Message-ID: <alpine.LNX.2.00.1707011008060.4989@nippy.intranet>
Date:   Sat, 1 Jul 2017 12:40:18 +1000 (AEST)
From:   Finn Thain <fthain@...egraphics.com.au>
To:     Ondrej Zary <linux@...nbow-software.org>
cc:     "James E.J. Bottomley" <jejb@...ux.vnet.ibm.com>,
        "Martin K. Petersen" <martin.petersen@...cle.com>,
        linux-scsi@...r.kernel.org, linux-kernel@...r.kernel.org,
        Michael Schmitz <schmitzmic@...il.com>
Subject: Re: [PATCH v5 0/6] g_NCR5380: PDMA fixes and cleanup

On Fri, 30 Jun 2017, Ondrej Zary wrote:

> > > No more log spamming on DTC but reads are corrupted even more than 
> > > before. The IRQ check after data transfer increases the chance of 
> > > catching an IRQ before the buffer could become ready.
> >
> > If we delay the IRQ check, that just means that CSR_GATED_53C80_IRQ 
> > will be detected a bit later (128 bytes later)... so not much 
> > difference.
> 
> The main difference is that my original patch read the CSR once and 
> then:
> - transfer data if the buffer is ready, ignoring any IRQ
> - if the buffer is not ready, check for an IRQ
> - if neither buffer is ready nor IRQ was asserted, check for timeout
> 

I think your pseudocode is equivalent to this code* from the patch,

	if (NCR5380_poll_politely2(hostdata, hostdata->c400_ctl_status,
	                           CSR_HOST_BUF_NOT_RDY, 0,
	                           hostdata->c400_ctl_status,
	                           CSR_GATED_53C80_IRQ,
	                           CSR_GATED_53C80_IRQ, HZ / 64) < 0)
		break;

	if (NCR5380_read(hostdata->c400_ctl_status) &
	    CSR_HOST_BUF_NOT_RDY)
		break;

	/* transfer data */

If this doesn't work on DTC devices, I think we have to special case them 
somehow.

> So yes, the IRQ will be detected 128 bytes later - but the IRQ is 
> asserted at 3968 bytes which means the transfer will then be complete.
> 

Yes, unless the IRQ was asserted 128 bytes into the transfer...

> > > This patch:
> > > --- a/drivers/scsi/g_NCR5380.c
> > > +++ b/drivers/scsi/g_NCR5380.c
> > > @@ -548,8 +548,10 @@ static inline int generic_NCR5380_precv(struct
> > >  		start += 128;
> > >
> > >  		if (NCR5380_read(hostdata->c400_ctl_status) &
> > > -		    CSR_GATED_53C80_IRQ)
> > > +		    CSR_GATED_53C80_IRQ) {
> > > +			printk("r irq at start=%d basr=0x%02x\n", start,
> > > NCR5380_read(BUS_AND_STATUS_REG)); break;
> > > +		}
> > >  	}
> > >
> > >  	residual = len - start;
> > >
> > > produces lots of these lines:
> > > [  896.194054] r irq at start=128 basr=0x98
> > > [  896.197758] r irq at start=3968 basr=0x98
> >
> > Assuming that the registers are available and valid, the value 0x98 
> > means BASR_END_DMA_TRANSFER | BASR_IRQ | BASR_PHASE_MATCH. There is no 
> > BASR_BUSY_ERROR here, so the cause of the CSR_GATED_53C80_IRQ must be 
> > that the 53c400 has terminated the transfer by asserting /EOP. That 
> > shouldn't happen before before the counters run down.
> >
> > It doesn't make sense. So maybe the 53c80 registers are not valid at 
> > this point? That means a phase mismatch can't be excluded... unlikely 
> > at 128 bytes into the transfer. Busy error? Also unlikely.
> >
> > I have to conclude that CSR_GATED_53C80_IRQ and BASR_END_DMA_TRANSFER 
> > can't be trusted on this board. I guess that's why you examine the 
> > BASR directly in your original algorithm but ignore 
> > BASR_END_DMA_TRANSFER.
> 
> Exactly, BASR_END_DMA_TRANSFER causes problems on DTC.
> 
> > It does look like some kind of timing issue: the "start" value above 
> > changes from one log message to the next. Who knows?
> 
> Only that two values (128 and 3968) appeared. 3968 = 4096-128, one block 
> before the end of transfer. So probably BASR_END_DMA_TRANSFER is 
> asserted one block earlier (i.e. when the last block of data is 
> transferred from the drive to the chip buffer). Haven't seen this 
> problem at start=128 before.
> 

