Tao Zhu [Thu, 16 Jan 2020 16:38:36 +0000 (16:38 +0000)]
net/ice: fix flow FDIR/switch memory leak
1. Fix ice FDIR and hash flow memory leak.
2. Fix the ice definition of LIST_FOR_EACH_ENTRY_SAFE not
save tmp which cause list deletion incompletely.
Fixes: 5f0978e96220 ("net/ice/base: add OS specific implementation") Fixes: f5cafa961fae ("net/ice: add flow director create and destroy") Fixes: 5ad3db8d4bdd ("net/ice: enable advanced RSS") Cc: stable@dpdk.org Signed-off-by: Tao Zhu <taox.zhu@intel.com> Reviewed-by: Simei Su <simei.su@intel.com> Reviewed-by: Yahui Cao <yahui.cao@intel.com> Acked-by: Xiaolong Ye <xiaolong.ye@intel.com>
Xiaolong Ye [Wed, 22 Jan 2020 03:58:10 +0000 (11:58 +0800)]
net/i40e/base: skip further adminq init for VF
Since VF has no need of firmware, we can skip further adminq init which
involves firmware operation, this patch fixes the testpmd segfault issue
when starting with i40e VF.
Fixes: d5e1a149362e ("net/i40e/base: check MAC type") Signed-off-by: Xiaolong Ye <xiaolong.ye@intel.com> Acked-by: Beilei Xing <beilei.xing@intel.com>
Somnath Kotur [Thu, 16 Jan 2020 13:04:55 +0000 (18:34 +0530)]
net/bnxt: fix max rings calculation
Max Tx rings count could be lesser than max Rx rings in some
cases, so take this into account as well.
Account for stat contexts available(one for each ring) along with
no: of completion rings(one for each ring) to cap the max no: of
Tx /Rx rings that can be possibly created.
Kalesh AP [Thu, 16 Jan 2020 13:04:52 +0000 (18:34 +0530)]
net/bnxt: fix a memory leak in port stop
The memory for mark table is allocated during port start.
But the allocated memory is freed only during port close
or driver unload which in turn causes a memory leakage
on each port start/stop.
Fixed it by moving the memory free to port stop.
Fixes: 94eb699bc82e ("net/bnxt: support flow mark action") Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Reviewed-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Kalesh AP [Thu, 16 Jan 2020 13:04:51 +0000 (18:34 +0530)]
net/bnxt: handle HW filter setting when port is stopped
Driver destroy the VNIC when the port is brought down.
Port HW filter setting such as promiscuos, allmulti and
VLAN filtering will be applied when port is started.
Fixed to return success silently for these callbacks
when port is stopped. Also fixed to clear "bp->dev_stopped"
before invoking bnxt_vlan_offload_set_op() in bnxt_dev_start_op().
Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Reviewed-by: Venkat Duvvuru <venkatkumar.duvvuru@broadcom.com>
Kalesh AP [Thu, 16 Jan 2020 13:04:50 +0000 (18:34 +0530)]
net/bnxt: fix VLAN strip
HWRM_VNIC_CFG command to configure VNIC dynamically with
traffic running is not working. Driver has to free and
recreate the VNIC and then reconfigure the VNIC filters.
Fixes: 7fe5668d2ea3 ("net/bnxt: support VLAN filter and strip") Cc: stable@dpdk.org Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Reviewed-by: Ajit Khaparde <ajit.khaparde@broadcom.com> Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com>
Kalesh AP [Thu, 16 Jan 2020 13:04:49 +0000 (18:34 +0530)]
net/bnxt: fix enable/disable VLAN filtering
There is no condition check for the user requested operation
for VLAN filtering. As a result, VLAN filtering is getting disabled
when the user enables/disables VLAN stripping on same port.
The function bnxt_hwrm_clear_l2_filter() didn't actually free
L2 filter in HW if the reference count of filter is zero.
Fixed it by incrementing the reference count of filter in
bnxt_alloc_filter() routine.
Because of the recent changes in bnxt_hwrm_clear_l2_filter(),
change was needed in the routine bnxt_set_default_mac_addr_op()
to destroy and re-create the default filter when the user
changes the default MAC of the port.
Kalesh AP [Thu, 16 Jan 2020 13:04:48 +0000 (18:34 +0530)]
net/bnxt: fix IOVA mapping
Use rte_malloc_virt2iova() to obtain the IO address of a
virtual address obtained through rte_malloc().
Fixed to use the iova address returned by rte_memzone_reserve_aligned()
as the call always returns with populating "mz->iova" with
rte_malloc_virt2iova(mz->addr).
Removed redundant rte_mem_lock_page() call to lock the pages.
Fixes: f55e12f33416 ("net/bnxt: support extended port counters") Cc: stable@dpdk.org Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com> Reviewed-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Kalesh AP [Thu, 16 Jan 2020 13:04:47 +0000 (18:34 +0530)]
net/bnxt: fix probe in FreeBSD
In FreeBSD environment, nic_uio driver does not support interrupts
and rte_intr_callback_register() will fail to register interrupts
which in turn causes bnxt driver probe failure.
