Fix the following unlikely but possible race:
CPU 1 CPU 2
------------------------------------------------------------------------
AR-request tasklet
lookup handler
unregister handler
free handler->callback_data or handler
call handler->callback
The application which registered the handler has no way to stop nodes
sending new requests to their address range, hence cannot prevent this
race.
Fix it simply by extending the address_handler_lock-protected region
from only around the lookup to around both lookup and call. We only
need to do so in the exclusive region handler; the FCP region handler
already holds the lock around the handler->callback call.
Alas this removes the current ability to execute the callback in
parallel on different CPUs if it was called for different FireWire cards
at the same time. (For a single card, the handler is already
serialized.) If this loss of a rather obscure feature is not tolerable,
a more complex fix would be required: Add a handler reference counter;
wait in fw_core_remove_address_handler() for this conter to become zero.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Target-like applications or peer-to-peer-like applications require the
global address handler registration which we have right now, or a per-
card registration. And node lookup, while it would be nice to have,
would be impossible in the brief time between self-ID-complete event and
completion of firewire-core's topology scanning.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Associate all log messages from firewire-core with the respective card
because some people have more than one card. E.g.
firewire_ohci 0000:04:00.0: added OHCI v1.10 device as card 0, 8 IR + 8 IT contexts, quirks 0x0
firewire_ohci 0000:05:00.0: added OHCI v1.10 device as card 1, 8 IR + 8 IT contexts, quirks 0x0
firewire_core: created device fw0: GUID 0814438400000389, S800
firewire_core: phy config: new root=ffc1, gap_count=5
firewire_core: created device fw1: GUID 0814438400000388, S800
firewire_core: created device fw2: GUID 0001d202e06800d1, S800
turns into
firewire_ohci 0000:04:00.0: added OHCI v1.10 device as card 0, 8 IR + 8 IT contexts, quirks 0x0
firewire_ohci 0000:05:00.0: added OHCI v1.10 device as card 1, 8 IR + 8 IT contexts, quirks 0x0
firewire_core 0000:04:00.0: created device fw0: GUID 0814438400000389, S800
firewire_core 0000:04:00.0: phy config: new root=ffc1, gap_count=5
firewire_core 0000:05:00.0: created device fw1: GUID 0814438400000388, S800
firewire_core 0000:04:00.0: created device fw2: GUID 0001d202e06800d1, S800
This increases the module size slightly; to keep this in check, turn the
former printk wrapper macros into functions. Their implementation is
largely copied from driver core's dev_printk counterparts.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Change the log line prefix from "firewire_net: " to "net firewire0: "
etc. for the case that several RFC 2734 interfaces are being used in the
same machine.
Note, the netdev_printk API is not very useful to firewire-net.
netdev_notice(net, "abc\n") would result in irritating messages like
"firewire_ohci 0000:0a:00.0: firewire0: abc". Nor would a dev_printk on
the fw_unit.device to which firewire-net is being bound be useful,
because there are generally multiple ones of those per interface (from
all RFC 2734 peers on the bus, the local node being only one of them).
In the initialization message of each interface, log the PCI device
name of the card which is parent of the netdevice instead of the GUID
of the peer which was semi-randomly used to establish the netdevice.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
On second thought, there is little reason to have driver name differ
from module name. Therefore, change
/sys/bus/firewire/drivers/net
/sys/bus/firewire/devices/fw0.0/driver -> [...]/net
/sys/module/firewire_net/drivers/firewire:net
to
/sys/bus/firewire/drivers/firewire_net
/sys/bus/firewire/devices/fw0.0/driver -> [...]/firewire_net
/sys/module/firewire_net/drivers/firewire:firewire_net
It is redundant but consistent with firewire-sbp2's recently changed
driver name.
I don't see this anywhere used, so it should not matter either way.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Commit eba9ebaaa2 "firewire: sbp2: use dev_printk API" changed
messages from e.g.
firewire_sbp2: fw3.0: logged in to LUN 0000 (0 retries)
to
sbp2 fw3.0: logged in to LUN 0000 (0 retries)
because the driver calls itself as "sbp2" when registering with driver
core and with SCSI core. This is of course confusing, so switch to the
name "firewire_sbp2" for driver core in order to match what lsmod and
/sys/module/ show. So we are back to
firewire_sbp2 fw3.0: logged in to LUN 0000 (0 retries)
in the kernel log.
