cascardo/linux.git
8 years agoMerge remote-tracking branch 'mkp-scsi/4.5/scsi-queue' into misc
James Bottomley [Thu, 14 Jan 2016 00:29:30 +0000 (16:29 -0800)]
Merge remote-tracking branch 'mkp-scsi/4.5/scsi-queue' into misc

8 years agoimm: Use new parport device model
Sudip Mukherjee [Tue, 5 Jan 2016 12:40:50 +0000 (18:10 +0530)]
imm: Use new parport device model

Modify imm driver to use the new parallel port device model.

Signed-off-by: Sudip Mukherjee <sudip@vectorindia.org>
Reviewed-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agomegaraid: Fix possible NULL pointer deference in mraid_mm_ioctl
Nicholas Krause [Tue, 5 Jan 2016 19:32:54 +0000 (14:32 -0500)]
megaraid: Fix possible NULL pointer deference in mraid_mm_ioctl

This adds the needed check after the call to the function
mraid_mm_alloc_kioc in order to make sure that this function has not
returned NULL and therefore makes sure we do not deference a NULL
pointer if one is returned by mraid_mm_alloc_kioc.  Further more add
needed comments explaining that this function call can return NULL if
the list head is empty for the pointer passed in order to allow furture
users to understand this required pointer check.

Signed-off-by: Nicholas Krause <xerofoify@gmail.com>
Acked-by: Sumit Saxena <sumit.saxena@avagotech.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agostorvsc: Fix typo in MODULE_PARM_DESC
Dan Carpenter [Fri, 8 Jan 2016 11:02:04 +0000 (14:02 +0300)]
storvsc: Fix typo in MODULE_PARM_DESC

The module_param is "storvsc_vcpus_per_sub_channel" so we need to use
that for MODULE_PARM_DESC() as well.

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agocxgbi: Typo in MODULE_PARM_DESC
Dan Carpenter [Fri, 8 Jan 2016 11:01:36 +0000 (14:01 +0300)]
cxgbi: Typo in MODULE_PARM_DESC

The module_param is "cxgb3i_rx_credit_thres" so the MODULE_PARM_DESC()
should match that.

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years ago3w-xxxx: Pass through compat mode ioctls
Ryan C. Underwood [Fri, 25 Dec 2015 05:45:27 +0000 (21:45 -0800)]
3w-xxxx: Pass through compat mode ioctls

This allows 32-bit userspace tools (tw_cli, smartctl) to work on a
64-bit system.  The command buffer is opaque to us, so, no word size
problems.

Signed-off-by: Ryan C. Underwood <nemesis@icequake.net>
Reviewed-by: Shane Seymour <shane.seymour@hpe.com>
Reviewed-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agohisi_sas: Use u64 for qw0 in free_device_v1_hw()
John Garry [Mon, 21 Dec 2015 11:35:09 +0000 (19:35 +0800)]
hisi_sas: Use u64 for qw0 in free_device_v1_hw()

By reading in itct.qw0 into a 32b variable the top 32 bits were being
lost.  In practice this was OK as they were zeroes.

Fixes: 27a3f229 ("hisi_sas: Add cq interrupt")

Signed-off-by: John Garry <john.garry@huawei.com>
Reviewed-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Reviewed-by: Shane Seymour <shane.seymour@hpe.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agohisi_sas: Fix typo in setup_itct_v1_hw()
John Garry [Mon, 21 Dec 2015 11:35:08 +0000 (19:35 +0800)]
hisi_sas: Fix typo in setup_itct_v1_hw()

We were doing a arithmetic comparison instead of logical shift by
accident.  Mis-programming the itct did not seem to make a difference to
operation.

Fixes: abda97c2fe874 ("hisi_sas: Add dev_found")

Signed-off-by: John Garry <john.garry@huawei.com>
Reviewed-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Reviewed-by: Shane Seymour <shane.seymour@hpe.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agohisi_sas: Fix v1 itct masks
John Garry [Mon, 21 Dec 2015 11:35:07 +0000 (19:35 +0800)]
hisi_sas: Fix v1 itct masks

The mask fields are for quad-words, so add ULL suffix.  Also
unreferenced ITCT_HDR_BREAK_REPLY and ITCT_HDR_MAX_BURST are removed.

Fixes: 50af155b6c ("hisi_sas: Add v1 hardware reg")

Signed-off-by: John Garry <john.garry@huawei.com>
Reviewed-by: Shane Seymour <shane.seymour@hpe.com>
Reviewed-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoipr: Fix out-of-bounds null overwrite
Insu Yun [Wed, 6 Jan 2016 17:44:01 +0000 (12:44 -0500)]
ipr: Fix out-of-bounds null overwrite

Return value of snprintf is not bound by size value, 2nd argument.
(https://www.kernel.org/doc/htmldocs/kernel-api/API-snprintf.html).
Return value is number of printed chars, can be larger than 2nd
argument.  Therefore, it can write null byte out of bounds ofbuffer.
Since snprintf puts null, it does not need to put additional null byte.

Signed-off-by: Insu Yun <wuninsu@gmail.com>
Reviewed-by: Shane Seymour <shane.seymour@hpe.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoscsi: add Synology to 1024 sector blacklist
Mike Christie [Thu, 7 Jan 2016 22:34:05 +0000 (16:34 -0600)]
scsi: add Synology to 1024 sector blacklist

Another iscsi target that cannot handle large IOs, but does not tell us
a limit.

The Synology iSCSI targets report:

Block limits VPD page (SBC):
  Write same no zero (WSNZ): 0
  Maximum compare and write length: 0 blocks
  Optimal transfer length granularity: 0 blocks
  Maximum transfer length: 0 blocks
  Optimal transfer length: 0 blocks
  Maximum prefetch length: 0 blocks
  Maximum unmap LBA count: 0
  Maximum unmap block descriptor count: 0
  Optimal unmap granularity: 0
  Unmap granularity alignment valid: 0
  Unmap granularity alignment: 0
  Maximum write same length: 0x0 blocks

and the size of the command it can handle seems to depend on how much
memory it can allocate at the time. This results in IO errors when
handling large IOs. This patch just has us use the old 1024 default
sectors for this target by adding it to the scsi blacklist. We do not
have good contacs with this vendors, so I have not been able to try and
fix on their side.

I have posted this a long while back, but it was not merged. This
version just fixes it up for merge/patch failures in the original
version.

Reported-by: Ancoron Luciferis <ancoron.luciferis@googlemail.com>
Reported-by: Michael Meyers <steltek@tcnnet.com>
Signed-off-by: Mike Christie <mchristi@redhat.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Cc: <stable@vger.kernel.org> # 4.1+
8 years agoMerge branch 'jejb-scsi' into misc
James Bottomley [Thu, 7 Jan 2016 23:51:13 +0000 (15:51 -0800)]
Merge branch 'jejb-scsi' into misc

8 years agoncr5380: Add support for HP C2502
Ondrej Zary [Sun, 3 Jan 2016 05:06:19 +0000 (16:06 +1100)]
ncr5380: Add support for HP C2502

HP C2502 cards (based on 53C400A chips) use different magic numbers for
software-based I/O address configuration than other cards.
The configuration is also extended to allow setting the IRQ.

Move the configuration to a new function magic_configure() and move
magic the magic numbers into an array. Add new magic numbers for these
HP cards and hp_c2502 module parameter to use them, e.g.:
modprobe g_NCR5380 ncr_irq=7 ncr_addr=0x280 hp_c2502=1

Tested with HP C2502 and DTCT-436P.

Signed-off-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix wait for 53C80 registers registers after PDMA
Ondrej Zary [Sun, 3 Jan 2016 05:06:18 +0000 (16:06 +1100)]
ncr5380: Fix wait for 53C80 registers registers after PDMA

The check for 53C80 registers accessibility was commented out because
it was broken (inverted). Fix and enable it.

