Add an extra parameter to the ring dequeue burst/bulk functions so that
those functions can optionally return the amount of remaining objs in the
ring. This information can be used by applications in a number of ways,
for instance, with single-consumer queues, it provides a max
dequeue size which is guaranteed to work.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Reviewed-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
Add an extra parameter to the ring enqueue burst/bulk functions so that
those functions can optionally return the amount of free space in the
ring. This information can be used by applications in a number of ways,
for instance, with single-producer queues, it provides a max
enqueue size which is guaranteed to work. It can also be used to
implement watermark functionality in apps, replacing the older
functionality with a more flexible version, which enables apps to
implement multiple watermark thresholds, rather than just one.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Reviewed-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
The bulk fns for rings returns 0 for all elements enqueued and negative
for no space. Change that to make them consistent with the burst functions
in returning the number of elements enqueued/dequeued, i.e. 0 or N.
This change also allows the return value from enq/deq to be used directly
without a branch for error checking.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Reviewed-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
Remove the watermark support. A future commit will add support for having
enqueue functions return the amount of free space in the ring, which will
allow applications to implement their own watermark checks, while also
being more useful to the app.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Reviewed-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
There was a compile time setting to enable a ring to yield when
it entered a loop in mp or mc rings waiting for the tail pointer update.
Build time settings are not recommended for enabling/disabling features,
and since this was off by default, remove it completely. If needed, a
runtime enabled equivalent can be used.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Reviewed-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
The debug option only provided statistics to the user, most of
which could be tracked by the application itself. Remove this as a
compile time option, and feature, simplifying the code.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Reviewed-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
The size and mask fields are duplicated in both the producer and
consumer data structures. Move them out of that into the top level
structure so they are not duplicated.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Reviewed-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
create a common structure to hold the metadata for the producer and
the consumer, since both need essentially the same information - the
head and tail values, the ring size and mask.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Reviewed-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
Users compiling DPDK should not need to know or care about the arrangement
of cachelines in the rte_ring structure. Therefore just remove the build
option and set the structures to be always split. On platforms with 64B
cachelines, for improved performance use 128B rather than 64B alignment
since it stops the producer and consumer data being on adjacent cachelines.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Reviewed-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
Changes in the thread layout described, with an updated diagram.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: John McNamara <john.mcnamara@intel.com>
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: John McNamara <john.mcnamara@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Now that we're printing out a page of stats every second to the console,
we should give the stats it's own core so that we don't interfere with
the performance of the Rx core.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
This patch tunes Rx, Tx, and rte_distributor_process() burst sizes to
maximize performance.
It also addresses some checkpatch issues.
The result is approximately 10% performance increase.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Give the distribution functionality it's own core for performance,
otherwise it's limited by the Rx core.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
On some machines, ports take several seconds to come up. This
patch causes the app to wait.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
This will allow us to see what's going on at various stages
throughout the sample app, with per-second visibility
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
This is the main switch over between the legacy API and the new
burst API. We rename all the functions in rte_distributor.c to remove
the _v1705, and we add in _v20 in the rte_distributor_v20.c
We also rename the rte_distributor_next.h as rte_distributor.h, as
this is now the public header.
At the same time, we need the autotests and sample app to compile
properly, hence those changes are in this patch also.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
In the next few patches, we'll want to test old and new API,
so here we're allowing different parameters to be passed to
the tests, instead of just a distributor struct.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Add an optimised version of the in-flight flow matching algorithm
using SIMD instructions. This should give up to 1.5x over the scalar
versions performance.
Falls back to scalar version if SSE4.2 not available
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
This patch includes the code for new burst-capable distributor library.
It also includes the rte_distributor_next.h file which will
be used as the public header once we add in the symbol versioning
for v20 and v1705 APIs, at which stage we will rename it to
rte_distributor.h.