This also changes
/sys/bus/firewire/drivers/sbp2
/sys/bus/firewire/devices/fw3.0/driver -> [...]/sbp2
/sys/module/firewire_sbp2/drivers/firewire:sbp2
to
/sys/bus/firewire/drivers/firewire_sbp2
/sys/bus/firewire/devices/fw3.0/driver -> [...]/firewire_sbp2
/sys/module/firewire_sbp2/drivers/firewire:firewire_sbp2
but "cat /sys/class/scsi_host/host27/proc_name" stays "sbp2" just in
case that proc_name is used by any userland.
The transport detection in lsscsi is not affected. (Tested with lsscsi
version 0.25.) Udev's /dev/disk/by-id and by-path symlinks are not
affected either.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
fw_unit device drivers invariably need to talk to the fw_unit's parent
(an fw_device) and grandparent (an fw_card). firewire-core already
maintains an fw_card reference for the entire lifetime of an fw_device.
Likewise, let firewire-core maintain an fw_device reference for the
entire lifetime of an fw_unit so that fw_unit drivers don't have to.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Add the dma_sync_single_* calls necessary to ensure proper cache
synchronization for isochronous data buffers on non-coherent
architectures.
Signed-off-by: Clemens Ladisch <clemens@ladisch.de>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
If a device's firmware initiates a bus reset by setting the IBR bit in
PHY register 1 without resetting the gap count field to 63 (and without
having sent a PHY configuration packet beforehand), the gap count of
this node will remain at the old value after the bus reset and thus be
inconsistent with the gap count on all other nodes.
The bus manager is supposed to detect the inconsistent gap count values
in the self ID packets and correct them by issuing another bus reset.
However, if the buggy device happens to be the cycle master, and if it
sends a cycle start packet immediately after the bus reset (which is
likely after a long bus reset), then the time between the end of the
selfID phase and the start of the cycle start packet will be based on
the too-small gap count value, so this gap will be too short to be
detected as a subaction gap by the other nodes. This means that the
cycle start packet will be assumed to be self ID data, and will be
stored after the actual self ID quadlets in the self ID buffer.
This garbage in the self ID buffer made firewire-core ignore all of the
self ID data, and thus prevented the Linux bus manager from correcting
the problem. Furthermore, because the bus reset handling was aborted
completely, asynchronous transfers would be no longer handled correctly,
and fw_run_transaction() would hang until the next bus reset.
To fix this, make the detection of inconsistent self IDs more
discriminating: If the invalid data in the self ID buffer looks like
a cycle start packet, we can assume that the previous data in the buffer
is correctly received self ID information, and process it normally.
(We inspect only the first quadlet of the cycle start packet, because
this value is different enough from any valid self ID quadlet, and many
controllers do not store the cycle start packet in five quadlets because
they expect self ID data to have an even number of quadlets.)
This bug has been observed when a bus-powered DesktopKonnekt6 is
switched off with its power button.
Signed-off-by: Clemens Ladisch <clemens@ladisch.de>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Change memory region to ohci "middle address space". This effectively
reduces the number of packets by 50%.
[Stefan R.:] This eliminates 1394 ack packets and improved throughput
by a few percent in some tests with an S400a connection with and without
gap count optimization. Since firewire-net taxes the AR-req DMA unit of
a FireWire controller much more than firewire-sbp2 (which uses the
middle address space with PCI posted writes too), this commit also
changes a related error printk into a ratelimited one as a precaution.
Side note: The IPv4-over-1394 drivers of Mac OS X 10.4, Windows XP SP3,
and the Thesycon 1394 bus driver for Windows all use the middle address
space too.