Well, if we ignore the 128 B short transfer as an aberrant outlier, and if 
we assume that BASR_END_DMA_TRANSFER can be made to work correctly, then 
perhaps we can tweak the timing so that a 4096 byte transfer is not cut 
short.

Alternatively, perhaps we could workaround the BASR_END_DMA_TRANSFER 
issue with a special case that ignores Gated IRQ when
	hostdata->board == BOARD_DTC3181E && start == len - 128

> > > This fixes the DTC read corruption, although I don't like the repeated
> > > ctl_status register reads:
> > > --- a/drivers/scsi/g_NCR5380.c
> > > +++ b/drivers/scsi/g_NCR5380.c
> > > @@ -533,7 +533,7 @@ static inline int generic_NCR5380_precv(struct
> > >  			break;
> > >
> > >  		if (NCR5380_read(hostdata->c400_ctl_status) &
> > > -		    CSR_HOST_BUF_NOT_RDY)
> > > +		    CSR_GATED_53C80_IRQ && (NCR5380_read(hostdata->c400_ctl_status) &
> > > CSR_HOST_BUF_NOT_RDY)) break;
> > >
> > >  		if (hostdata->io_port && hostdata->io_width == 2)
> >
> > But that means the transfer will continue even when 
> > CSR_HOST_BUF_NOT_RDY. Your original algorithm doesn't attempt that. 
> > Neither does the algorithm in the datasheet. We should try to omit 
> > this change.
> 
> Sorry for that, it's a bug. I got lost in the conditions.
> 
> > > @@ -546,10 +546,6 @@ static inline int generic_NCR5380_precv(struct
> > >  		memcpy_fromio(dst + start,
> > >  			hostdata->io + NCR53C400_host_buffer, 128);
> > >  		start += 128;
> > > -
> > > -		if (NCR5380_read(hostdata->c400_ctl_status) &
> > > -		    CSR_GATED_53C80_IRQ)
> > > -			break;
> > >  	}
> > >
> > >  	residual = len - start;
> >
> > I think we should keep the CSR_GATED_53C80_IRQ check for the other 
> > boards, if this bogus BASR_END_DMA_TRANSFER problem is confined to 
> > DTC436.
> >
> > How about this change? (to be applied on top of 6/6)
> 
> It does not apply. Changed while() {} -> do {} while() manually.
> 

Sorry, I sent you the wrong diff again.

