Instead of allocating memory for address range mappings (ARM) and handling all
reads and writes to said memory, the new ARM manager calls a function for each
received request with the data, transaction code, and allocation information.
The ARM tag manager now must handle validation of memory accesses and
retreive/write the data. This allows implementations to use network resources
or generated data as memory.
The ARM manager no longer automatically sends response packets. It is the
responsibility of the user to send response packets using
raw1394_send_rw_response().
The interface is not implemented for raw1394 and will probably never be
implemented for raw1394. It is for firewire-cdev (modern Linux) only.
The __uint32_t type is not a standard type, and not necessarily
provided by all C libraries. However, <stdint.h> provides a set of
standard types, which are guaranteed to be available with all C
libraries.
This patch therefore changes __uint32_t to uint32_t, which allows to
fix the compilation with the Musl C library.
This patch is already used by the Alpine Linux distribution and the
Buildroot build system.
Original patch from
http://git.alpinelinux.org/cgit/aports/plain/main/libraw1394/fix-types.patch.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Align fw_handle buffer for 64-bit access. This fixes SIGBUS on SPARC
when capturing DV stream with "dvgrab".
[Stefan R: If libraw1394 is compiled for 32 bit userland, struct
fw_handle.buffer was only 32 bit aligned. Various *__u64 accesses
happen to the buffer, and those accesses require 64 bit alignment on
some CPU architectures. The bug certainly affected all libraw1394
client applications on such architectures with 32 bit userland.]
Signed-off-by: Aaro Koskinen <aaro.koskinen@iki.fi>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
This bugfix grew out of an extended investigation into a problem encountered
by a small number of people running FFADO. FFADO would report that the tx
iso cycle number supplied to the iso tx callback seemingly went backwards -
something which should not ordinarily occur. The bug seemed to be sensitive
to timing and in some cases would disappear when debug traces were inserted
into either FFADO or libraw1394. In essence, libraw1394 was requesting tx
data for cycles which had already been requested.
Initial discussions can be found in the thread "Problem with RME FF800. Can
not start jackd" on the ffado-user mailing list. A followup investigation
is tracked in FFADO ticket number 379
(http://subversion.ffado.org/ticket/379) and referenced in the thread
"Revisiting backward cycle number sequence (ticket 379)" on ffado-devel.
The latter mailing list thread includes a lengthy explanation of what I
think is happening.
To summarise, the root of the problem seems to be that on certain machines
under certain conditions, something causes the kernel to post an iso tx
event at a time when fewer than irq_interval packets have been transmitted.
Unfortunately it has not been possible to determine the underlying cause of
this. Whatever the cause, tests carried out with the reporter of ticket 379
have shown that it is occurring. As a result, the adjustment to
libraw1394's packet_count must be done with reference to the number of
packets reported as transmitted by the kernel instead of simply assuming
that irq_interval packets have been sent.
A patch implementing this fix is at the end of this post. This fixes the
problem when the newer ABI is in use, which provides tx packet timestamps
(and thus an indication of the number of packets actually transmitted) to
userspace. It does not address the problem when the older ABI is used, but
given the nature of the problem I don't think it's possible to fix it
without access to the timestamps (or at least without some way to determine
the number of packets really transmitted).
Testing by "juanramon" (see ticket 379) has demonstrated that it fixes the
"backward cycle number" problem on his machine.
Thanks to Andreas Hehn and "juanramon" for their invaluable help in tracking
this down.
Signed-off-by: Jonathan Woithe <jwoithe@just42.net>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
A temporarily allocated buffer which is used to pass data from libraw1394's
event loop to the Address Range Mapping callback was never freed.