Signed-off-by: Stephan Gatzka <stephan@gatzka.org>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Use kernel.h's convenience macros. Also omit a printk that should never
happen and won't matter much if it ever happened.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Takes less source code and machine code, and less runtime with PHYs
other than TSB41BA3D (e.g. TSB81BA3 with device ID 0x831304 which takes
one instead of six read_paged_phy_reg now).
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Fix: phy_reg_mutex must be held over the write/read_phy_reg pair which
gets PHY port status.
Only print to the log when a TSB41BA3D was found. By far most TSB82AA2
cards have a TSB81BA3, and firewire-ohci can keep quiet about that.
Shorten some strings and comments. Change some whitespace.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
This patch implements a work around for the Texas Instruments PHY
TSB41BA3D. This phy has a bug at least in combination with the TI LLCs
TSB82AA2B and TSB12LV26. The selfid coming from the locally connected
phy is not propagated into the selfid buffer of the OHCI (see
http://www.ti.com/litv/pdf/sllz059 for details). The main idea is to
construct the selfid ourselves.
Signed-off-by: Stephan Gatzka <stephan@gatzka.org>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Code inside bus_reset_work may now sleep. This is a prerequisite to
support a phy from Texas Instruments cleanly. The patch to support this
phy will be submitted later.
Signed-off-by: Stephan Gatzka <stephan@gatzka.org>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
sbp2_release_target() is folded into its primary user, sbp2_remove().
The only other caller, a failure path in sbp2_probe(), now uses
sbp2_remove(). This adds unnecessary cancel_delayed_work_sync() calls
to that failure path but results in less code and text.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Implement sbp2_queue_work(), which is now a very simple accessor to one
of the struct sbp2_logical_unit members, right after the definition of
struct sbp2_logical_unit.
Put the sbp2_reconnect() implementation right after the sbp2_login()
implementation. They are both part of the SBP-2 access protocol.
Implement the driver methods sbp2_probe(), spp2_update(), sbp2_remove()
in this order, reflecting the lifetime of an SBP-2 target.
Place the sbp2_release_target() implementation right next to
sbp2_remove() which is its primary user, and after sbp2_probe() which is
the counterpart to sbp2_release_target().
There are no changes to the implementations here, or at least not meant
to be.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Since commit 0278ccd9d5 "firewire: sbp2:
fix panic after rmmod with slow targets", the lifetime of an sbp2_target
instance does no longer extent past the return of sbp2_remove().
Therefore it is no longer necessary to call fw_unit_get/put() and
fw_device_get/put() in sbp2_probe/remove().
Furthermore, said commit also ensures that lu->work is not going to be
executed or requeued at a time when the sbp2_target is no longer in use.
Hence there is no need for sbp2_target reference counting for lu->work.
Other concurrent contexts:
- Processes which access the sysfs of the SCSI host device or of one
of its subdevices are safe because these interfaces are all removed
by scsi_remove_device/host() in sbp2_release_target().
- SBP-2 command block ORB transactions are finished when
scsi_remove_device() in sbp2_release_target() returns.
- SBP-2 management ORB transactions are finished when
cancel_delayed_work_sync(&lu->work) before sbp2_release_target()
returns.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
This fixes https://bugs.launchpad.net/ubuntu/+source/linux/+bug/801719 .
An O2Micro PCI Express FireWire controller,
"FireWire (IEEE 1394) [0c00]: O2 Micro, Inc. Device [1217:11f7] (rev 05)"
which is a combination device together with an SDHCI controller and some
sort of storage controller, misses SBP-2 status writes from an attached
FireWire HDD. This problem goes away if MSI is disabled for this
FireWire controller.
The device reportedly does not require QUIRK_CYCLE_TIMER.
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de> (amended changelog)
Cc: <stable@kernel.org>
If firewire-sbp2 starts a login to a target that doesn't complete ORBs
in a timely manner (and has to retry the login), and the module is
removed before the operation times out, you end up with a null-pointer
dereference and a kernel panic.
[SR: This happens because sbp2_target_get/put() do not maintain
module references. scsi_device_get/put() do, but at occasions like
Chris describes one, nobody holds a reference to an SBP-2 sdev.]
