Re: [PATCH 76/71] ncr5380: Enable PDMA for DTC chips

From: Finn Thain
Date: Sat Dec 05 2015 - 22:41:08 EST



On Fri, 4 Dec 2015, Ondrej Zary wrote:

> Add I/O register mapping for DTC chips and enable PDMA mode.
>
> These chips have 16-bit wide HOST BUFFER register (counter register at
> offset 0x0d increments by 2 on each HOST BUFFER read). Detect it
> automatically.
>
> Large PIO transfers crash at least the DTCT-436P chip (all reads result
> in 0xFF) so this patch actually makes it work.
>
> The chip also crashes when we bang the C400 host status register too
> heavily after PDMA write - a small udelay is needed.
>
> Tested on DTCT-436P and verified that it does not break 53C400A.
>
> Signed-off-by: Ondrej Zary <linux@xxxxxxxxxxxxxxxxxxxx>
> ---
> drivers/scsi/g_NCR5380.c | 59 ++++++++++++++++++++++++++++------------------
> drivers/scsi/g_NCR5380.h | 4 +++-
> 2 files changed, 39 insertions(+), 24 deletions(-)
>
> diff --git a/drivers/scsi/g_NCR5380.c b/drivers/scsi/g_NCR5380.c
> index 85da3c2..9816b81 100644
> --- a/drivers/scsi/g_NCR5380.c
> +++ b/drivers/scsi/g_NCR5380.c
> @@ -328,7 +328,7 @@ static int __init generic_NCR5380_detect(struct scsi_host_template *tpnt)
> ports = ncr_53c400a_ports;
> break;
> case BOARD_DTC3181E:
> - flags = FLAG_NO_PSEUDO_DMA;
> + flags = FLAG_NO_DMA_FIXUP;

Nice!

> ports = dtc_3181e_ports;
> break;
> }
> @@ -412,10 +412,12 @@ static int __init generic_NCR5380_detect(struct scsi_host_template *tpnt)
> hostdata->c400_blk_cnt = 1;
> hostdata->c400_host_buf = 4;
> }
> - if (overrides[current_override].board == BOARD_NCR53C400A) {
> + if (overrides[current_override].board == BOARD_NCR53C400A ||
> + overrides[current_override].board == BOARD_DTC3181E) {
> hostdata->c400_ctl_status = 9;
> hostdata->c400_blk_cnt = 10;
> hostdata->c400_host_buf = 8;
> + hostdata->c400_host_idx = 13;
> }
> #else
> instance->base = overrides[current_override].NCR5380_map_name;
> @@ -430,8 +432,20 @@ static int __init generic_NCR5380_detect(struct scsi_host_template *tpnt)
> if (NCR5380_init(instance, flags))
> goto out_unregister;
>
> +#ifndef SCSI_G_NCR5380_MEM
> + /* read initial value of index register */
> + i = NCR5380_read(hostdata->c400_host_idx);
> + /* read something from host buffer */
> + NCR5380_read(hostdata->c400_host_buf);
> + /* I/O width = index register increment */
> + hostdata->io_width = NCR5380_read(hostdata->c400_host_idx) - i;
> + if (hostdata->io_width < 0)
> + hostdata->io_width += 128;
> +#endif

Will the result depend on the initial state of the chip, such as
CSR_TRANS_DIR or CSR_HOST_BUF_NOT_RDY?

It is possible to generate an interrupt with the buffer read, which should
be masked at the chip.

This buffer read may cause the 53C400 control logic to signal the 53C80
core. I suppose it is unlikely to cause any signalling on the SCSI bus
unless the 53C80 happened to be in DMA mode.

The possibility that io_width == 0 is not handled; wouldn't this result
indicate that PDMA shouldn't be used?

io_width can be calculated without a conditional statement, which may be
easier to read.

Can we be confident that detection will fail for all devices that don't
support word-sized IO, to avoid a regression?

The patch seems to assume that no memory-mapped card needs word-sized IO
for PDMA. Can you confirm?

The previous version of this patch was simpler and more predictable. You
enabled word-size IO for DTC3181E which is testable. Does this version
benefit any other cards?