Signed-off-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Enable PDMA for DTC chips
Ondrej Zary [Sun, 3 Jan 2016 05:06:17 +0000 (16:06 +1100)]
ncr5380: Enable PDMA for DTC chips

Add I/O register mapping for DTC chips and enable PDMA mode.

These chips have 16-bit wide HOST BUFFER register and it must be read
by 16-bit accesses (we lose data otherwise).

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 on 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@rainbow-software.org>
Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Enable PDMA for NCR53C400A
Ondrej Zary [Sun, 3 Jan 2016 05:06:16 +0000 (16:06 +1100)]
ncr5380: Enable PDMA for NCR53C400A

Add I/O register mapping for NCR53C400A and enable PDMA mode to
improve performance and fix non-working IRQ.

Tested with HP C2502 (and user-space enabler).

Signed-off-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Use runtime register mapping
Ondrej Zary [Sun, 3 Jan 2016 05:06:15 +0000 (16:06 +1100)]
ncr5380: Use runtime register mapping

Convert compile-time C400_ register mapping to runtime mapping.
This removes the weird negative register offsets and allows adding
additional mappings.

While at it, convert read/write loops into insb/outsb.

Signed-off-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix pseudo DMA transfers on 53C400
Ondrej Zary [Sun, 3 Jan 2016 05:06:14 +0000 (16:06 +1100)]
ncr5380: Fix pseudo DMA transfers on 53C400

Pseudo-DMA (PDMA) has been broken for ages, resulting in hangs on
53C400-based cards.

According to 53C400 datasheet, PDMA transfer length must be a multiple
of 128. Check if that's true and use PIO if it's not.

This makes PDMA work on 53C400 (Canon FG2-5202).

Signed-off-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Cleanup whitespace and parentheses
Finn Thain [Sun, 3 Jan 2016 05:06:13 +0000 (16:06 +1100)]
ncr5380: Cleanup whitespace and parentheses

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoatari_NCR5380: Merge changes from NCR5380.c
Finn Thain [Sun, 3 Jan 2016 05:06:12 +0000 (16:06 +1100)]
atari_NCR5380: Merge changes from NCR5380.c

In the past, atari_NCR5380.c was overlooked by those working on NCR5380.c
and this caused needless divergence. All of the changes in this patch were
taken from NCR5380.c.

This removes some unimportant discrepancies between the two core driver
forks so that 'diff' can be used to reveal the important ones, to
facilitate reunification.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Merge changes from atari_NCR5380.c
Finn Thain [Sun, 3 Jan 2016 05:06:11 +0000 (16:06 +1100)]
ncr5380: Merge changes from atari_NCR5380.c

In the past, NCR5380.c was overlooked by those working on atari_NCR5380.c
and this caused needless divergence. All of the changes in this patch were
taken from atari_NCR5380.c.

This removes some unimportant discrepancies between the two core driver
forks so that 'diff' can be used to reveal the important ones, to
facilitate reunification.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix whitespace in comments using regexp
Finn Thain [Sun, 3 Jan 2016 05:06:10 +0000 (16:06 +1100)]
ncr5380: Fix whitespace in comments using regexp

Hanging indentation was a poor choice for the text inside comments. It
has been used in the wrong places and done badly elsewhere. There is
little consistency within any file. One fork of the core driver uses
tabs for this indentation while the other uses spaces. Better to use
flush-left alignment throughout.

This patch is the result of the following substitution. It replaces tabs
and spaces at the start of a comment line with a single space.

perl -i -pe 's,^(\t*[/ ]\*)[ \t]+,$1 ,' drivers/scsi/{atari_,}NCR5380.c

This removes some unimportant discrepancies between the two core driver
forks so that the important ones become obvious, to facilitate
reunification.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix trailing whitespace using regexp
Finn Thain [Sun, 3 Jan 2016 05:06:09 +0000 (16:06 +1100)]
ncr5380: Fix trailing whitespace using regexp

This patch is the result of the following substitution. It removes any
tabs and spaces at the end of a line.

perl -i -pe 's,[\t ]+$,,' drivers/scsi/{atari_,}NCR5380.c

This removes some unimportant discrepancies between the two core driver
forks so that the important ones become obvious, to facilitate
reunification.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Cleanup comments
Finn Thain [Sun, 3 Jan 2016 05:06:08 +0000 (16:06 +1100)]
ncr5380: Cleanup comments

The CVS revision log is not nearly as useful as the history/history.git
repo, so remove it. Roman's commentary at the top of his driver repeats
the same information elsewhere in the file so remove it. Also remove
some other redundant or obsolete comments.

Both the driver and the datasheets confusingly refer to a DMA access
for a SCSI WRITE command as a "DMA write". Similarly a SCSI READ command
is called a "DMA read". This is the opposite of the usual convention.
Thankfully, the chip documentation and driver code also use "DMA send" and
"DMA receive", so adopt this terminology.

This removes some unimportant discrepancies between the two core driver
forks so that 'diff' can be used to reveal the important ones, to
facilitate reunification.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix soft lockups
Finn Thain [Sun, 3 Jan 2016 05:06:07 +0000 (16:06 +1100)]
ncr5380: Fix soft lockups

Because of the rudimentary design of the chip, it is necessary to poll the
SCSI bus signals during PIO and this tends to hog the CPU. The driver will
accept new commands while others execute, and this causes a soft lockup
because the workqueue item will not terminate until the issue queue is
emptied.

When exercising dmx3191d using sequential IO from dd, the driver is sent
512 KiB WRITE commands and 128 KiB READs. For a PIO transfer, the rate is
is only about 300 KiB/s, so these are long-running commands. And although
PDMA may run at several MiB/s, interrupts are disabled for the duration
of the transfer.

Fix the unresponsiveness and soft lockup issues by calling cond_resched()
after each command is completed and by limiting max_sectors for drivers
that don't implement real DMA.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoatari_scsi, sun3_scsi: Remove global Scsi_Host pointer
Finn Thain [Sun, 3 Jan 2016 05:06:06 +0000 (16:06 +1100)]
atari_scsi, sun3_scsi: Remove global Scsi_Host pointer

This refactoring removes two global Scsi_Host pointers. This
improves consistency with other ncr5380 drivers. Adopting the same
conventions as the other drivers makes them easier to read.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoatari_NCR5380: Eliminate HOSTNO macro
Finn Thain [Sun, 3 Jan 2016 05:06:05 +0000 (16:06 +1100)]
atari_NCR5380: Eliminate HOSTNO macro

Keep the two core driver forks in sync.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoatari_NCR5380: Remove HOSTNO macro from printk() and seq_printf() calls
Finn Thain [Sun, 3 Jan 2016 05:06:04 +0000 (16:06 +1100)]
atari_NCR5380: Remove HOSTNO macro from printk() and seq_printf() calls

Remove the HOSTNO macro that is peculiar to atari_NCR5380.c and
contributes to the problem of divergence of the NCR5380 core drivers.
Keep NCR5380.c in sync.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Implement new eh_bus_reset_handler
Finn Thain [Sun, 3 Jan 2016 05:06:03 +0000 (16:06 +1100)]
ncr5380: Implement new eh_bus_reset_handler

NCR5380.c lacks a sane eh_bus_reset_handler. The atari_NCR5380.c code is
much better but it should not throw out the issue queue (that would be
a host reset) and it neglects to set the result code for commands that it
throws out. Fix these bugs and keep the two core drivers in sync.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix EH during arbitration and selection
Finn Thain [Sun, 3 Jan 2016 05:06:02 +0000 (16:06 +1100)]
ncr5380: Fix EH during arbitration and selection

During arbitration and selection, the relevant command is invisible to
exception handlers and can be found only in a pointer on the stack of a
different thread.

When eh_abort_handler can't find a given command, it can't decide whether
that command was completed already or is still in arbitration or selection
phase. But it must return either SUCCESS (e.g. command completed earlier)
or FAILED (could not abort the nexus, try bus reset).

The solution is to make sure all commands belonging to the LLD are always
visible to exception handlers. Add another scsi_cmnd pointer to the
hostdata struct to track the command in arbitration or selection phase.