The new distributor code contains a very similar API to the legacy code,
but now sends bursts of up to 8 mbufs to each worker. Flow ID's are
reduced to 15 bits for an optimal flow matching algorithm.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
We'll be adding internal implementation definitions in here
that are common to both burst and legacy APIs.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Move files out of the way so that we can replace with new
versions of the distributor library. Files are named in
such a way as to match the symbol versioning that we will
apply for backward ABI compatibility.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
When a C file for a library/driver is including the public header files for
that library, those need to be included as local includes using quotes
rather than angle-brackets. Without doing so, parallel builds can fail, as
the compiler will only look for those headers in the global include folder
rather than locally, and the build system does not enforce that the headers
for a lib are installed before the rest of the lib is compiled.
Fixes: 097ab0bac0 ("crypto/scheduler: add API")
Fixes: 503e9c5afb ("crypto/scheduler: register as vdev driver")
Fixes: 31439ee72b ("crypto/scheduler: add API implementations")
Cc: stable@dpdk.org
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Xenvirt driver uses the cmdline lib for parsing ether addresses so add it
as a dependency to fix builds.
Fixes: feb9f680cd ("mk: optimize directory dependencies")
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Rather than reading the size directly from the ring structure, use the
dedicated ring function for that purpose.
Previous commits to do this only did so for the null crypto driver which
was the only one compiled in by default, but all other drivers need to be
similarly updated.
Fixes: b11c78a2e0 ("crypto/null: use ring size function")
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Rather than querying the number of CPUs on the system multiple times, and
printing out the number each time, just query the value from sysctl once
and store it for future reuse.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
The -e flag to readlink doesn't exist on FreeBSD so change it to -f instead
which is present on both BSD and Linux. Error reported is:
readlink: illegal option -- e
usage: readlink [-fn] [file ...]
usage: dirname string [...]
./devtools/check-git-log.sh: /git-log-fixes.sh: not found
Fixes: 814c8822ef ("scripts: check cc stable mailing list in commit")
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Before this patch, the management of dependencies between directories
had several issues:
- the generation of .depdirs, done at configuration is slow: it can take
more than one minute on some slow targets (usually ~10s on a standard
PC without -j).
- for instance, it is possible to express a dependency like:
- app/foo depends on lib/librte_foo
- and lib/librte_foo depends on app/bar
But this won't work because the directories are traversed with a
depth-first algorithm, so we have to choose between doing 'app' before
or after 'lib'.
- the script depdirs-rule.sh is too complex.
- we cannot use "make -d" for debug, because the output of make is used for
the generation of .depdirs.
This patch moves the DEPDIRS-* variables in the upper Makefile, making
the dependencies much easier to calculate. A DEPDIRS variable is still
used to process library dependencies in LDLIBS.
After this commit, "make config" is almost immediate.
Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
Tested-by: Robin Jarry <robin.jarry@6wind.com>
Tested-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Add support to the vHostdriver for the new API to force free consumed
buffers on Tx ring. vHost does not cache the mbufs so there is no work
to do.
Signed-off-by: Billy McFall <bmcfall@redhat.com>
Acked-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Keith Wiles <keith.wiles@intel.com>
Add support to the e1000 igb driver for the new API to force free
consumed buffers on Tx ring. This API is independent of the tx_rs_thresh
setting. With this API, buffers should be free even if tx_rs_thresh is
not met.
e1000 igb driver does not implement a tx_rs_thresh to free mbufs, it
frees a slot in the ring as needed. However, it could be implemented at
some future date.
Signed-off-by: Billy McFall <bmcfall@redhat.com>
Acked-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Acked-by: Keith Wiles <keith.wiles@intel.com>
Add a new API to force free consumed buffers on Tx ring. API will return
the number of packets freed (0-n) or error code if feature not supported
(-ENOTSUP) or input invalid (-ENODEV).
