc. That lower limit might
be due to a DMA limitation on the fastlane board. We could accommodate
the different limit for this board by using a board-specific
dma_length_limit() callback...
case for any of the cards the zorro_esp drives, it might be better to
lower the max length to 61440 (64k-4k) so the residual is a page.
For the benefit of keeping the code simple, and avoid retesting the
fastlane board, that might indeed be the better solution.
Cheers,
Michael
James