Replace 'retain_dma_irq' with the new 'selecting' pointer, to bring
atari_NCR5380.c into line with NCR5380.c.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Implement new eh_abort_handler
Finn Thain [Sun, 3 Jan 2016 05:06:01 +0000 (16:06 +1100)]
ncr5380: Implement new eh_abort_handler

Introduce a new eh_abort_handler implementation. This one attempts to
follow all of the rules relating to EH handlers. There is still a known
bug: during selection, a command becomes invisible to the EH handlers
because it only appears in a pointer on the stack of a different thread.
This bug is addressed in a subsequent patch.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix autosense bugs
Finn Thain [Sun, 3 Jan 2016 05:06:00 +0000 (16:06 +1100)]
ncr5380: Fix autosense bugs

NCR5380_information_transfer() may re-queue a command for autosense,
after calling scsi_eh_prep_cmnd(). This creates several possibilities:

1. Reselection may intervene before the re-queued command gets processed.
   If the reconnected command then undergoes autosense, this causes the
   scsi_eh_save data from the previous command to be overwritten.

2. After NCR5380_information_transfer() calls scsi_eh_prep_cmnd(),
   a new REQUEST SENSE command may arrive. This would be queued ahead
   of any command already undergoing autosense, which means the
   scsi_eh_save data might be restored to the wrong command.

3. After NCR5380_information_transfer() calls scsi_eh_prep_cmnd(),
   eh_abort_handler() may abort the command. But the scsi_eh_save data is
   not discarded, which means the scsi_eh_save data might be incorrectly
   restored to the next REQUEST SENSE command issued.

This patch adds a new autosense list so that commands that are re-queued
because of a CHECK CONDITION result can be kept apart from the REQUEST
SENSE commands that arrive via queuecommand.

This patch also adds a function dedicated to dequeueing and preparing the
next command for processing. By refactoring the main loop in this way,
scsi_eh_save takes place when an autosense command is dequeued rather
than when re-queued.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Refactor command completion
Finn Thain [Sun, 3 Jan 2016 05:05:59 +0000 (16:05 +1100)]
ncr5380: Refactor command completion

Implement a 'complete_cmd' function to complete commands. This is needed
by the following patch; the new function provides a site for the logic
needed to correctly handle REQUEST SENSE commands.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Use standard list data structure
Finn Thain [Sun, 3 Jan 2016 05:05:58 +0000 (16:05 +1100)]
ncr5380: Use standard list data structure

The NCR5380 drivers have a home-spun linked list implementation for
scsi_cmnd structs that uses cmd->host_scribble as a 'next' pointer. Adopt
the standard list_head data structure and list operations instead. Remove
the eh_abort_handler rather than convert it. Doing the conversion would
only be churn because the existing EH handlers don't work and get replaced
in a subsequent patch.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove redundant volatile qualifiers
Finn Thain [Sun, 3 Jan 2016 05:05:57 +0000 (16:05 +1100)]
ncr5380: Remove redundant volatile qualifiers

The hostdata struct is now protected by a spin lock so the volatile
qualifiers are redundant. Remove them.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove LIST and REMOVE macros
Finn Thain [Sun, 3 Jan 2016 05:05:56 +0000 (16:05 +1100)]
ncr5380: Remove LIST and REMOVE macros

Printing command pointers can be useful when debugging queues. Other than
that, the LIST and REMOVE macros are just clutter. These macros are
redundant now that NDEBUG_QUEUES causes pointers to be printed, so remove
them.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Use dsprintk() for queue debugging
Finn Thain [Sun, 3 Jan 2016 05:05:55 +0000 (16:05 +1100)]
ncr5380: Use dsprintk() for queue debugging

Print the command pointers in the log messages for debugging queue data
structures. The LIST and REMOVE macros can then be removed.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Use shost_priv helper
Finn Thain [Sun, 3 Jan 2016 05:05:54 +0000 (16:05 +1100)]
ncr5380: Use shost_priv helper

Make use of the shost_priv() helper. Remove HOSTDATA and SETUP_HOSTDATA
macros because they harm readability.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove H_NO macro and introduce dsprintk
Finn Thain [Sun, 3 Jan 2016 05:05:53 +0000 (16:05 +1100)]
ncr5380: Remove H_NO macro and introduce dsprintk

Replace all H_NO and some HOSTNO macros (both peculiar to atari_NCR5380.c)
with a new dsprintk macro that's more useful and more consistent. The new
macro avoids a lot of boilerplate in new code in subsequent patches. Keep
NCR5380.c in sync. Remaining HOSTNO macros are removed as side-effects
of subsequent patches.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove command list debug code
Finn Thain [Sun, 3 Jan 2016 05:05:52 +0000 (16:05 +1100)]
ncr5380: Remove command list debug code

Some NCR5380 hosts offer a .show_info method to access the contents of
the various command list data structures from a procfs file. When NDEBUG
is set, the same information is sent to the console during EH.

The two core drivers, atari_NCR5380.c and NCR5380.c differ here. Because
it is just for debugging, the easiest way to fix the discrepancy is
simply remove this code.

The only remaining users of NCR5380_show_info() and NCR5380_write_info()
are drivers that define PSEUDO_DMA. The others have no use for the
.show_info method, so don't initialize it.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Change instance->host_lock to hostdata->lock
Finn Thain [Sun, 3 Jan 2016 05:05:51 +0000 (16:05 +1100)]
ncr5380: Change instance->host_lock to hostdata->lock

NCR5380.c presently uses the instance->host_lock spin lock. Convert this
to a new spin lock that protects the NCR5380_hostdata struct.

atari_NCR5380.c previously used local_irq_save/restore() rather than a
spin lock. Convert this to hostdata->lock in irq mode. For SMP platforms,
the interrupt handler now also acquires the spin lock.

This brings all locking in the two core drivers into agreement.

Adding this locking also means that a bunch of volatile qualifiers can be
removed from the members of the NCR5380_hostdata struct. This is done in
a subsequent patch.

Proper locking will allow the abort handler to locate a command being
aborted. This is presently impossible if the abort handler is invoked when
the command has been moved from a queue to a pointer on the stack. (If
eh_abort_handler can't determine whether a command has been completed
or is still being processed then it can't decide whether to return
success or failure.)

The hostdata spin lock is now held when calling NCR5380_select() and
NCR5380_information_transfer(). Where possible, the lock is dropped for
polling and PIO transfers.

Clean up the now-redundant SELECT_ENABLE_REG writes, that used to provide
limited mutual exclusion between information_transfer() and reselect().

Accessing hostdata->connected without data races means taking the lock;
cleanup these accesses.

The new spin lock falls away for m68k and other UP builds, so this should
have little impact there. In the SMP case the new lock should be
uncontested even when the SCSI bus is contested.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove redundant ICR_ARBITRATION_LOST test and eliminate FLAG_DTC3181E
Finn Thain [Sun, 3 Jan 2016 05:05:50 +0000 (16:05 +1100)]
ncr5380: Remove redundant ICR_ARBITRATION_LOST test and eliminate FLAG_DTC3181E

Remove FLAG_DTC3181E. It was used to suppress a final Arbitration Lost
(SEL asserted) test that isn't actually needed. The test was suppressed
because it causes problems for DTC436 and DTC536 chips. It takes place
after the host wins arbitration, so SEL has been asserted. These chips
can't seem to tell whether it was the host or another bus device that
did so.

This questionable final test appears in a flow chart in an early NCR5380
datasheet. It was removed from later documents like the DP5380 datasheet.