This patch cancels pending work for each unit in sbp2_remove(), which
hopefully means there are no extra references around that prevent us
from unloading. This fixes my crash.
Signed-off-by: Chris Boot <bootc@bootc.net>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
* 'fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/ieee1394/linux1394-2.6:
firewire: ohci: fix DMA unmapping in an error path
firewire: cdev: fix 32 bit userland on 64 bit kernel compat corner cases
Some older Panasonic made camcorders (Panasonic AG-EZ30 and NV-DX110,
Grundig Scenos DLC 2000) reject requests with ack_busy_X if a request is
sent immediately after they sent a response to a prior transaction.
This causes firewire-core to fail probing of the camcorder with "giving
up on config rom for node id ...". Consequently, programs like kino or
dvgrab are unaware of the presence of a camcorder.
Such transaction failures happen also with the ieee1394 driver stack
(of the 2.4...2.6 kernel series until 2.6.36 inclusive) but with a lower
likelihood, such that kino or dvgrab are generally able to use these
camcorders via the older driver stack. The cause for firewire-ohci's or
firewire-core's worse behavior is not yet known. Gap count optimization
in firewire-core is not the cause. Perhaps the slightly higher latency
of transaction completion in the older stack plays a role. (ieee1394:
AR-resp DMA context tasklet -> packet completion ktread -> user process;
firewire-core: tasklet -> user process.)
This change introduces retries and delays after ack_busy_X into
firewire-core's Config ROM reader, such that at least firewire-core's
probing and /dev/fw* creation are successful. This still leaves the
problem that userland processes are facing transaction failures.
gscanbus's built-in retry routines deal with them successfully, but
neither kino's nor dvgrab's do ever succeed.
But at least DV capture with "dvgrab -noavc -card 0" works now. Live
video preview in kino works too, but not actual capture.
One way to prevent Configuration ROM reading failures in application
programs is to modify libraw1394 to synthesize read responses by means
of firewire-core's Configuration ROM cache. This would only leave
CMP and FCP transaction failures as a potential problem source for
applications.
Reported-and-tested-by: Thomas Seilund <tps@netmaster.dk>
Reported-and-tested-by: René Fritz <rene@colorcube.de>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
If request_irq failed, we would pass wrong arguments to
dma_free_coherent. https://bugzilla.redhat.com/show_bug.cgi?id=728185
Reported-by: Mads Kiilerich
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Clemens points out that we need to use compat_ptr() in order to safely
cast from u64 to addresses of a 32-bit usermode client.
Before, our conversion went wrong
- in practice if the client cast from pointer to integer such that
sign-extension happened, (libraw1394 and libdc1394 at least were not
doing that, IOW were not affected)
or
- in theory on s390 (which doesn't have FireWire though) and on the
tile architecture, regardless of what the client does.
The bug would usually be observed as the initial get_info ioctl failing
with "Bad address" (EFAULT).
Reported-by: Carl Karsten <carl@personnelware.com>
Reported-by: Clemens Ladisch <clemens@ladisch.de>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
This allows us to move duplicated code in <asm/atomic.h>
(atomic_inc_not_zero() for now) to <linux/atomic.h>
Signed-off-by: Arun Sharma <asharma@fb.com>
Reviewed-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: Mike Frysinger <vapier@gentoo.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Between open(2) of a /dev/fw* and the first FW_CDEV_IOC_GET_INFO
ioctl(2) on it, the kernel already queues FW_CDEV_EVENT_BUS_RESET events
to be read(2) by the client. The get_info ioctl is practically always
issued right away after open, hence this condition only occurs if the
client opens during a bus reset, especially during a rapid series of bus
resets.
The problem with this condition is twofold:
- These bus reset events carry the (as yet undocumented) @closure
value of 0. But it is not the kernel's place to choose closures;
they are privat to the client. E.g., this 0 value forced from the
kernel makes it unsafe for clients to dereference it as a pointer to
a closure object without NULL pointer check.
- It is impossible for clients to determine the relative order of bus
reset events from get_info ioctl(2) versus those from read(2),
except in one way: By comparison of closure values. Again, such a
procedure imposes complexity on clients and reduces freedom in use
of the bus reset closure.
