table. Previously, the ddb code knew of each linker set of auxiliary
commands and which explicit command list they were tied to. These changes
add a simple command_table struct that contains both the static list of
commands and the pointers for any auxiliary linker set of additional
commands. This also makes it possible for other arbitrary command tables
to be defined in other parts of the kernel w/o having to edit ddb itself.
The DB_SET macro has also been trimmed down to just creating an entry in
a linker set. A new DB_FUNC macro does what the old DB_SET did which is
to not only add an entry to the linker set but also to include a function
prototype for the function being added. With these changes, it's now also
possible to create aliases for ddb functions using DB_SET() directly if
desired.
buildkernel: provide a real but dummy name to ${DEPENDFILE}
so that the relevant exists() check in bsd.prog.mk fails and
ensures that ${GENHDRS} are built before any other objects.
MFC after: 3 days
chips where setting the FAILDIS bit is not effective. While here,
try again to make it clear that reported parity errors indicate
a failure of some PCI device *other than* the aic7xxx controller.
timer reset rather than the timer of an SCB still pending on the
controller after recovery completed. This should correct timeout
loops seen in the field.
over iteration of their multicast address lists when synchronizing the
hardware address filter with the network stack-maintained list.
Problem reported by: Ed Maste (emaste at phaedrus dot sandvine dot ca>
MFC after: 1 week
aic7xxx.c:
Allow print_reg() to be called with a NULL column.
aic79xx.c:
Correct new usage of SCB_GET_TAG().
aic7xxx.c:
Fix stray ahd that snuck in here.
close holes in detecting busfrees that occur after a packetized target
transitions to a non-packetized phase. The most common case where this
occurs is when a target is externally reset so the controller believes
a packetzied negotiation agreement is still in effect. Unfortunately,
disabling this feature seems to cause problems for the 7901B. Re-enable
ehanced busfree detection for this part until I can get my hands on a
samble to figure out if the old workaround is necessary and, if so, how
to make it work correctly.
to us was to help out the Linux port, but really just invited overflow.
In fact, the request sense timer was overflowing prior to this change making
it much shorter than intended.
aic_osm_lib.h:
Be more careful about overflow in all timer/timeout primitives.
Add constants for SPI protocol delays that are needed for
target mode.
aic7xxx.c:
Correct a target mode issue that caused an occassional
spurious REQ to be seen on the bus when performing manual
message processing (e.g. transfer rate negotiation).
Enforce phase change bus settle rules with explicit
delays when performing manual message processing in
target mode. The sequencer already did this for
"fast-path", target mode message processing.
the driver to issue a bus reset more quickly than intended. We want to
*wait* if we find another SCB that could be the cause of this timeout,
not proceed to a bus reset.
Noticed by: kan
The ISA probe uses an identify routine to probe all slot locations from
1 to 14 that do not conflict with other allocated resources. This required
making aic7770.c part of the driver core when compiled as a module.
aic7xxx.c:
aic79xx.c:
aic_osm_lib.c:
Use aic_scb_timer_start() consistently to start the watchdog timer.
This removes a few places that verbatum copied the code in
aic_scb_timer_start().
During recovery processing, allow commands to still be queued to
the controller. The only requirement we have is that our recovery
command be queued first - something the code already guaranteed.
The only other change required to make this work is to prevent
timers from being started for these newly queued commands.
Approved by: re
code was adjusting twice for the instruction pointer indicating
the *next* instruction to execute. The aic79xx driver had a similar
bug, but was fixed some time ago.
by a transaction performing a driver handled message sequence (an
scb with the MK_MESSAGE flag set).
SCBs that perform host managed messaging must always be
at the head of their per-target selection queue so that
the firmware knows to manually assert ATN if the current
negotiation agreement is packetized. In the past we
guaranteed this by queuing these SCBs separarately in
the execution queue. This exposes the system to potential
command reordering in two cases:
1) Another SCB for the same ITL nexus is queued that does
not have the MK_MESSAGE flag set. This SCB will be
queued to the per-target list which can be serviced
before the MK_MESSAGE scb that preceeded it.
2) If the target cannot accept all of the commands in the
per-target selection queue in one selection, the remainder
is queued to the tail of the selection queues so as to
effect round-robin scheduling. This could allow the
MK_MESSAGE scb to be sent to the target before the
requeued commands.
This commit changes the firmware policy to defer queuing
MK_MESSAGE SCBs into the selection queues until this can
be done without affecting order. This means that the
target's selection queue is either empty, or the last
SCB on the execution queue is also a MK_MESSAGE SCB.
During any wait, the firmware halts the download of new
SCBs so only a single "holding location" is required.