By the time this final test takes place, the driver has already tested
the Arbitration Lost bit several times. The first test happens 3 us after
BUS FREE (or longer due to register access delays). The protocol requires
that a device stop signalling within 1.8 us after BUS FREE unless it won
arbitration, in which case it must assert SEL, which is detected 1.2 us
later by the first Arbitration Lost test.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoatari_NCR5380: Fix queue_size limit
Finn Thain [Sun, 3 Jan 2016 05:05:49 +0000 (16:05 +1100)]
atari_NCR5380: Fix queue_size limit

When a target reports a QUEUE_FULL condition it causes the driver to
update the 'queue_size' limit with the number of currently allocated tags.
At least, that's what's supposed to happen, according to the comments.
Unfortunately the terms in the assignment are swapped. Fix this and
cleanup some obsolete comments.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix and cleanup scsi_host_template initializers
Finn Thain [Sun, 3 Jan 2016 05:05:48 +0000 (16:05 +1100)]
ncr5380: Fix and cleanup scsi_host_template initializers

Add missing .module initializer. Use distinct .proc_name values for the
g_NCR5380 and g_NCR5380_mmio modules. Remove pointless CAN_QUEUE and
CMD_PER_LUN override macros. Cleanup whitespace and code style.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix NDEBUG_NO_DATAOUT flag
Finn Thain [Sun, 3 Jan 2016 05:05:47 +0000 (16:05 +1100)]
ncr5380: Fix NDEBUG_NO_DATAOUT flag

NDEBUG_NO_DATAOUT should not disable DATA IN phases too. Fix this.
(This bug has long been fixed in atari_NCR5380.c.)

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Cleanup #include directives
Finn Thain [Sun, 3 Jan 2016 05:05:46 +0000 (16:05 +1100)]
ncr5380: Cleanup #include directives

Remove unused includes (stat.h, signal.h, proc_fs.h) and move includes
needed by the core drivers into the common header (delay.h etc).

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix off-by-one bug in extended_msg[] bounds check
Finn Thain [Sun, 3 Jan 2016 05:05:45 +0000 (16:05 +1100)]
ncr5380: Fix off-by-one bug in extended_msg[] bounds check

Fix the array bounds check when transferring an extended message from the
target.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Standardize reselection handling
Finn Thain [Sun, 3 Jan 2016 05:05:44 +0000 (16:05 +1100)]
ncr5380: Standardize reselection handling

Bring the two NCR5380_reselect() implementations into agreement.

Replace infinite loops in atari_NCR5380.c with timeouts, as per NCR5380.c.

Remove 'abort' flag in NCR5380.c as per atari_NCR5380.c -- if reselection
fails, there may be no MESSAGE IN phase so don't attempt data transfer.

During selection, don't interfere with the chip registers after a
reselection interrupt intervenes.

Clean up some trivial issues with code style, comments and printk.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Replace READ_OVERRUNS macro with FLAG_NO_DMA_FIXUPS
Finn Thain [Sun, 3 Jan 2016 05:05:43 +0000 (16:05 +1100)]
ncr5380: Replace READ_OVERRUNS macro with FLAG_NO_DMA_FIXUPS

The workarounds for chip errata appear twice, in slightly different
forms. One is used when defined(REAL_DMA) || defined(REAL_DMA_POLL), the
other when defined(PSEUDO_DMA). In the PDMA case, the workarounds have
been made conditional on FLAG_NO_DMA_FIXUPS. Do the same for the DMA case,
to eliminate the READ_OVERRUNS macro.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Replace redundant flags with FLAG_NO_DMA_FIXUP
Finn Thain [Sun, 3 Jan 2016 05:05:42 +0000 (16:05 +1100)]
ncr5380: Replace redundant flags with FLAG_NO_DMA_FIXUP

The flags DMA_WORKS_RIGHT, FLAG_NCR53C400 and FLAG_HAS_LAST_BYTE_SENT
all mean the same thing, i.e. the chip is not a 538[01]. (More recent
devices such as the 53C80 have a 'Last Byte Sent' bit in the Target
Command Register as well as other fixes for End-of-DMA errata.)

These flags have no additional meanings since previous cleanup patches
eliminated the NCR53C400 macro, moved g_NCR5380-specific code out of the
core driver and standardized interrupt handling.

Use the FLAG_NO_DMA_FIXUP flag to suppress End-of-DMA errata workarounds,
for those cards and drivers that make use of the TCR_LAST_BYTE_SENT bit.
Remove the old flags.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Introduce NCR5380_poll_politely2
Finn Thain [Sun, 3 Jan 2016 05:05:41 +0000 (16:05 +1100)]
ncr5380: Introduce NCR5380_poll_politely2

SCSI bus protocol sometimes requires monitoring two related conditions
simultaneously. Enhance NCR5380_poll_politely() for this purpose, and
put it to use in the arbitration algorithm. It will also find use in
pseudo DMA.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Standardize interrupt handling
Finn Thain [Sun, 3 Jan 2016 05:05:40 +0000 (16:05 +1100)]
ncr5380: Standardize interrupt handling

Because interrupt handling is crucial to the core driver(s), all wrapper
drivers need to agree on this code. This patch removes discrepancies.

NCR5380_intr() in NCR5380.c has the following pointless loop that differs
from the code in atari_NCR5380.c.

done = 1;
do {
/* ... */
} while (!done);

The 'done' flag gets cleared when a reconnected command is to be processed
from the work queue. But in NCR5380.c, the flag is also used to cause the
interrupt conditions to be re-examined. Perhaps this was because
NCR5380_reselect() was expected to cause another interrupt, or perhaps
the remaining present interrupt conditions need to be handled after the
NCR5380_reselect() call?

Actually, both possibilities are bogus, as is the loop itself. It seems
have been overlooked in the hit-and-miss removal of scsi host instance
list iteration many years ago; see history/history.git commit 491447e1fcff
("[PATCH] next NCR5380 updates") and commit 69e1a9482e57 ("[PATCH] fix up
NCR5380 private data"). See also my earlier patch, "Always retry
arbitration and selection".

The datasheet says, "IRQ can be reset simply by reading the Reset
Parity/Interrupt Register". So don't treat the chip IRQ like a
level-triggered interrupt. Of the conditions that set the IRQ flag,
some are level-triggered and some are edge-triggered, which means IRQ
itself must be edge-triggered.

Some interrupt conditions are latched and some are not. Before clearing
the chip IRQ flag, clear all state that may cause it to be raised. That
means clearing the DMA Mode and Busy Monitor bits in the Mode Register
and clearing the host ID in the Select Enable register.

Also clean up some printk's and some comments. Keep atari_NCR5380.c and
NCR5380.c in agreement.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove UNSAFE macro
Finn Thain [Sun, 3 Jan 2016 05:05:39 +0000 (16:05 +1100)]
ncr5380: Remove UNSAFE macro

Configuring core drivers using macros like this one prevents re-unifying
the core driver forks, and prevents implementing the core driver as a
library or a platform driver.

The UNSAFE macro in particular is a poor workaround for the problem of
interrupt latency. Releasing the locks complicates things because then we
would have to handle the possibility of EH handler invocation during a
PDMA transfer.

The comments say that instead of using this macro, "you're going to be
better off twiddling with transfersize". I agree. Remove this stuff.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Standardize work queueing algorithm
Finn Thain [Sun, 3 Jan 2016 05:05:38 +0000 (16:05 +1100)]
ncr5380: Standardize work queueing algorithm

The complex main_running/queue_main mechanism is peculiar to
atari_NCR5380.c. It isn't SMP safe and offers little value given that
the work queue already offers concurrency management. Remove this
complexity to bring atari_NCR5380.c closer to NCR5380.c.

It is not a good idea to call the information transfer state machine from
queuecommand because, according to Documentation/scsi/scsi_mid_low_api.txt
that could happen in soft irq context. Fix this.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Use work_struct instead of delayed_work
Finn Thain [Sun, 3 Jan 2016 05:05:37 +0000 (16:05 +1100)]
ncr5380: Use work_struct instead of delayed_work

Each host instance now has it's own work queue so the main() work item can
sleep when necessary. That means we can use a simple work item rather than
a delayed work item. This brings NCR5380.c closer to atari_NCR5380.c.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Dont wait for BUS FREE after disconnect
Finn Thain [Sun, 3 Jan 2016 05:05:36 +0000 (16:05 +1100)]
ncr5380: Dont wait for BUS FREE after disconnect

When there is a queued command and no connected command, NCR5380_select()
is called and arbitration begins. The chip waits for BUS FREE once the
MR_ARBITRATE bit in the mode register is enabled. That means there is
no need to wait for BUS FREE after disconnecting.