This was pointed out by the following valgrind trace:
3067120 (3066560 direct, 560 indirect) bytes in 10952 blocks are definitely lost in loss record 36 of 36
at 0x4029F6F : malloc ()
by 0x405B1B5 : ??? (in usr/lib/libraw1394.so.11.0.1)
by 0x405B492 : ??? (in usr/lib/libraw1394.so.11.0.1)
by 0x405BF24 : fw_loop_iterate (in usr/lib/libraw1394.so.11.0.1)
by 0x405C197 : ??? (in usr/lib/libraw1394.so.11.0.1)
by 0x405D6F8 : fw_write (in usr/lib/libraw1394.so.11.0.1)
by 0x405A292 : raw1394_write (in usr/lib/libraw1394.so.11.0.1)
by 0x805A0F2 : main (main.cpp:121)
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Since dual-stack capability was added to libraw1394, raw1394_new_handle()
and raw1394_new_handle_on_port() began to alter errno even when succeeding.
This breaks old application code which contains the bug of checking for
failure in errno rather than in the return code of said functions, or
similar bugs with wrong assumptions about errno.
While those applications should be fixed, it may not always be possible or
feasible to do so. Hence add a workaround to libraw1394 which saves and
restores errno in said two functions.
From a superficial review of dispatch.c, it seems that these two functions
are the only ones where such a workaround may be needed. However, this may
not be true if any fw_XYZ() function implementation differs from the
counterpart ieee1394_XYZ() function in the way that the former alters errno
during successful execution while the latter does not. To be clear,
altering errno in absence of failure is absolutely allowed in library code
(except for signal handlers), yet it may be unexpected and be perceived as
a library or kernel regression if the application client code is buggy in
this regard.
Reported-by: Vladimir Romanov <blueboar2@gmail.com>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
"the error code of the ioctl" is always -1. And ioctl() sets errno for us.
So let's say this in a way which is less likely to be misunderstood.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
This lets initiators of isochronous streams or asynchronous streams from
or to the local node figure out what maximum speed can be configured.
Furthermore it can be used to display connection speeds for informative
purposes without having to perform topology analysis (in case of 1394a
buses) or extensive phy port status queries (in case of 1394b buses).
To be in line with other existing libraw1394 APIs which use nodeid_t
variables, this API identifies a node only via a card:nodeID tuple which
is unsafe against generation changes. A node can only be properly
identified by card:generation:nodeID tuples. However, since this new
API extension and libraw1394 as a whole is mainly aimed towards existing
libraw1394 client code bases rather than new developments, I decided
against making this call race free but somewhat more difficult to use in
typical existing client code.
A unit test for the new call is added to testlibraw as well.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
This catapults the libraw1394 API into the year 2002.
Actually, passing speed codes of 3...5 into the relevant libraw1394
functions should be working already since the kernel gained 1394b
support a long time ago and libraw1394 does not check values.
The added definitions are only for clarity and to fully match the
argument type in the function declarations.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
This can save some memory. (The library user can always round max
packet size himself if it's needed for some reason.)
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
This is an extension relative to raw1394_read_cycle_timer().
It lets the client choose a system clock other than CLOCK_REALTIME.
Use case: http://subversion.ffado.org/ticket/242
The underlying ioctl supports reading the system clock with nanoseconds
resolution. The new libraw1394 call sticks with microseconds resolution
though. This makes transition from (or parallel use with)
raw1394_read_cycle_timer() easier. Besides, the call itself takes longer
than a microsecond, primarily due to a costly MMIO read (on many
controllers even three or more MMIO reads).
Unit tests with CLOCK_MONOTONIC and CLOCK_MONOTONIC_RAW are added to
testlibraw as well.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
To conform with 'size' arguments of other libraw1394 calls, change the one in
raw1394_add_config_rom_descriptor() from quadlets to bytes. This breaks runtime
compatibility with potential clients that were written against B.J.'s original
patch, hence reorder arguments just to break compatibility also at compile time.
Change errno to ENOSYS (function not implemented) when called while running on
top of raw1394.
Allow &token to be NULL for convenience of clients which don't require
raw1394_remove_config_rom_descriptor().
Add exhaustive documentation. Much of it is copied from the documentation of
the underlying ioctl.