Luckily, MK_MESSAGE SCBs are rare and typically occur only
during CAM's bus probe where only one command is outstanding
at a time. However, during some recovery scenarios, the
reordering *could* occur.
aic79xx.c:
Update ahd_search_qinfifo() and helper routines to
search for pending MK_MESSAGE scbs and properly
restitch the execution queue if either the MK_MESSAGE
SCB is being aborted, or the MK_MESSAGE SCB can be
queued due to the execution queue draining due to
aborts.
Enable LQOBUSFREE status to assert an interrupt.
This should be redundant since a BUSFREE interrupt
should always occur along with an LQOBUSFREE event,
but on the Rev A, this doesn't seem to be guaranteed.
When a PPR request is rejected when a previously
existing packetized agreement is in place, assume
that the target has been reset without our knowledge
and revert to async/narrow transfers. This corrects
two issues: the stale ENATNO setting that was used
to send the PPR is cleared so the firmware is not
confused by a future packetized selection with
ATN asserted but no MK_MESSAGE flag in the SCB and
it speeds up recovery by aborting any pending
packetized transactions that by definition are now
dead.
When re-queueing SCBs after a failed negotiation
attempt, ensure command ordering by freezing the
device queue first.
Traverse the list of pending SCBs rather than the
whole SCB array on the controller when pushing
MK_MESSAGE flag changes out to the controller.
The original code was optimized for the aic7xxx
controllers where there are fewer controller slots
then pending SCBs and the firmware picks SCB
slots. For the U320 controller, the hope is
that we have fewer pending SCBs then the 512
slots on the controller.
Enhance some diagnostics.
Factor out some common code.
aic79xx.h:
Add prototype for new ahd_done_with_status() that is
used to factor out some commone code.
aic79xx.reg:
Add definisions for the pending MK_MESSAGE SCB.
aic79xx.seq:
Defer MK_MESSAGE SCB queing to the execution queue
so as to preserve command ordering. Re-arrange some
of the selection processing code so the above change
had no performance impact on the common code path.
Close a few critical section holes.
When entering a non-packetized phase, manually enable
busfree interrupts, since the controller hardware
does not do this automatically.
aic79xx_inline.h:
Enhance logging for queued SCBs.
aic79xx_osm.c:
Add new a new DDB ahd command, ahd_dump, which
invokes the ahd_dump_card_state() routine on the
unit specified with the ahd_sunit DDB command.
aic79xx_pci.c:
Turn on the BUSFREEREV bug for the Rev B. controller.
This is required to close the busfree during non-packetized
phase hole.
for unknown events.
A number of modules return EINVAL in this instance, and I have left
those alone for now and instead taught MOD_QUIESCE to accept this
as "didn't do anything".
Allow 500us between pauses in ahd_pause_and_flushwork().
The maximum we will wait is now 500ms.
In the same routine, remove any attempt to clear ENSELO.
Let the firmware do it once the current selection has
completed. This avoids some race conditions having to
do with non-packetized completions and the auto-clearing
of ENSELO on packetized completions.
Also avoid attempts to clear critical sections when
interrups are pending. We are going to loop again
anyway, so clearing critical sections is a waste of
time. It also may not be possible to clear a critical
section if the source of the interrupt was a SEQINT.
aic79xx_pci.c:
Use the Generic 9005 mask when looking for generic 7901B
parts. This allows the driver to attach to 7901B parts
on motherboards using a non-Adaptec subvendor ID.
aic79xx_inline.h:
Test for the SCBRAM_RD_BUG against the bugs
field, not the flags field in the softc.
aic79xx.c:
Cancel pending transactions on devices that
respond with a selection timeout. This decreases
the duration of timeout recovery when a device
disappears.
aic79xx.c:
Don't bother forcing renegotiation on a selection
timeout now that we use the device reset handler
to abort any pending commands on the target.
The device reset handler already takes us down
to async narrow and forces a renegotiation.
In the device reset handlers, only send a
BDR sent async event if the status is not
CAM_SEL_TIMEOUT. This avoids sending this
event in the selection timeout case
aic79xx.c:
Modify the Core timeout handler to verify that another
command has the potential to timeout before passing off
a command timeout as due to some other command. This
safety measure is added in response to a timeout recovery
failure on H2B where it appears that incoming reselection
status was lost during a drive pull test. In that case,
the recovery handler continued to wait for the command
that was active on the bus indefinetly. While the root
cause of the above issue is still being determined seems
a prudent safeguard.
aic79xx_pci.c:
Add a specific probe entry for the Dell OEM 39320(B).
aic79xx.c:
aic79xx.h:
aic79xx.reg:
aic79xx.seq:
Modify the aic79xx firmware to never cross a cacheline or
ADB boundary when DMA'ing completion entries to the host.