There is presently no polling for BUS FREE after sending an ABORT or
other message that might lead to disconnection. It only happens after
COMMAND COMPLETE or DISCONNECT messages, which seems inconsistent.
Remove the polling for !BSY in the COMMAND COMPLETE and DISCONNECT
cases.

BTW, the comments say "avoid nasty timeouts" and perhaps BUS FREE polling
was somehow helpful back in Linux v0.99.14u, when it was introduced.
The relevant timeout is presently 1 second (for bus arbitration).

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoatari_NCR5380: Use arbitration timeout
Finn Thain [Sun, 3 Jan 2016 05:05:35 +0000 (16:05 +1100)]
atari_NCR5380: Use arbitration timeout

Allow target selection to fail with a timeout instead of waiting in
infinite loops. This gets rid of the unused NCR_TIMEOUT macro, it is more
defensive and has proved helpful in debugging.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoatari_NCR5380: Set do_abort() timeouts
Finn Thain [Sun, 3 Jan 2016 05:05:34 +0000 (16:05 +1100)]
atari_NCR5380: Set do_abort() timeouts

Use timeouts in do_abort() in atari_NCR5380.c instead of infinite loops.
Also fix the kernel-doc comment. Keep the two core driver forks in sync.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix bus phase in do_abort()
Finn Thain [Sun, 3 Jan 2016 05:05:33 +0000 (16:05 +1100)]
ncr5380: Fix bus phase in do_abort()

NCR5380_poll_politely() returns either 0 (success) or -ETIMEDOUT. However,
in do_abort(), the return value is incorrectly taken to be the status
register value. This means that the bus is put into DATA OUT phase instead
of MESSAGE OUT. Fix this.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix !REQ timeout in do_abort()
Finn Thain [Sun, 3 Jan 2016 05:05:32 +0000 (16:05 +1100)]
ncr5380: Fix !REQ timeout in do_abort()

NCR5380_poll_politely() never returns -1. That means do_abort() can fail
to handle a timeout after waiting for the target to negate REQ. Fix this
and cleanup other NCR5380_poll_politely() call sites.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Add missing break after case MESSAGE_REJECT
Finn Thain [Sun, 3 Jan 2016 05:05:31 +0000 (16:05 +1100)]
ncr5380: Add missing break after case MESSAGE_REJECT

MESSAGE REJECT does not imply DISCONNECT: the target is about to enter
MESSAGE IN or MESSAGE OUT phase.

This bug fix comes from atari_NCR5380.c. Unfortunately it never made it
into the original NCR5380.c core driver.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove references to linked commands
Hannes Reinecke [Sun, 3 Jan 2016 05:05:30 +0000 (16:05 +1100)]
ncr5380: Remove references to linked commands

Some old drivers partially implemented support for linked commands using
a "proposed" next_link pointer in struct scsi_cmnd that never actually
existed. Remove this code.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Drop DEF_SCSI_QCMD macro
Finn Thain [Sun, 3 Jan 2016 05:05:29 +0000 (16:05 +1100)]
ncr5380: Drop DEF_SCSI_QCMD macro

Remove the DEF_SCSI_QCMD macro (already removed from atari_NCR5380.c). The
lock provided by DEF_SCSI_QCMD is only needed for queue data structures.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Add missing lock in eh_abort_handler
Finn Thain [Sun, 3 Jan 2016 05:05:28 +0000 (16:05 +1100)]
ncr5380: Add missing lock in eh_abort_handler

The host spin lock needs to be acquired by NCR5380_abort() before it calls
NCR5380_select(). This patch doesn't actually fix the EH issues in this
driver but it does avoid this:

BUG: spinlock already unlocked on CPU#0, kworker/u4:1/14
 lock: 0xc0c0f834, .magic: dead4ead, .owner: <none>/-1, .owner_cpu: -1
 CPU: 0 PID: 14 Comm: kworker/u4:1 Not tainted 3.15.5 #5
 Workqueue: scsi_tmf_4 scmd_eh_abort_handler
 Call Trace:
 [ef885d70] [c0008acc] show_stack+0x70/0x1bc (unreliable)
 [ef885db0] [c0492a00] dump_stack+0x84/0x684
 [ef885dc0] [c006f314] spin_dump+0xd0/0xe8
 [ef885dd0] [c006f460] do_raw_spin_unlock+0xd4/0xd8
 [ef885df0] [c0491c8c] _raw_spin_unlock_irq+0x10/0x3c
 [ef885e00] [f381fe3c] NCR5380_select+0x3e4/0x6e8 [dmx3191d]
 [ef885e40] [f382026c] NCR5380_abort+0x12c/0x190 [dmx3191d]
 [ef885e60] [c02fec9c] scmd_eh_abort_handler+0x100/0x460
 [ef885e80] [c0046470] process_one_work+0x16c/0x420
 [ef885ea0] [c0046870] worker_thread+0x14c/0x430
 [ef885ed0] [c004e4f4] kthread+0xd8/0xec
 [ef885f40] [c00124d4] ret_from_kernel_thread+0x5c/0x64

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Fix NCR5380_transfer_pio() result
Finn Thain [Sun, 3 Jan 2016 05:05:27 +0000 (16:05 +1100)]
ncr5380: Fix NCR5380_transfer_pio() result

According to the SCSI-2 draft revision 10L, atari_NCR5380.c is correct
when it says that the phase lines are valid up until ACK is negated
following the transmission of the last byte in MESSAGE IN phase. This is
true for all information transfer phases, from target to initiator.

Sample the phase bits in STATUS_REG so that NCR5380_transfer_pio() can
return the correct result. The return value is presently unused (perhaps
because of bugs like this) but this change at least fixes the caller's
phase variable, which is passed by reference.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Rework disconnect versus poll logic
Finn Thain [Sun, 3 Jan 2016 05:05:26 +0000 (16:05 +1100)]
ncr5380: Rework disconnect versus poll logic

The atari_NCR5380.c and NCR5380.c core drivers differ in their handling of
target disconnection. This is partly because atari_NCR5380.c had all of
the polling and sleeping removed to become entirely interrupt-driven, and
it is partly because of damage done to NCR5380.c after atari_NCR5380.c was
forked. See commit 37cd23b44929 ("Linux 2.1.105") in history/history.git.

The polling changes that were made in v2.1.105 are questionable at best:
if REQ is not already asserted when NCR5380_transfer_pio() is invoked, and
if the expected phase is DATA IN or DATA OUT, the function will schedule
main() to execute after USLEEP_SLEEP jiffies and then return. The problems
here are the expected REQ timing and the sleep interval*. Avoid this issue
by using NCR5380_poll_politely() instead of scheduling main().

The atari_NCR5380.c core driver requires the use of the chip interrupt and
always permits target disconnection. It sets the cmd->device->disconnect
flag when a device disconnects, but never tests this flag.

The NCR5380.c core driver permits disconnection only when
instance->irq != NO_IRQ. It sets the cmd->device->disconnect flag when
a device disconnects and it tests this flag in a couple of places:

1. During NCR5380_information_transfer(), following COMMAND OUT phase,
   if !cmd->device->disconnect, the initiator will take a guess as to
   whether or not the target will then choose to go to MESSAGE IN phase
   and disconnect. If the driver guesses "yes", it will schedule main()
   to execute after USLEEP_SLEEP jiffies and then return there.

   Unfortunately the driver may guess "yes" even after it has denied
   the target the disconnection privilege. When the target does not
   disconnect, the sleep can be beneficial, assuming the sleep interval
   is appropriate (mostly it is not*).

   And even if the driver guesses "yes" correctly, and the target would
   then disconnect, the driver still has to go through the MESSAGE IN
   phase in order to get to BUS FREE phase. The main loop can do nothing
   useful until BUS FREE, and sleeping just delays the phase transition.