Signed-off-by: Billy McFall <bmcfall@redhat.com>
Acked-by: Keith Wiles <keith.wiles@intel.com>
The rte_eal_init function will now pass failure reason hints to the
application. To help app developers decipher this, add some brief
information about what the codes are indicating.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
For now, exit the init. It's likely that even aborting the initialization
is premature in this case, as it may be possible to proceed even if one
bus or another is not available.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Even if one vdev should fail, there's no need to prevent further
processing. Log the error, and reflect it to the higher levels to
decide.
Seems like it's possible to continue. At least, the error is reflected
properly in the logs. A user could then go and correct or investigate
the situation.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Some devices may be inaccessible for a variety of reasons, or the
PCI-bus may be unavailable causing the whole thing to fail. Still,
better to continue attempts at probes.
Since PCI isn't neccessarily required, it may be possible to simply log
the error and continue on letting the user check the logs and restart
the application when things have failed.
This will usually be an issue because of permissions. However, it could
also be caused by OOM. In either case, errno will contain the
underlying cause.
For linux, it is safe to re-init the system here, so allow the
application to take corrective action and reinit.
For BSD, this is not the case, for other reasons, including hugepage
allocation has already happened, and needs to be properly uninitialized.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Plugins are useful and important. However, it seems crazy to abort
everything just because they don't initialize properly.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
There could be some confusion as to why the call failed - this change
will always reflect the value of the error in rte_error.
When initializing the interrupt thread, there are a number of possible
reasons for failure - some of which are correctable by the application.
Do not panic() needlessly, and give the application a change to reflect
this information to the user.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
After code inspection, there is no way for eal_timer_init() to fail. It
simply returns 0 in all cases. As such, this test could either go-away
or stay here as 'future-proofing'.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
When log initialization fails, it's generally because the fopencookie
failed. While this is rare in practice, it could happen, and it is
likely because of memory pressure. So, flag the error, and allow the
user to retry.
Memory init can only fail when access to hugepages (either as primary or
secondary process) fails (and that is usually permissions). Since the
manner of failure is not reversible, we cannot allow retry.
There are some theoretical racy conditions in the system that _could_
cause early tailq init to fail; however, no need to panic the
application. While it can't continue using DPDK, it could make better
alerts to the user.
rte_eal_alarm_init() call uses the linux timerfd framework to create a
poll()-able timer using standard posix file operations. This could fail
for a few reasons given in the man-pages, but many could be
corrected by the user application. No need to panic.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
When memzone initialization fails, report the error to the calling
application rather than panic(). Without a good way of detaching /
releasing hugepages, at this point the application will have to restart.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
It's possible that the application could take a corrective action here,
and either prompt the user for different arguments, or at least perform
a better logging. Exiting this early prevents any useful information
gathering from the application layer.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
When attempting to scan hugepages, signal to the eal that an error has
occurred, rather than performing a panic.
If we fail to acquire hugepage information, simply signal an error to
the application. This clears the run_once counter, allowing the user or
application to take a corrective action and retry.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
This adds a new API to check for the eal cpu versions.
It's now possible to gracefully exit the application, or for
applications which support non-dpdk datapaths working in concert with
DPDK datapaths, there no longer is the possibility of exiting for
unsupported CPUs.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
There may be no way to gracefully recover, but the application
should be notified that a failure happened, rather than completely
aborting. This allows the user to proceed with a "slow-path" type
solution.
After this change, the EAL CPU NUMA node resolution step can no longer
emit an rte_panic. This aligns with the code in rte_eal_init, which
expects failures to return an error code.
Signed-off-by: Aaron Conole <aconole@redhat.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
The FreeBSD implementation wasn't registering new devices
with the device framework on start up. However, common
code attempts to unregister them on shutdown which causes
a SEGFAULT. This fix makes the FreeBSD code do the same
thing as the Linux code for registration.
Fixes: 13a1317d3b ("pci: create device list and fallback on its members")
Cc: stable@dpdk.org
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Acked-by: Shreyansh Jain <shreyansh.jain@nxp.com>