> +
> if (overrides[current_override].board == BOARD_NCR53C400 ||
> - overrides[current_override].board == BOARD_NCR53C400A)
> + overrides[current_override].board == BOARD_NCR53C400A ||
> + overrides[current_override].board == BOARD_DTC3181E)
> NCR5380_write(hostdata->c400_ctl_status, CSR_BASE);
>
> NCR5380_maybe_reset_bus(instance);
> @@ -558,11 +572,10 @@ static inline int NCR5380_pread(struct Scsi_Host *instance, unsigned char *dst,
> while (NCR5380_read(hostdata->c400_ctl_status) & CSR_HOST_BUF_NOT_RDY);
>
> #ifndef SCSI_G_NCR5380_MEM
> - {
> - int i;
> - for (i = 0; i < 128; i++)
> - dst[start + i] = NCR5380_read(hostdata->c400_host_buf);
> - }
> + if (hostdata->io_width == 2)
> + insw(instance->io_port + hostdata->c400_host_buf, dst + start, 64);
> + else
> + insb(instance->io_port + hostdata->c400_host_buf, dst + start, 128);
> #else
> /* implies SCSI_G_NCR5380_MEM */
> memcpy_fromio(dst + start,
> @@ -579,11 +592,10 @@ static inline int NCR5380_pread(struct Scsi_Host *instance, unsigned char *dst,
> }
>
> #ifndef SCSI_G_NCR5380_MEM
> - {
> - int i;
> - for (i = 0; i < 128; i++)
> - dst[start + i] = NCR5380_read(hostdata->c400_host_buf);
> - }
> + if (hostdata->io_width == 2)
> + insw(instance->io_port + hostdata->c400_host_buf, dst + start, 64);
> + else
> + insb(instance->io_port + hostdata->c400_host_buf, dst + start, 128);
> #else
> /* implies SCSI_G_NCR5380_MEM */
> memcpy_fromio(dst + start,
> @@ -642,10 +654,10 @@ static inline int NCR5380_pwrite(struct Scsi_Host *instance, unsigned char *src,
> while (NCR5380_read(hostdata->c400_ctl_status) & CSR_HOST_BUF_NOT_RDY)
> ; // FIXME - timeout
> #ifndef SCSI_G_NCR5380_MEM
> - {
> - for (i = 0; i < 128; i++)
> - NCR5380_write(hostdata->c400_host_buf, src[start + i]);
> - }
> + if (hostdata->io_width == 2)
> + outsw(instance->io_port + hostdata->c400_host_buf, src + start, 64);
> + else
> + outsb(instance->io_port + hostdata->c400_host_buf, src + start, 128);
> #else
> /* implies SCSI_G_NCR5380_MEM */
> memcpy_toio(hostdata->iomem + NCR53C400_host_buffer,
> @@ -657,12 +669,11 @@ static inline int NCR5380_pwrite(struct Scsi_Host *instance, unsigned char *src,
> if (blocks) {
> while (NCR5380_read(hostdata->c400_ctl_status) & CSR_HOST_BUF_NOT_RDY)
> ; // FIXME - no timeout
> -
> #ifndef SCSI_G_NCR5380_MEM
> - {
> - for (i = 0; i < 128; i++)
> - NCR5380_write(hostdata->c400_host_buf, src[start + i]);
> - }
> + if (hostdata->io_width == 2)
> + outsw(instance->io_port + hostdata->c400_host_buf, src + start, 64);
> + else
> + outsb(instance->io_port + hostdata->c400_host_buf, src + start, 128);
> #else
> /* implies SCSI_G_NCR5380_MEM */
> memcpy_toio(hostdata->iomem + NCR53C400_host_buffer,
> @@ -682,8 +693,10 @@ static inline int NCR5380_pwrite(struct Scsi_Host *instance, unsigned char *src,
> /* All documentation says to check for this. Maybe my hardware is too
> * fast. Waiting for it seems to work fine! KLL
> */
> - while (!(i = NCR5380_read(hostdata->c400_ctl_status) & CSR_GATED_53C80_IRQ))
> + while (!(i = NCR5380_read(hostdata->c400_ctl_status) & CSR_GATED_53C80_IRQ)) {
> + udelay(4); /* DTC436 chip hangs without this */
> ; // FIXME - no timeout
> + }

When you added the braces, the lone semicolon became redundant. But I
think the entire loop is bogus.

Why do we wait for CSR_GATED_53C80_IRQ? I can understand testing it during
a transfer but why afterwards? (The core driver could test for the IRQ
flag in the Bus and Status Register, at the end of a DMA, if this scsi
host has no IRQ line.)

The comments and the 53C400 datasheet say we should instead wait for
CSR_53C80_REG. I agree. The g_NCR5380 wrapper driver can't safely return
control to the core driver unless the 53C80 registers are available.

The algorithm in the datasheet waits for CSR_53C80_REG before checking
BASR_END_DMA_TRANSFER, checking interrupt flags, disabling DMA mode etc.

g_NCR5380.c has an '#if 0' around the BASR_END_DMA_TRANSFER check, because
it doesn't wait for CSR_53C80_REG. Does NCR's algorithm work with your
cards?

>
> /*
> * I know. i is certainly != 0 here but the loop is new. See previous
> diff --git a/drivers/scsi/g_NCR5380.h b/drivers/scsi/g_NCR5380.h
> index c5e57b7..b3936aa 100644
> --- a/drivers/scsi/g_NCR5380.h
> +++ b/drivers/scsi/g_NCR5380.h
> @@ -44,7 +44,9 @@
> #define NCR5380_implementation_fields \
> int c400_ctl_status; \
> int c400_blk_cnt; \
> - int c400_host_buf;
> + int c400_host_buf; \
> + int c400_host_idx; \
> + int io_width;
>
> #else
> /* therefore SCSI_G_NCR5380_MEM */
>

--
--
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/