2. If !cmd->device->disconnect and REQ is not already asserted when
   NCR5380_information_transfer() is invoked, the function polls for REQ
   for USLEEP_POLL jiffies. If REQ is not asserted, it then schedules
   main() to execute after USLEEP_SLEEP jiffies and returns.

   The idea is apparently to yeild the CPU while waiting for REQ.
   This is conditional upon !cmd->device->disconnect, but there seems
   to be no rhyme or reason for that. For example, the flag may be
   unset because disconnection privilege was denied because the driver
   has no IRQ. Or the flag may be unset because the device has never
   needed to disconnect before. Or if the flag is set, disconnection
   may have no relevance to the present bus phase.

Another deficiency of the existing algorithm is as follows. When the
driver has no IRQ, it prevents disconnection, and generally polls and
sleeps more than it would normally. Now, if the driver is going to poll
anyway, why not allow the target to disconnect? That way the driver can do
something useful with the bus instead of polling unproductively!

Avoid this pointless latency, complexity and guesswork by using
NCR5380_poll_politely() instead of scheduling main().

* For g_NCR5380, the time intervals for USLEEP_SLEEP and USLEEP_POLL are
  200 ms and 10 ms, respectively. They are 20 ms and 200 ms respectively
  for the other NCR5380 drivers. There doesn't seem to be any reason for
  this discrepancy. The timing seems to have no relation to the type of
  adapter. Bizarrely, the timing in g_NCR5380 seems to relate only to one
  particular type of target device. This patch attempts to solve the
  problem for all NCR5380 drivers and all target devices.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Implement NCR5380_dma_xfer_len and remove LIMIT_TRANSFERSIZE macro
Finn Thain [Sun, 3 Jan 2016 05:05:25 +0000 (16:05 +1100)]
ncr5380: Implement NCR5380_dma_xfer_len and remove LIMIT_TRANSFERSIZE macro

Follow the example of the atari_NCR5380.c core driver and adopt the
NCR5380_dma_xfer_len() hook. Implement NCR5380_dma_xfer_len() for dtc.c
and g_NCR5380.c to take care of the limitations of these cards. Keep the
default for drivers using PSEUDO_DMA.

Eliminate the unused macro LIMIT_TRANSFERSIZE.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Always retry arbitration and selection
Finn Thain [Sun, 3 Jan 2016 05:05:24 +0000 (16:05 +1100)]
ncr5380: Always retry arbitration and selection

If NCR5380_select() returns -1, it means arbitration was lost or selection
failed and should be retried. If the main loop simply terminates when there
are still commands on the issue queue, they will remain queued until they
expire.

Fix this by clearing the 'done' flag after selection failure or lost
arbitration.

The "else break" clause in NCR5380_main() that gets removed here appears
to be a vestige of a long-gone loop that iterated over host instances.
See commit 491447e1fcff ("[PATCH] next NCR5380 updates") in
history/history.git.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Eliminate selecting state
Finn Thain [Sun, 3 Jan 2016 05:05:23 +0000 (16:05 +1100)]
ncr5380: Eliminate selecting state

Linux v2.1.105 changed the algorithm for polling for the BSY signal
in NCR5380_select() and NCR5380_main().

Presently, this code has a bug. Back then, NCR5380_set_timer(hostdata, 1)
meant reschedule main() after sleeping for 10 ms. Repeated 25 times this
provided the recommended 250 ms selection time-out delay. This got broken
when HZ became configurable.

We could fix this but there's no need to reschedule the main loop. This
BSY polling presently happens when the NCR5380_main() work queue item
calls NCR5380_select(), which in turn schedules NCR5380_main(), which
calls NCR5380_select() again, and so on.

This algorithm is a deviation from the simpler one in atari_NCR5380.c.
The extra complexity and state is pointless. There's no reason to
stop selection half-way and return to to the main loop when the main
loop can do nothing useful until selection completes.

So just poll for BSY. We can sleep while polling now that we have a
suitable workqueue.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Sleep when polling, if possible
Finn Thain [Sun, 3 Jan 2016 05:05:22 +0000 (16:05 +1100)]
ncr5380: Sleep when polling, if possible

When in process context, sleep during polling if doing so won't add
significant latency. In interrupt context or if the lock is held, poll
briefly then give up. Keep both core drivers in sync.

Calibrate busy-wait iterations to allow for variation in chip register
access times between different 5380 hardware implementations.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Introduce unbound workqueue
Finn Thain [Sun, 3 Jan 2016 05:05:21 +0000 (16:05 +1100)]
ncr5380: Introduce unbound workqueue

Allocate a work queue that will permit busy waiting and sleeping. This
means NCR5380_init() can potentially fail, so add this error path.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Cleanup bogus {request,release}_region() calls
Finn Thain [Sun, 3 Jan 2016 05:05:20 +0000 (16:05 +1100)]
ncr5380: Cleanup bogus {request,release}_region() calls

Commit 8b801ead3d7a ("[ARM] rpc: update Acorn SCSI drivers to modern ecard
interfaces") neglected to remove a request_region() call in cumana_1.c.

Commit eda32612f7b2 ("[PATCH] give all LLDD driver a ->release method") in
history/history.git added some pointless release_region() calls in dtc.c,
pas16.c and t128.c.

Fix these issues.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Eliminate USLEEP_WAITLONG delay
Finn Thain [Sun, 3 Jan 2016 05:05:19 +0000 (16:05 +1100)]
ncr5380: Eliminate USLEEP_WAITLONG delay

Linux 2.1.105 introduced the USLEEP_WAITLONG delay, apparently "needed for
Mustek scanners". It is intended to stall the issue queue for 5 seconds.
There are a number of problems with this.

1. Only g_NCR5380 enables the delay, which implies that the other five
   drivers using the NCR5380.c core driver remain incompatible with
   Mustek scanners.

2. The delay is not implemented by atari_NCR5380.c, which is problematic
   for re-unifying the two core driver forks.

3. The delay is implemented using NCR5380_set_timer() which makes it
   unreliable. A new command queued by the mid-layer cancels the delay.

4. The delay is applied indiscriminately in several situations in which
   NCR5380_select() returns -1. These are-- reselection by the target,
   failure of the target to assert BSY, and failure of the target to
   assert REQ. It's clear from the comments that USLEEP_WAITLONG is not
   relevant to the reselection case. And reportedly, these scanners do
   not disconnect.

5. atari_NCR5380.c was forked before Linux 2.1.105, so it was spared some
   of the damage done to NCR5380.c. In this case, the atari_NCR5380.c core
   driver was more standard-compliant and may not have needed any
   workaround like the USLEEP_WAITLONG kludge. The compliance issue was
   addressed in the previous patch.

If these scanners still don't work, we need a better solution. Retrying
selection until EH aborts a command offers equivalent robustness. Bugs in
the existing driver prevent EH working correctly but this is addressed in
a subsequent patch. Remove USLEEP_WAITLONG.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Keep BSY asserted when entering SELECTION phase
Finn Thain [Sun, 3 Jan 2016 05:05:18 +0000 (16:05 +1100)]
ncr5380: Keep BSY asserted when entering SELECTION phase

NCR5380.c is not compliant with the SCSI-2 standard (at least, not with
the draft revision 10L that I have to refer to). The selection algorithm
in atari_NCR5380.c is correct, so use that.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Proceed with next command after NCR5380_select() calls scsi_done
Finn Thain [Sun, 3 Jan 2016 05:05:17 +0000 (16:05 +1100)]
ncr5380: Proceed with next command after NCR5380_select() calls scsi_done

If a target disappears from the SCSI bus, NCR5380_select() may
subsequently fail with a time-out. In this situation, scsi_done is
called and NCR5380_select() returns 0. Both hostdata->connected and
hostdata->selecting are NULL and the main loop should proceed with
the next command in the issue queue. Clarify this logic.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Always escalate bad target time-out in NCR5380_select()
Finn Thain [Sun, 3 Jan 2016 05:05:16 +0000 (16:05 +1100)]
ncr5380: Always escalate bad target time-out in NCR5380_select()

Remove the restart_select and targets_present variables introduced in
Linux v1.1.38. The former was used only for a questionable debug printk
and the latter "so we can call a select failure a retryable condition".
Well, retrying select failure in general is a different problem to a
target that doesn't assert BSY. We need to handle these two cases
differently; the latter case can be left to the SCSI ML.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Use return instead of goto in NCR5380_select()
Finn Thain [Sun, 3 Jan 2016 05:05:15 +0000 (16:05 +1100)]
ncr5380: Use return instead of goto in NCR5380_select()

The "failed" label in NCR5380_select() is not helpful. Some failures
return 0, others -1. Use return instead of goto to improve clarity and
brevity, like atari_NCR5380.c does. Fix the relevant comments.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove redundant register writes
Finn Thain [Sun, 3 Jan 2016 05:05:14 +0000 (16:05 +1100)]
ncr5380: Remove redundant register writes

Remove the duplicate write to the Select Enable Register that appeared
in v1.1.38.