So, change the ABI to suppress queuing of bus reset events before the
first FW_CDEV_IOC_GET_INFO ioctl was issued by the client.
Note, this ABI change cannot be version-controlled. The kernel cannot
distinguish old from new clients before the first FW_CDEV_IOC_GET_INFO
ioctl.
We will try to back-merge this change into currently maintained stable/
longterm series, and we only document the new behaviour. The old
behavior is now considered a kernel bug, which it basically is.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Cc: <stable@kernel.org>
On Jun 27 Linus Torvalds wrote:
> The correct error code for "I don't understand this ioctl" is ENOTTY.
> The naming may be odd, but you should think of that error value as a
> "unrecognized ioctl number, you're feeding me random numbers that I
> don't understand and I assume for historical reasons that you tried to
> do some tty operation on me".
[...]
> The EINVAL thing goes way back, and is a disaster. It predates Linux
> itself, as far as I can tell. You'll find lots of man-pages that have
> this line in it:
>
> EINVAL Request or argp is not valid.
>
> and it shows up in POSIX etc. And sadly, it generally shows up
> _before_ the line that says
>
> ENOTTY The specified request does not apply to the kind of object
> that the descriptor d references.
>
> so a lot of people get to the EINVAL, and never even notice the ENOTTY.
[...]
> At least glibc (and hopefully other C libraries) use a _string_ that
> makes much more sense: strerror(ENOTTY) is "Inappropriate ioctl for
> device"
So let's correct this in the <linux/firewire-cdev.h> ABI while it is
still young, relative to distributor adoption.
Side note: We return -ENOTTY not only on _IOC_TYPE or _IOC_NR mismatch,
but also on _IOC_SIZE mismatch. An ioctl with an unsupported size of
argument structure can be seen as an unsupported version of that ioctl.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Cc: <stable@kernel.org>
When firewire-ohci is bound to a Pinnacle MovieBoard, eventually a
"Register access failure" is logged and an interrupt storm or a kernel
panic happens. https://bugzilla.kernel.org/show_bug.cgi?id=36622
Until this is sorted out (if that is going to succeed at all), let's
just prevent firewire-ohci from touching these devices.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Cc: <stable@kernel.org>
The software reset in firewire-ohci's pci_remove does not have a great
prospect of success if the card was already physically removed at this
point. So let's skip the 500 ms that were spent in retries here.
Also, replace a defined constant by its open-coded value. This is not a
constant from a specification but an arbitrarily chosen retry limit. It
was only used in this single place.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Detect and handle ejection of FireWire CardBus cards in PHY register
accesses:
- The last attempt of firewire-core to reset the bus during shutdown
caused a spurious "firewire_ohci: failed to write phy reg" error
message in the log. Skip this message as well as the prior retry
loop that needlessly took 100 milliseconds.
- In the unlikely case that a PHY register was read right after card
ejection, a bogus value was obtained and possibly acted upon.
Instead, fail the read attempt.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Stopping an isochronous reception DMA context takes two loop iterations
in context_stop on several controllers (JMicron, NEC, VIA). But there
is no extra delay necessary between these two reg_read trials; the MMIO
reads themselves are slow enough. Hence bring back the behavior from
before commit dd6254e5c0 "firewire: ohci:
remove superfluous posted write flushes" on these controllers by means
of an "if (i)" condition.
Isochronous context stop is performed in preemptible contexts (and only
rarely), hence this change is of little impact. (Besides, Agere and TI
controllers always, or almost always, have the context stopped already
at the first ContextControl read.)
More important is asynchronous transmit context stop, which is performed
while local interrupts are disabled (on the two AT DMAs in
bus_reset_tasklet, i.e. after a self-ID-complete event). In my
experience with several controllers, tested with a usermode AT-request
transmitter as well as with FTP transmission over firewire-net, the AT
contexts were luckily already stopped at the first ContextControl read,
i.e. never required another MMIO read let alone mdelay. A possible
explanation for this is that the controllers which I tested perhaps stop
AT DMA before they perform the self-ID reception DMA.