In PCI mode, at least in 32/33 configurations, the SCB
DMA engine may lose its place in the data-stream should
the target force a retry on something other than an
8byte aligned boundary. In PCI-X mode, we do this to
avoid split transactions since many chipsets seem to be
unable to format proper split completions to continue
the data transfer.
The above change allows us to drop our completion entries
from 8 bytes to 4. We were using 8 byte entries to ensure
that PCI retries could only occur on an 8byte aligned
boundary. Now that the sequencer guarantees this by splitting
up completions, we can safely drop the size to 4 bytes (2
byte tag, one byte SG_RESID, one byte pad).
Both the split-completion and PCI retry problems only show
up under high tag load when interrupt coalescing is being
especially effective. The switch from a 2byte completion
entry to an 8 byte entry to solve the PCI problem increased
the chance of incurring a split in PCI-X mode when multiple
transactions were completed at once. Dropping the completion
size to 4 bytes also means that we can complete more commands
in a single DMA (128byte FIFO -> 32 commands instead of 16).
aic79xx.c:
Modify the SCSIINT handler to defer clearing
sequencer critical sections to the individual
interrupt handlers. This allows us to
immediately disable any outgoing selections in
the case of an unexpected busfree so we don't
inadvertantly clear ENSELO *after* a new selection
has started. Doing so may cause the sequencer
to miss a successful selection.
In ahd_update_pending_scbs(), only clear ENSELO if
the bus is currently busy and a selection is not
already in progress or the sequencer has yet to
handle a pending selection. While we want to ensure
that the selection for the SCB at the head of the
selection queue is restarted so that any change in
negotiation request can take effect, we can't clobber
pending selection state without confusing the sequencer
into missing a selection.
sequencer interrupt codes. These codes are only
relevant to the code that was last being executed
and that context is cleared when we reset the
program counter. This addresses a race condition
between a sequencer interrupt and any SCSI event
that causes us to restart the sequencer.
o When running the untagged-Q, we must start the
timer for any transaction we queue.
o Give the firmware half a millisecond between
pauses to flush work out. This should give us
around half a second of total delay before flagging
an issue with pausing and flushing controller work.
Only attempt to clear critical sections if there
are no pending interrupts in the pause and flush
loop. If the sequencer has issued an INTSTAT, we
may not be able to step out of the critical section.
o Cancel pending transactions on devices that
respond with a selection timeout. This decreases
the duration of timeout recovery when a device
disappears.
Don't bother forcing renegotiation on a selection
timeout now that we use the device reset handler
to abort any pending commands on the target.
The device reset handler already takes us down
to async narrow and forces a renegotiation.
o In the device reset handlers, only send a
BDR sent async event if the status is not
CAM_SEL_TIMEOUT. This avoids sending this
event in the selection timeout case.
o Modify the Core timeout handler to verify that another
command has the potential to timeout before passing off
a command timeout as due to some other command.
to build the kernel. It doesn't affect the operation if gcc.
Most of the changes are just adding __INTEL_COMPILER to #ifdef's, as
icc v8 may define __GNUC__ some parts may look strange but are
necessary.