Add example code which doubles as unit test in testlibraw.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
This adds support of the firewire-core (juju) ABI to add and remove config ROM directories
or descriptors. The raw1394 ABI supports similar requests, but for now we leave the two
functions unimplemented when running on top of raw1394.
Signed-off-by: Benjamin Buchalter <bj@mhlabs.com>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de> (whitespace changes)
This guarantees that all features of libraw1394 are actually built in.
Before, some features and fixes would be silently dropped if too old
system headers (typically provided by a package called linux-headers)
were present.
An alternative would be to keep using system headers but add warnings
during the ./configure stage if old headers were encountered. But this
helps only the person who builds libraw1394 (if there is a person
involved at all), not the users who have no reliable way to determine
how the library binary was built.
Another alternative would be to change the former soft dependency on
certain linux-headers versions into a hard dependency, i.e. fail the
build in absence of too old headers. This would add an inconvenience
in setting up the build environment though: The system headers would
have to be updated or a private copy of linux/firewire-*.h be specified
by way of the --with-fw-dir configure switch.
Anyhow. The libraw1394 sources now already bring a suitable copy of
the two header files. The --with-fw-dir configure switch is no longer
useful and is removed.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Add copies of the Linux kernel header files but don't #include them just yet.
These are exact copies of the files from linux-3.4. In case that we eventually
want updates from later kernels, a diff of the kernel's firewire-c*.h between
v3.4 and then should do the trick.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
If read() on the inotify handle gave us several events at once, and handling
one of them resulted in whatever error, there is little reason not to try
handling the rest of the events.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
If multiple inotify events are presented, process *all* of them.
This can happen when several device adds are pushed simultaneously.
handle_inotify() was refactored.
Signed-off-by: Peter Hurley <phurley@charter.net>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
If an error is encountered while adding a new device in inotify
handling, make sure the fd is marked invalid (-1).
Signed-off-by: Peter Hurley <phurley@charter.net>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Applications or higher-level libraries have retry strategies of their own
in place, but they don't work too well sometimes. For example, old
Panasonic camcorders require pauses in the order of several milliseconds
between response of a former transaction and request of the next one,
but libavc1394 and libiec61883 retry already after 20 microseconds.
This change cures all FCP transaction failures ("send oops") in kino and
dvgrab that I was getting with Panasonic NV-DX110. According to reports,
Panasonic AG-EZ30 and Grundig Scenos DLC 2000 were affected too.
The additional latency in raw1394_read/write/lock/lock64 appears to be
the better alternative compared to terminal I/O failures. Besides, a
caller of this blocking request API should at least in theory be
prepared to cope with transaction durations in the order of a few seconds.
IEEE 1394 specifies split transaction timeouts of up to 8 seconds. An
application which needs more control should use the non-blocking request
API, i.e. raw1394_start_read/write/lock/lock64.
We specifically only retry after ack-busy, not after any of the other
types of transaction failures that may or may not succeed if retried.
This change is only done in the firewire-core backend (a.k.a. juju).
The same could be added to the raw1394 backend (a.k.a. linux1394) but is
not as important there, perhaps because transaction completion latency
in the ieee1394 core very much increases the success rate of existing
retry code in libavc1394 and friends.
Note, this does not fix every and all FCP transaction problems. There
are e.g. certain JVC camcorders which do not properly complete FCP
transactions if an application frequently polls for status or requests
status right before a control request, even with an order of magnitude
greater delays than used in this patch.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Reviewed-by: Dan Dennedy <dan@dennedy.org>
The kernel already read each node's Configuration ROM and cached it.
So let all libraw1394 clients read from that cache instead of having
to perform all those transactions all over again.
This reduces bus traffic at application start-up and at each bus reset.
It also makes all Configuration ROM accesses fool-proof and robust.
This together with the kernel patch "firewire: core: handle ack_busy
when fetching the Config ROM" lets me use an old Panasonic camcorder
which requires us to keep pauses between response and request ---
longer than librom1394's retry pause --- with dvgrab (though still
with frequent failures of write requests to FCP_COMMAND, i.e. with lots
of "send oops" noise in the console and occasionally having to repeat
key-presses in interactive mode).