But we cannot be sure about that and should keep the interrupts-disabled
busy loop as short as possible. Hence, query the ContextControl
register in 1000 udelay(10) intervals instead of 10 udelay(1000)
intervals. I understand from an estimation by Clemens Ladisch that
stopping a busy DMA context should take microseconds or at worst tens of
microseconds, not milliseconds.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
The call to flush_writes() in context_stop() is superfluous because
another register read is done immediately afterwards.
The call to flush_writes() in ar_context_run() does not need to be done
individually for each AR context, so move it to ohci_enable(). This
also makes ohci_enable() clearer because it no longer depends on a side
effect of ar_context_run() to flush its own register writes.
Finally, the setting of a context's wake bit does not need to be flushed
because neither the driver logic nor the API require the CPU to wait for
this action. This removes the last MMIO reads from the packet queueing
code paths.
Signed-off-by: Clemens Ladisch <clemens@ladisch.de>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Fixing a deprecation, replacing __attribute__((packed)) with __packed.
It was deprecated for portability, specifically to avoid GCC specific
code. See commit 82ddcb0405.
Signed-off-by: August Lilleaas <august@augustl.com>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de> (added include compiler.h)
The struct sbp2_logical_unit.work items can all be executed in parallel
but are not reentrant. Furthermore, reconnect or re-login work must be
executed in a WQ_MEM_RECLAIM workqueue.
Hence replace the old single-threaded firewire-sbp2 workqueue by a
concurrency-managed but non-reentrant workqueue with rescuer.
firewire-core already maintains one, hence use this one.
In earlier versions of this change, I observed occasional failures of
parallel INQUIRY to an Initio INIC-2430 FireWire 800 to dual IDE bridge.
More testing indicates that parallel INQUIRY is not actually a problem,
but too quick successions of logout and login + INQUIRY, e.g. a quick
sequence of cable plugout and plugin, can result in failed INQUIRY.
This does not seem to be something that should or could be addressed by
serialization.
Another dual-LU device to which I currently have access to, an
OXUF924DSB FireWire 800 to dual SATA bridge with firmware from MacPower,
has been successfully tested with this too.
This change is beneficial to environments with two or more FireWire
storage devices, especially if they are located on the same bus.
Management tasks that should be performed as soon and as quickly as
possible, especially reconnect, are no longer held up by tasks on other
devices that may take a long time, especially login with INQUIRY and sd
or sr driver probe.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
We do not need slab allocations for ORB pointer write transactions
anymore in order to satisfy streaming DMA mapping constraints, thanks to
commit da28947e7e "firewire: ohci: avoid separate DMA mapping for
small AT payloads".
(Besides, the slab-allocated buffers that firewire-sbp2 used to provide
for 8-byte write requests were still not fully portable since they
shared a cacheline with unrelated CPU-accessed data.)
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
firewire-sbp2 already takes care for internal serialization where
required (ORB list accesses), and it does not use cmd->serial_number
internally. Hence it is safe to not grab the shost lock around
queuecommand.