Additional changes:
- in_cksum.[ch]:
* use a generic C version instead of the assembly version in the !gcc
case (ASM code breaks with the optimizations icc does)
-> no bad checksums with an icc compiled kernel
Help from: andre, grehan, das
Stolen from: alpha version via ppc version
The entire checksum code should IMHO be replaced with the DragonFly
version (because it isn't guaranteed future revisions of gcc will
include similar optimizations) as in:
---snip---
Revision Changes Path
1.12 +1 -0 src/sys/conf/files.i386
1.4 +142 -558 src/sys/i386/i386/in_cksum.c
1.5 +33 -69 src/sys/i386/include/in_cksum.h
1.5 +2 -0 src/sys/netinet/igmp.c
1.6 +0 -1 src/sys/netinet/in.h
1.6 +2 -0 src/sys/netinet/ip_icmp.c
1.4 +3 -4 src/contrib/ipfilter/ip_compat.h
1.3 +1 -2 src/sbin/natd/icmp.c
1.4 +0 -1 src/sbin/natd/natd.c
1.48 +1 -0 src/sys/conf/files
1.2 +0 -1 src/sys/conf/files.amd64
1.13 +0 -1 src/sys/conf/files.i386
1.5 +0 -1 src/sys/conf/files.pc98
1.7 +1 -1 src/sys/contrib/ipfilter/netinet/fil.c
1.10 +2 -3 src/sys/contrib/ipfilter/netinet/ip_compat.h
1.10 +1 -1 src/sys/contrib/ipfilter/netinet/ip_fil.c
1.7 +1 -1 src/sys/dev/netif/txp/if_txp.c
1.7 +1 -1 src/sys/net/ip_mroute/ip_mroute.c
1.7 +1 -2 src/sys/net/ipfw/ip_fw2.c
1.6 +1 -2 src/sys/netinet/igmp.c
1.4 +158 -116 src/sys/netinet/in_cksum.c
1.6 +1 -1 src/sys/netinet/ip_gre.c
1.7 +1 -2 src/sys/netinet/ip_icmp.c
1.10 +1 -1 src/sys/netinet/ip_input.c
1.10 +1 -2 src/sys/netinet/ip_output.c
1.13 +1 -2 src/sys/netinet/tcp_input.c
1.9 +1 -2 src/sys/netinet/tcp_output.c
1.10 +1 -1 src/sys/netinet/tcp_subr.c
1.10 +1 -1 src/sys/netinet/tcp_syncache.c
1.9 +1 -2 src/sys/netinet/udp_usrreq.c
1.5 +1 -2 src/sys/netinet6/ipsec.c
1.5 +1 -2 src/sys/netproto/ipsec/ipsec.c
1.5 +1 -1 src/sys/netproto/ipsec/ipsec_input.c
1.4 +1 -2 src/sys/netproto/ipsec/ipsec_output.c
and finally remove
sys/i386/i386 in_cksum.c
sys/i386/include in_cksum.h
---snip---
- endian.h:
* DTRT in C++ mode
- quad.h:
* we don't use gcc v1 anymore, remove support for it
Suggested by: bde (long ago)
- assym.h:
* avoid zero-length arrays (remove dependency on a gcc specific
feature)
This change changes the contents of the object file, but as it's
only used to generate some values for a header, and the generator
knows how to handle this, there's no impact in the gcc case.
Explained by: bde
Submitted by: Marius Strobl <marius@alchemy.franken.de>
- aicasm.c:
* minor change to teach it about the way icc spells "-nostdinc"
Not approved by: gibbs (no reply to my mail)
- bump __FreeBSD_version (lang/icc needs to know about the changes)
Incarnations of this patch survive gcc compiles since a loooong time,
I use it on my desktop. An icc compiled kernel works since Nov. 2003
(exceptions: snd_* if used as modules), it survives a build of the
entire ports collection with icc.
Parts of this commit contains suggestions or submissions from
Marius Strobl <marius@alchemy.franken.de>.
Reviewed by: -arch
Submitted by: netchild
Intel C/C++ compiler (lang/icc) to build the kernel.
The icc CPUTYPE CFLAGS use icc v7 syntax, icc v8 moans about them, but
doesn't abort. They also produce CPU specific code (new instructions
of the CPU, not only CPU specific scheduling), so if you get coredumps
with signal 4 (SIGILL, illegal instruction) you've used the wrong
CPUTYPE.
Incarnations of this patch survive gcc compiles and my make universe.
I use it on my desktop.
To use it update share/mk, add
/usr/local/intel/compiler70/ia32/bin (icc v7, works)
or
/usr/local/intel_cc_80/bin (icc v8, doesn't work)
to your PATH, make sure you have a new kernel compile directory
(e.g. MYKERNEL_icc) and run
CFLAGS="-O2 -ip" CC=icc make depend
CFLAGS="-O2 -ip" CC=icc make
in it.
Don't compile with -ipo, the build infrastructure uses ld directly to
link the kernel and the modules, but -ipo needs the link step to be
performed with Intel's linker.
Problems with icc v8:
- panic: npx0 cannot be emulated on an SMP system
- UP: first start of /bin/sh results in a FP exception
Parts of this commit contains suggestions or submissions from
Marius Strobl <marius@alchemy.franken.de>.
Reviewed by: silence on -arch
Submitted by: netchild
aic79xx.seq:
Convert the COMPLETE_DMA_SCB list to an "stailq". This allows us to
safely keep the SCB that is currently being DMA'ed back the host on
the head of the list while processing completions off of the bus. The
newly completed SCBs are appended to the tail of the queue. In the
past, we just dequeued the SCB that was in flight from the list, but
this could result in a lost completion should the host perform certain
types of error recovery that must cancel all in-flight SCB DMA operations.