Also remove the redundant write to Initiator Command Register prior to
calling do_abort().

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove unused hostdata->aborted flag
Finn Thain [Sun, 3 Jan 2016 05:05:13 +0000 (16:05 +1100)]
ncr5380: Remove unused hostdata->aborted flag

The aborted flag was introduced in v1.1.38 but never used. Remove it.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Simplify bus reset handlers
Finn Thain [Sun, 3 Jan 2016 05:05:12 +0000 (16:05 +1100)]
ncr5380: Simplify bus reset handlers

Make use of do_reset() in the bus reset handler in atari_NCR5380.c. The
version in NCR5380.c already does so. Keep them in sync.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoatari_NCR5380: Remove RESET_BOOT, CONFIG_ATARI_SCSI_TOSHIBA_DELAY and CONFIG_ATARI_SC...
Finn Thain [Sun, 3 Jan 2016 05:05:11 +0000 (16:05 +1100)]
atari_NCR5380: Remove RESET_BOOT, CONFIG_ATARI_SCSI_TOSHIBA_DELAY and CONFIG_ATARI_SCSI_RESET_BOOT

The atari_NCR5380.c core driver now takes care of bus reset upon driver
initialization if required (same as NCR5380.c). Move the Toshiba CD-ROM
support into the core driver, enabled with a host flag, so that all
NCR5380 drivers can make use of it.

Drop the RESET_BOOT macros and the ATARI_SCSI_RESET_BOOT and
ATARI_SCSI_TOSHIBA_DELAY Kconfig symbols, which are now redundant.

Remove the atari_scsi_reset_boot(), mac_scsi_reset_boot() and
sun3_scsi_reset_boot() routines. None of this duplicated code is needed
now that all drivers can use NCR5380_maybe_reset_bus().

This brings atari_scsi, mac_scsi and sun3_scsi into line with all of the
other NCR5380 drivers.

The bus reset may raise an interrupt. That would be new behaviour for
atari_scsi only when CONFIG_ATARI_SCSI_RESET_BOOT=n. The ST DMA interrupt
is not assigned to atari_scsi at this stage, so
CONFIG_ATARI_SCSI_RESET_BOOT=y may well be problematic already.
Regardless, do_reset() now raises and clears the interrupt within
local_irq_save/restore which should avoid problems.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoatari_NCR5380: Reset bus on driver initialization if required
Finn Thain [Sun, 3 Jan 2016 05:05:10 +0000 (16:05 +1100)]
atari_NCR5380: Reset bus on driver initialization if required

Merge the bus reset code from NCR5380.c into atari_NCR5380.c. This allows
for removal of a lot of duplicated code conditional on the RESET_BOOT
macro (in the next patch).

The atari_NCR5380.c fork lacks the do_reset() and NCR5380_poll_politely()
routines from NCR5380.c, so introduce them. They are indispensible. Keep
the two implementations in sync.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Move NCR53C400-specific code
Finn Thain [Sun, 3 Jan 2016 05:05:09 +0000 (16:05 +1100)]
ncr5380: Move NCR53C400-specific code

Move board-specific code like this,
NCR5380_write(C400_CONTROL_STATUS_REG, CSR_BASE);
from the core driver to the board driver. Eliminate the NCR53C400 macro
from the core driver. Removal of all macros like this one will be
necessary in order to have one core driver that can support all kinds of
boards.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Split NCR5380_init() into two functions
Finn Thain [Sun, 3 Jan 2016 05:05:08 +0000 (16:05 +1100)]
ncr5380: Split NCR5380_init() into two functions

This patch splits the NCR5380_init() function into two parts, similar
to the scheme used with atari_NCR5380.c. This avoids two problems.

Firstly, NCR5380_init() may perform a bus reset, which would cause the
chip to assert IRQ. The chip is unable to mask its bus reset interrupt.
Drivers can't call request_irq() before calling NCR5380_init(), because
initialization must happen before the interrupt handler executes. If
driver initialization causes an interrupt it may be problematic on some
platforms. To avoid that, first move the bus reset code into
NCR5380_maybe_reset_bus().

Secondly, NCR5380_init() contains some board-specific interrupt setup code
for the NCR53C400 that does not belong in the core driver. In moving this
code, better not re-order interrupt initialization and bus reset. Again,
the solution is to move the bus reset code into NCR5380_maybe_reset_bus().

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove NCR5380_instance_name macro
Finn Thain [Sun, 3 Jan 2016 05:05:07 +0000 (16:05 +1100)]
ncr5380: Remove NCR5380_instance_name macro

This macro makes the code cryptic. Remove it.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove NCR5380_local_declare and NCR5380_setup macros
Finn Thain [Sun, 3 Jan 2016 05:05:06 +0000 (16:05 +1100)]
ncr5380: Remove NCR5380_local_declare and NCR5380_setup macros

The NCR5380_local_declare and NCR5380_setup macros exist to define and
initialize a particular local variable, to provide the address of the
chip registers needed for the driver's implementation of its
NCR5380_read/write register access macros.

In cumana_1 and macscsi, these macros generate pointless code like this,
struct Scsi_Host *_instance;
_instance = instance;

In pas16, the use of NCR5380_read/write in pas16_hw_detect() requires that
the io_port local variable has been defined and initialized, but the
NCR5380_local_declare and NCR5380_setup macros can't be used for that
purpose because the Scsi_Host struct has not yet been instantiated.

Moreover, these macros were removed from atari_NCR5380.c long ago and
now they constitute yet another discrepancy between the two core driver
forks.

Remove these "optimizations".

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove more pointless macros
Finn Thain [Sun, 3 Jan 2016 05:05:05 +0000 (16:05 +1100)]
ncr5380: Remove more pointless macros

ASM macro is never defined. rtrc in pas16.c is not used.
NCR5380_map_config, do_NCR5380_intr, do_t128_intr and do_pas16_intr
are unused. NCR_NOT_SET harms readability. Remove them.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Eliminate PDEBUG*, TDEBUG* and DTCDEBUG* macros
Finn Thain [Sun, 3 Jan 2016 05:05:04 +0000 (16:05 +1100)]
ncr5380: Eliminate PDEBUG*, TDEBUG* and DTCDEBUG* macros

Replace {P,T,DTC}DEBUG_INIT with NDEBUG_INIT. Remove dead debugging
code, including code that's conditional upon *DEBUG_TRANSFER.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoncr5380: Remove redundant static variable initializers
Finn Thain [Sun, 3 Jan 2016 05:05:03 +0000 (16:05 +1100)]
ncr5380: Remove redundant static variable initializers

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agoatari_scsi: Fix SCSI host ID setting
Finn Thain [Sun, 3 Jan 2016 05:05:02 +0000 (16:05 +1100)]
atari_scsi: Fix SCSI host ID setting

The NVRAM location of this byte is 16, as documented in
http://toshyp.atari.org/en/004009.html

This was confirmed by Michael Schmitz, by setting the SCSI host ID
under EmuTOS and then checking the value in /proc/driver/nvram and
/dev/nvram under Linux.