While we are at housekeeping, drop a redundant struct member:
sbp2_command_orb.done is set once in a hot path and dereferenced once in
a hot path. We can as well dereference sbp2_command_orb.cmd->scsi_done
instead.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
firewire-core manages the following types of work items:
fw_card.br_work:
- resets the bus on a card and possibly sends a PHY packet before that
- does not sleep for long or not at all
- is scheduled via fw_schedule_bus_reset() by
- firewire-ohci's pci_probe method
- firewire-ohci's set_config_rom method, called by kernelspace
protocol drivers and userspace drivers which add/remove
Configuration ROM descriptors
- userspace drivers which use the bus reset ioctl
- itself if the last reset happened less than 2 seconds ago
fw_card.bm_work:
- performs bus management duties
- usually does not (but may in corner cases) sleep for long
- is scheduled via fw_schedule_bm_work() by
- firewire-ohci's self-ID-complete IRQ handler tasklet
- firewire-core's fw_device.work instances whenever the root node
device was (successfully or unsuccessfully) discovered,
refreshed, or rediscovered
- itself in case of resource allocation failures or in order to
obey the 125ms bus manager arbitration interval
fw_device.work:
- performs node probe, update, shutdown, revival, removal; including
kernel driver probe, update, shutdown and bus reset notification to
userspace drivers
- usually sleeps moderately long, in corner cases very long
- is scheduled by
- firewire-ohci's self-ID-complete IRQ handler tasklet via the
core's fw_node_event
- firewire-ohci's pci_remove method via core's fw_destroy_nodes/
fw_node_event
- itself during retries, e.g. while a node is powering up
iso_resource.work:
- accesses registers at the Isochronous Resource Manager node
- usually does not (but may in corner cases) sleep for long
- is scheduled via schedule_iso_resource() by
- the owning userspace driver at addition and removal of the
resource
- firewire-core's fw_device.work instances after bus reset
- itself in case of resource allocation if necessary to obey the
1000ms reallocation period after bus reset
fw_card.br_work instances should not, and instances of the others must
not, be executed in parallel by multiple CPUs -- but were not protected
against that. Hence allocate a non-reentrant workqueue for them.
fw_device.work may be used in the memory reclaim path in case of SBP-2
device updates. Hence we need a workqueue with rescuer and cannot use
system_nrt_wq.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Reviewed-by: Tejun Heo <tj@kernel.org>
When queueing iso packets, the run time is dominated by the two
MMIO accesses that set the DMA context's wake bit. Because most
drivers submit packets in batches, we can save much time by
removing all but the last wakeup.
The internal kernel API is changed to require a call to
fw_iso_context_queue_flush() after a batch of queued packets.
The user space API does not change, so one call to
FW_CDEV_IOC_QUEUE_ISO must specify multiple packets to take
advantage of this optimization.
In my measurements, this patch reduces the time needed to queue
fifty skip packets from userspace to one sixth on a 2.5 GHz CPU,
or to one third at 800 MHz.
Signed-off-by: Clemens Ladisch <clemens@ladisch.de>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
We do not need slab allocations anymore in order to satisfy
streaming DMA mapping constraints, thanks to commit da28947e7e
"firewire: ohci: avoid separate DMA mapping for small AT payloads".
(Besides, the slab-allocated buffers that firewire-core, firewire-sbp2,
and firedtv used to provide for 8-byte write and lock requests were
still not fully portable since they crossed cacheline boundaries or
shared a cacheline with unrelated CPU-accessed data. snd-firewire-lib
got this aspect right by using an extra kmalloc/ kfree just for the
8-byte transaction buffer.)
This change replaces kmalloc'ed lock transaction scratch buffers in
firewire-core, firedtv, and snd-firewire-lib by local stack allocations.
Perhaps the most notable result of the change is simpler locking because
there is no need to serialize usages of preallocated per-device buffers
anymore. Also, allocations and deallocations are simpler.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Acked-by: Clemens Ladisch <clemens@ladisch.de>
Current implementation of ohci_set_config_rom() uses a deferred
bus reset via fw_schedule_bus_reset(). If clients add multiple
unit descriptors to the config_rom in quick succession, the
deferred bus reset may not have fired before succeeding update
requests have come in. This can lead to an incorrect partial
update of the config_rom for both addition and removal of
config_rom descriptors, as the ohci_set_config_rom() routine
will return -EBUSY if a previous pending update has not been
completed yet; the requested update just gets dropped on the floor.
This patch recognizes that the "in-flight" update can be modified
until it has been processed by the bus-reset, and the locking
in the bus_reset_tasklet ensures that the update is done atomically
with respect to modifications made by ohci_set_config_rom(). The
-EBUSY error case is simply removed.
[Stefan R: The bug always existed at least theoretically. But it
became easy to trigger since 2.6.36 commit 02d37bed18 "firewire: core:
integrate software-forced bus resets with bus management" which
introduced long mandatory delays between janitorial bus resets.]
Signed-off-by: Benjamin Buchalter <bj@mhlabs.com>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de> (trivial style changes)
Cc: <stable@kernel.org> # 2.6.36.y and newer