For simplicity of implementation, only the blocking raw1394_read() is
modified, not the nonblocking raw1394_start_read().
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Signed-off-by: Dan Dennedy <dan@dennedy.org>
Libraw1394 uses raw1394's convention of cycle numbers always being less
than one second, i.e., in the range 0..7999.
Firewire-core uses raw cycle numbers as used by the hardware, i.e., with
several additional bits for the seconds. This was correctly handled
when presenting timestamps returned by the kernel to the application,
but the application's start_on_cycle value was passed directly to the
kernel.
To fix this, do the same calculations that ohci1394 did internally,
i.e., interpret the start_on_cycle value as relative to the current
seconds value of the cycle timer.
Signed-off-by: Clemens Ladisch <clemens@ladisch.de>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Libraw1394's attempt to optimize away the packet queueing ioctl syscall
overhead was a little bit too successful: when used with FFADO's
streaming system, it ended up delaying packets after their intended
transmission time.
For now, to ensure correctness, don't try to optimize anything.
This makes FFADO playback on Juju with my DICE work.
Signed-off-by: Clemens Ladisch <clemens@ladisch.de>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
When faced with bogus config ROM read responses from an audio device
that did not support block requests as advertized, libffado's csr1212
code was able to recover when running on top of raw1394 but corrupted
its config ROM cache when running on top of firewire-core.
http://subversion.ffado.org/ticket/299
While the actual cause was a combination of firmware bug of the device
and flaw in csr1212.c of libffado, the much less graceful behavior when
running on firewire-core was obviously due to libraw1394's
firewire-core backend. Hence,
- do not write into the client's buffer if rcode != RCODE_COMPLETE,
- do not copy more data than the actual response contained.
The latter safeguard is not overly effective though. The libraw1394 API
has no means to inform a client about the error case that a responder
node sent less bytes than were requested. (The case that the responder
sent more bytes than requested is covered up by the kernel already.)
Should we synthesize an I/O failure? Does not sound ideal either.
However, such a size mismatch should never happen; the important part of
this change is the RCODE_COMPLETE check.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
If multiple cards are installed, firewire-core will emit requests from
nodes on any of the cards to clients. This is not expected by
libraw1394 clients since a raw1394handle_t is bound to a single card
alias port.
On kernel 2.6.36 and newer we can filter out requests from other cards.
Note that we still need to call the response ioctl in order to release
kernel resources associated with an inbound transaction.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
The firewire-core (juju) backend of libraw1394 installs address range
mappings on the default ioctl fd, i.e. a file that represents a random
device on the chosen port. It receives incoming requests from any
sender node via this address range mapping. Due to a kernel ABI
limitation, the sender node ID is not known though. So far libraw1394
simply assumed the node ID of the device that provided the default
ioctl fd. This only works if there is only one accessible fd on the
entire bus.
This limitation caused for example libffado to fail to work with
another AV/C or IIDC device attached to the bus, because node IDs of
FCP requests and FCP responses did not match since the latter were
wrong. FCP clients which did not check sender node IDs were seemingly
not affected by this bug. The bug is fixed by a kernel ABI extension
in Linux 2.6.36. This libraw1394 change implements libraw1394's
counterpart to this ABI extension.
Hence this libraw1394 fix requires
- kernel-headers 2.6.36 or later at build time of libraw1394
- kernel 2.6.36 or later at runtime.
Otherwise, libraw1394 simply degrades to the faulty previous behaviour.
Side note: The change of IMPLEMENTED_CDEV_ABI_VERSION to 4 requires
that we fill in struct fw_cdev_allocate.region_end which was added in
the ABI v4.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Since linux/firewire-cdev.h header file and libraw1394 sources are
distributed separately, it is wrong to fill in a constant from that
header into the FW_CDEV_IOC_GET_INFO ioctl as the ABI version which
libraw1394 supports. This may not be forward compatible if an old
libraw1394 is compiled with a new kernel header and ran on top of a
kernel that implements new features that require a compatible
userland.