Signed-off-by: Finn Thain <fthain@telegraphics.com.au>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Michael Schmitz <schmitzmic@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agocxlflash: Enable device id for future IBM CXL adapter
Manoj Kumar [Mon, 14 Dec 2015 21:07:43 +0000 (15:07 -0600)]
cxlflash: Enable device id for future IBM CXL adapter

This drop enables a future card with a device id of 0x0600 to be
recognized by the cxlflash driver.

As per the design, the Accelerator Function Unit (AFU) for this new IBM
CXL Flash Adapter retains the same host interface as the previous
generation. For the early prototypes of the new card, the driver with
this change behaves exactly as the driver prior to this behaved with the
earlier generation card. Therefore, no card specific programming has
been added. These card specific changes can be staged in later if
needed.

Signed-off-by: Manoj N. Kumar <manoj@linux.vnet.ibm.com>
Acked-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agocxlflash: Resolve oops in wait_port_offline
Manoj Kumar [Mon, 14 Dec 2015 21:07:23 +0000 (15:07 -0600)]
cxlflash: Resolve oops in wait_port_offline

If an async error interrupt is generated, and the error requires the FC
link to be reset, it cannot be performed in the interrupt context. So a
work element is scheduled to complete the link reset in a process
context. If either an EEH event or an escalation occurs in between when
the interrupt is generated and the scheduled work is started, the MMIO
space may no longer be available. This will cause an oops in the worker
thread.

[  606.806583] NIP kthread_data+0x28/0x40
[  606.806633] LR wq_worker_sleeping+0x30/0x100
[  606.806694] Call Trace:
[  606.806721] 0x50 (unreliable)
[  606.806796] wq_worker_sleeping+0x30/0x100
[  606.806884] __schedule+0x69c/0x8a0
[  606.806959] schedule+0x44/0xc0
[  606.807034] do_exit+0x770/0xb90
[  606.807109] die+0x300/0x460
[  606.807185] bad_page_fault+0xd8/0x150
[  606.807259] handle_page_fault+0x2c/0x30
[  606.807338] wait_port_offline.constprop.12+0x60/0x130 [cxlflash]

To prevent the problem space area from being unmapped, when there is
pending work, a mapcount (using the kref mechanism) is held.  The
mapcount is released only when the work is completed.  The last
reference release is tied to the unmapping service.

Signed-off-by: Manoj N. Kumar <manoj@linux.vnet.ibm.com>
Acked-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Reviewed-by: Uma Krishnan <ukrishn@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agocxlflash: Fix to resolve cmd leak after host reset
Manoj Kumar [Mon, 14 Dec 2015 21:07:02 +0000 (15:07 -0600)]
cxlflash: Fix to resolve cmd leak after host reset

After a few iterations of resetting the card, either during EEH
recovery, or a host_reset the following is seen in the logs.  cxlflash
0008:00: cxlflash_queuecommand: could not get a free command

At every reset of the card, the commands that are outstanding are being
leaked.  No effort is being made to reap these commands.  A few more
resets later, the above error message floods the logs and the card is
rendered totally unusable as no free commands are available.

Iterated through the 'cmd' queue and printed out the 'free' counter and
found that on each reset certain commands were in-use and stayed in-use
through subsequent resets.

To resolve this issue, when the card is reset, reap all the commands
that are active/outstanding.

Signed-off-by: Manoj N. Kumar <manoj@linux.vnet.ibm.com>
Acked-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agocxlflash: Removed driver date print
Uma Krishnan [Mon, 14 Dec 2015 21:06:33 +0000 (15:06 -0600)]
cxlflash: Removed driver date print

Having a date for the driver requires it to be updated quite
often. Removing the date which is not necessary. Also made
use of the existing symbol to print the driver name.

Signed-off-by: Uma Krishnan <ukrishn@linux.vnet.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Acked-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agocxlflash: Fix to avoid virtual LUN failover failure
Matthew R. Ochs [Mon, 14 Dec 2015 20:55:44 +0000 (14:55 -0600)]
cxlflash: Fix to avoid virtual LUN failover failure

Applications which use virtual LUN's that are backed by a physical LUN
over both adapter ports may experience an I/O failure in the event of a
link loss (e.g. cable pull).

Virtual LUNs may be accessed through one or both ports of the adapter.
This access is encoded in the translation entries that comprise the
virtual LUN and used by the AFU for load-balancing I/O and handling
failover scenarios. In a link loss scenario, even though the AFU is able
to maintain connectivity to the LUN, it is up to the application to
retry the failed I/O. When applications are unaware of the virtual LUN's
underlying topology, they are unable to make a sound decision of when to
retry an I/O and therefore are forced to make their reaction to a failed
I/O absolute. The result is either a failure to retry I/O or increased
latency for scenarios where a retry is pointless.

To remedy this scenario, provide feedback back to the application on
virtual LUN creation as to which ports the LUN may be accessed. LUN's
spanning both ports are candidates for a retry in a presence of an I/O
failure.

Signed-off-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Acked-by: Manoj Kumar <manoj@linux.vnet.ibm.com>
Reviewed-by: Uma Krishnan <ukrishn@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agocxlflash: Fix to escalate LINK_RESET also on port 1
Manoj Kumar [Mon, 14 Dec 2015 20:55:09 +0000 (14:55 -0600)]
cxlflash: Fix to escalate LINK_RESET also on port 1

The original fix to escalate a 'login timed out' error to a LINK_RESET
was only made for one of the two ports on the card. This fix resolves
the same issue for the second port (port 1).

Signed-off-by: Manoj N. Kumar <manoj@linux.vnet.ibm.com>
Acked-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Reviewed-by: Uma Krishnan <ukrishn@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agostorvsc: Tighten up the interrupt path
K. Y. Srinivasan [Wed, 23 Dec 2015 21:15:51 +0000 (13:15 -0800)]
storvsc: Tighten up the interrupt path

On the interrupt path, we repeatedly establish the pointer to the
storvsc_device. While the compiler does inline get_in_stor_device() (and
other static functions) in the call chain in the interrupt path, the
compiler is repeatedly inlining the call to get_in_stor_device() each
time it is invoked.  The return value of get_in_stor_device() can be
cached in the interrupt path since there is higher level serialization
in place to ensure correct handling when the module unload races with
the processing of an incoming message from the host.  Optimize this code
path by caching the pointer to storvsc_device and passing it as an
argument.

Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Long Li <longli@microsoft.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Alex Ng <alexng@microsoft.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agostorvsc: Refactor the code in storvsc_channel_init()
K. Y. Srinivasan [Wed, 23 Dec 2015 21:15:50 +0000 (13:15 -0800)]
storvsc: Refactor the code in storvsc_channel_init()

The function storvsc_channel_init() repeatedly interacts with the host
to extract various channel properties. Refactor this code to eliminate
code repetition.

Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Long Li <longli@microsoft.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Alex Ng <alexng@microsoft.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agostorvsc: Properly support Fibre Channel devices
K. Y. Srinivasan [Wed, 23 Dec 2015 21:15:49 +0000 (13:15 -0800)]
storvsc: Properly support Fibre Channel devices

For FC devices managed by this driver, atttach the appropriate transport
template. This will allow us to create the appropriate sysfs files for
these devices. With this we can publish the wwn for both the port and the node.

Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Long Li <longli@microsoft.com>
Tested-by: Alex Ng <alexng@microsoft.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
8 years agostorvsc: Fix a bug in the layout of the hv_fc_wwn_packet
K. Y. Srinivasan [Wed, 23 Dec 2015 21:15:48 +0000 (13:15 -0800)]
storvsc: Fix a bug in the layout of the hv_fc_wwn_packet

The hv_fc_wwn_packet is exchanged over vmbus. Make the definition in
Linux match the Windows definition.

Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Long Li <longli@microsoft.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Alex Ng <alexng@microsoft.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>