> > diff --git a/drivers/scsi/g_NCR5380.c b/drivers/scsi/g_NCR5380.c
> > index 3948f522b4e1..8e80379cfaaa 100644
> > --- a/drivers/scsi/g_NCR5380.c
> > +++ b/drivers/scsi/g_NCR5380.c
> > @@ -525,16 +525,22 @@ static inline int generic_NCR5380_precv(struct
> > NCR5380_hostdata *hostdata, NCR5380_write(hostdata->c400_blk_cnt, len /
> > 128);
> >
> >  	do {
> > -		if (NCR5380_poll_politely2(hostdata, hostdata->c400_ctl_status,
> > -		                           CSR_HOST_BUF_NOT_RDY, 0,
> > -		                           hostdata->c400_ctl_status,
> > -		                           CSR_GATED_53C80_IRQ,
> > -		                           CSR_GATED_53C80_IRQ, HZ / 64) < 0)
> > -			break;
> > -
> > -		if (NCR5380_read(hostdata->c400_ctl_status) &
> > -		    CSR_HOST_BUF_NOT_RDY)
> > -			break;
> > +		if (hostdata->board == BOARD_DTC3181E) {
> > +			/* Ignore bogus CSR_GATED_53C80_IRQ */
> > +			if (NCR5380_poll_politely(hostdata, hostdata->c400_ctl_status,
> > +			                          CSR_HOST_BUF_NOT_RDY, 0, HZ / 64) < 0)
> > +				break;
> > +		} else {
> > +			if (NCR5380_poll_politely2(hostdata, hostdata->c400_ctl_status,
> > +			                           CSR_HOST_BUF_NOT_RDY, 0,
> > +			                           hostdata->c400_ctl_status,
> > +			                           CSR_GATED_53C80_IRQ,
> > +			                           CSR_GATED_53C80_IRQ, HZ / 64) < 0)
> > +				break;
> > +			if (NCR5380_read(hostdata->c400_ctl_status) &
> > +			    CSR_HOST_BUF_NOT_RDY)
> > +				break;
> > +		}
> >
> >  		if (hostdata->io_port && hostdata->io_width == 2)
> >  			insw(hostdata->io_port + hostdata->c400_host_buf,
> 
> It works but it looks really ugly (too many ifs). It also means that we 
> detect disconnects only by waiting for timeout.
> 

You're right about disconnect detection. When I wrote this code, I 
reasoned that the timeout may slow down CD-ROM reads (because they often 
disconnect) but disconnect/reconnect is slow anyway. The impact would need 
to be measured. Anyway, there are a couple of alternatives:

1) Fix the timing on DTC so that Gated IRQ can be polled. That would allow 
   us to use the same algorithm for all boards (i.e. the code snippet 
   marked * above).
 
2) Poll BASR even if that register may be inaccessible (no CSR_53C80_REG 
   flag). As I said, I'm reluctant to attempt this.

3) The test for hostdata->board == BOARD_DTC3181E && start == len - 128 
   should avoid this issue, since CD-ROMs disconnect on 2048-byte 
   boundaries. I'll try this in v6.

> > @@ -546,10 +552,6 @@ static inline int generic_NCR5380_precv(struct
> > NCR5380_hostdata *hostdata, memcpy_fromio(dst + start,
> >  				hostdata->io + NCR53C400_host_buffer, 128);
> >  		start += 128;
> > -
> > -		if (NCR5380_read(hostdata->c400_ctl_status) &
> > -		    CSR_GATED_53C80_IRQ)
> > -			break;
> >  	} while (start < len);
> >
> >  	residual = len - start;
> > @@ -600,6 +602,12 @@ static inline int generic_NCR5380_psend(struct
> > NCR5380_hostdata *hostdata, break;
> >
> >  		if (NCR5380_read(hostdata->c400_ctl_status) &
> > +		    CSR_HOST_BUF_NOT_RDY && start > 0) {
> > +			start -= 128;
> > +			break;
> > +		}
> > +
> > +		if (NCR5380_read(hostdata->c400_ctl_status) &
> >  		    CSR_GATED_53C80_IRQ)
> >  			break;
> >
> 
> This works too, although I'm not sure if the conditions are 100% 
> correct. I'm getting lost again. That means that we roll back 128 bytes 
> & "break" when the buffer is not ready and we already wrote something. 
> We also "break" when IRQ arrives. But what if the buffer is not ready 
> and start == 0?
> 

If start == 0 the residual ends up being the entire transfer. I think it 
is correct. I will see whether this logic can be expressed more clearly.

> > @@ -615,8 +623,7 @@ static inline int generic_NCR5380_psend(struct
> > NCR5380_hostdata *hostdata, start += 128;
> >  	} while (start < len);
> >
> > -	residual = max(len - start,
> > -	               (int)NCR5380_read(hostdata->c400_blk_cnt) * 128);
> > +	residual = len - start;
> >
> >  	if (residual != 0) {
> >  		/* 53c80 interrupt or transfer timeout. Reset 53c400 logic. */
> 

Thanks for testing.

-- 

Powered by blists - more mailing lists