Switch from using a 16bit completion entry, holding just the tag and the
completion valid bit, to a 64bit completion entry that also contains a
"status packet valid" indicator. This solves two problems:
o The SCB DMA engine on at least Rev B. silicon does not properly deal
with a PCI disconnect that occurs at a non-64bit aligned offset in the
chips "source buffer". When the transfer is resumed, the DMA engine
continues at the correct offset, but may wrap to the head of the buffer
causing duplicate completions to be reported to the host. By using a
completion buffer in host memory that is 64bit aligned and using 64bit
completion entries, such disconnects should only occur at aligned addresses.
This assumes that the host bridge will only disconnect on cache-line
boundaries and that cache-lines are multpiles of 64bits.
o By embedding the status information in the completion entry we can avoid
an extra memory reference to the HSCB for commands that complete without
error.
Use the comparison of a "host freeze count" and a "sequencer freeze count"
to allow the host to process most SCBs that complete with non-zero status
without having to clear critical sections. Instead the host can just pause the
sequencer, performs any necessary cleanup in the waiting for selection list,
increments its freeze count on the controller, and unpauses. This is only
possible because the sequencer defers completions of SCBs with bad status
until after all pending selections have completed. The sequencer then avoids
referencing any data structures the host may touch during completion of the
SCB until the freeze counts match.
aic79xx.c:
Change the strategy for allocating our sentinal HSCB for the QINFIFO. In
the past, this allocation was tacked onto the QOUTFIFO allocation. Now that
the qoutfifo has grown to accomodate larger completion entries, the old
approach will result in a 64byte allocation that costs an extra page of
coherent memory. We now do this extra allocation via ahd_alloc_scbs()
where the "unused space" can be used to allocate "normal" HSCBs.
In our packetized busfree handler, use the ENSELO bit to differentiate
between packetized and non-packetized unexpected busfree events that
occur just after selection, but before the sequencer has had the oportunity
to service the selection.
When cleaning out the waiting for selection list, use the SCSI mode
instead of the command channel mode. The SCB pointer in the command
channel mode may be referenced by the SCB dma engine even while the
sequencer is paused, whereas the SCSI mode SCB pointer is only accessed
by the sequencer.
Print the "complete on qfreeze" sequencer SCB completion list in
ahd_dump_card_state(). This list holds all SCB completions that are deferred
until a pending select-out qfreeze event has taken effect.
aic79xx.h:
Add definitions and structures to handle the new SCB completion scheme.
Add a controller flag that indicates if the controller is in HostRAID
mode.
aic79xx.reg:
Remove macros used for toggling from one data fifo mode to the other.
They have not been in use for some time.
Add scratch ram fields for our new qfreeze count scheme, converting
the complete dma list into an "stailq", and providing for the "complete
on qfreeze" SCB completion list. Some other fields were moved to retain
proper field alignment (alignment >= field size in bytes).
aic79xx.seq:
Add code to our idle loop to:
o Process deferred completions once a qfreeze event has taken full
effect.
o Thaw the queue once the sequencer and host qfreeze counts match.
Generate 64bit completion entries passing the SCB_SGPTR field as the
"good status" indicator. The first bit in this field is only set if
we have a valid status packet to send to the host.
Convert the COMPLETE_DMA_SCB list to an "stailq".
When using "setjmp" to register an idle loop handler, do not combine
the "ret" with the block move to pop the stack address in the same
instruction. At least on the A, this results in a return to the setjmp
caller, not to the new address at the top of the stack. Since we want
the latter (we want the newly registered handler to only be invoked from
the idle loop), we must use a separate ret instruction.
Add a few missing critical sections.
Close a race condition that can occur on Rev A. silicon. If both FIFOs
happen to be allocated before the sequencer has a chance to service the
FIFO that was allocated first, we must take special care to service the
FIFO that is not active on the SCSI bus first. This guarantees that a
FIFO will be freed to handle any snapshot requests for the FIFO that is
still on the bus. Chosing the incorrect FIFO will result in deadlock.
Update comments.
aic79xx_inline.h
Correct the offset calculation for the syncing of our qoutfifo.
Update ahd_check_cmdcmpltqueues() for the larger completion entries.
aic79xx_pci.c:
Attach to HostRAID controllers by default. In the future I may add a
sysctl to modify the behavior, but since FreeBSD does not have any
HostRAID drivers, failing to attach just results in more email and
bug reports for the author.
MFC After: 1week
to each other.
Correct the recovery thread's loop so that it
will terminate properly on shutdown. We also
clear the recovery_thread proc pointer so that
any additional calls to aic_terminate_recovery_thread()
will not attempt to kill a thread that doesn't
exist. Lastly, code the loop so that termination
will still be successfull even if the termination
request occurs just prior to us entering the loop
or while the recovery thread is off recovering
commands.