OK, the damage is already done in released versions of libraw1394.
Hence the FW_CDEV_VERSION of the kernel header file is not going to
be updated anymore in future kernel versions. (Only the version
internally to firewire-core will be incremented further.)
But let's remove the buggy usage of FW_CDEV_VERSION nevertheless.
Developers of other firewire-cdev client programs might look at
libraw1394 sources. The libraw1394 sources should not teach them
how to do it wrong.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
because ieee1394_destroy_handle does it too. Otherwise, clients which
rely on the ieee1394 backend behaviour leak memory when running on the
juju backend.
Also add the ability to call fw_iso_shutdown multiple times or before
a successful context initialization. Faulty clients might rely on it
based on ieee1394 backend behaviour.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Libraw1394 must not rely on the kernel always handing out the value 0
as handle of the (first) allocated isochronous I/O context. For now
this assumption is true but it may not stay that way forever.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Also add errno = ENOMEM because it is said that that some malloc
implementations might miss to do so.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
When EPOLLHUP event is received in fw_loop_iterate(), it is or'd
with EPOLLERR. The EPOLLHUP event was then overlooked in
handle_device_event() with unpredictable-but-generally bad results.
This problem has been rediscovered several times.
http://thread.gmane.org/gmane.linux.kernel.firewire.devel/13330http://thread.gmane.org/gmane.linux.kernel.firewire.devel/13779
Reported-by: B.J. Buchalter <bj@mhlabs.com>
Reported-by: Michael Thireos <mthireos@vanteon.com>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
If running on top of the raw1394 kernel driver (any kernel version) or
on top of firewire-core (kernel version <= 2.6.29), raw1394_start_*
functions will return a value > 0 on success, not == 0. Only with
firewire-core of kernel 2.6.30 or later, == 0 is returned on success.
The exact value depends on which driver is used, on CPU architecture,
and on request payload size in case of some types of requests. In any
case, only that the value is > or == 0 on success (but == -1 on
failure) is significant to libraw1394 client applications.
This mismatch between documentation and implementation was already
present in older libraw1394 versions, including v1.x. For the time
being, do not change the implementation, only adjust the documentation.
Reported-by: Dr. David Alan Gilbert
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Reported by Guus Sliepen: "make doc" failed due to missing doctype,
unknown elements, and duplicate element IDs in libraw1394.sgml.
The fix is to declare a recent DTD (matching the one which is used
in current Linux kernel documentation docbooks) and to make the
conflicting element IDs unique.
The latter part of the fix is just temporary. In order to avoid the
conflict when the documentation is updated the next time, also fix the
kerneldoc comments of the respective API elements: These are typedefs,
hence kernel-doc needs their comments prepended by "typedef ".
Tested with Gentoo's docbook-xml-dtd 4.5, docbook-xsl-stylesheets
1.75.2, docbook-sgml-utils 0.6.14, and openjade 1.3.2-r1.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Signed-off-by: Dan Dennedy <dan@dennedy.org>
The number of packets is a 4th of the number of header bytes (in case of
ABI version 1). Also, wrap after an increment over 8000.
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
This change is essentially cosmetic: Set fields of structs passed to
the kernel via ioctl so that valgrind will not complain about them.
Signed-off-by: Jay Fenlason <fenlason@redhat.com>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de> (changelog, comments)
More accurately report the cycle on which isochronous packets were
received. Only affects libraw1394 when used with kernel 2.6.29 or
older.
Signed-off-by: Jay Fenlason <fenlason@redhat.com>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de> (changelog, whitespace)
Peter McGoron
Thomas Petazzoni
Aaro Koskinen
Jonathan Woithe
Lee Cewd
Stefan Richter
Igor Kuzmin
B.J. Buchalter
Clemens Ladisch
Guus Sliepen
Peter Hurley
Jay Fenlason