Fixed driver to ignore interrupt callback failures in FreeBSD.
Also fixed to not use a dedicated completion ring for async events
from FW and process these events on RXQ0 in FreeBSD.
ppc64le failed when using large physical memory. I found problems in my two
commits in the past.
In commit e072d16f8920 ("vfio: fix expanding DMA area in ppc64le"), I added
a sanity check using a mapped address to resolve an issue around expanding
IOMMU window, but this was not enough, since memory allocation can return
memory anywhere dependent on memory fragmentation. DPDK may still skip DMA
mapping and attempts to unmap non-mapped DMA during expanding IOMMU window.
As a result, SPDK apps using large physical memory frequently failed to
proceed the communication with NVMe and/or went into an infinite loop.
The root cause of the bug was in a gap between memory segments managed by
DPDK and firmware-level DMA mapping. DPDK's memory segments don't contain
the state of DMA mapping, and so, the memesg_walk cannot determine if an
iterated memory segment is mapped or not. This resulted in incorrect DMA
maps and unmaps.
At this time, I added the code to avoid iterating non-mapped memory
segments during DMA mapping. The memseg_walk iterates over memory segments
marked as "used", and so, the code sets memory segments that will be
mapped or unmapped as "free" transiently.
The commit db90b4969e2e ("vfio: retry creating sPAPR DMA window") allows
retring different page levels and sizes to create DMA window. However, this
allows page sizes different from hugepage sizes. This inconsistency caused
failures at the time of DMA mapping after the window creation. This patch
fixes to retry only different page levels.
Fixes: e072d16f8920 ("vfio: fix expanding DMA area in ppc64le") Fixes: db90b4969e2e ("vfio: retry creating sPAPR DMA window") Cc: stable@dpdk.org Signed-off-by: Takeshi Yoshimura <tyos@jp.ibm.com> Reviewed-by: David Christensen <drc@linux.vnet.ibm.com>
Timer, LPM and Distributor libraries no longer use function versioning
and therefore do not need separate build for static and shared version
of libraries.
This patch removes use_function_versioning from their meson build files
and corresponding include from the sources.
Amit Gupta [Mon, 3 Feb 2020 19:49:08 +0000 (13:49 -0600)]
test/hash: split into shorter subtests
hash_readwrite test was taking too much time to complete in Travis.
Test is split into functional and perf test.
perf test is being moved under perf testsuites in meson.
Signed-off-by: Amit Gupta <agupta3@marvell.com> Acked-by: Yipeng Wang <yipeng1.wang@intel.com> Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Thomas Monjalon [Wed, 5 Feb 2020 15:39:23 +0000 (16:39 +0100)]
buildtools: allow pedantic empty pmdinfo
If a driver has no info generated by pmdinfogen,
and if this driver is compiled as pedantic,
then an failure would occur:
drivers/rte_common_mlx5.pmd.c:1: error:
ISO C forbids an empty translation unit [-Werror=pedantic]
Such error is triggered with the new mlx5 common directory.
In order to allow an "empty driver info" compiled in pedantic mode,
the script generating .pmd.c file is modified to add a static string
unconditionnaly.
The minimal generated code is:
static __attribute__((unused)) const char *generator =
"/path/to/dpdk/buildtools/gen-pmdinfo-cfile.sh";
Signed-off-by: Thomas Monjalon <thomas@monjalon.net> Acked-by: Bruce Richardson <bruce.richardson@intel.com>
API makes think that rte_cryptodev_info_get() cannot return
a value >= 3 (RTE_CRYPTO_AEAD_LIST_END in 19.11).
20.02-rc1 was returning 3 (RTE_CRYPTO_AEAD_CHACHA20_POLY1305).
So the ABI compatibility contract was broken.
It could be solved with some function versioning,
but because a lack of time, the feature is reverted for now.
This reverts following commits:
- 6c9f3b347e21 ("cryptodev: add Chacha20-Poly1305 AEAD algorithm")
- 2c512e64d600 ("crypto/qat: support Chacha Poly")
- d55e01f579e1 ("test/crypto: add Chacha Poly cases")
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
David Marchand [Sun, 2 Feb 2020 21:08:34 +0000 (22:08 +0100)]
devtools: add ABI checks
For normal developers, those checks are disabled.
Enabling them requires a configuration that will trigger the ABI dumps
generation as part of the existing devtools/test-build.sh and
devtools/test-meson-builds.sh scripts.
Those checks are enabled in the CI for the default meson options on x86
and aarch64 so that proposed patches are validated via our CI robot.
A cache of the ABI is stored in travis jobs to avoid rebuilding too
often.
Checks can be informational only, by setting ABI_CHECKS_WARN_ONLY when
breaking the ABI in a future release.
Explicit suppression rules have been added on internal structures
exposed to crypto drivers as the current ABI policy does not apply to
them.
This could be improved in the future by carefully splitting the headers
content with application and driver "users" in mind.
We currently have issues reported for librte_crypto recent changes for
which suppression rules have been added too.
Mellanox glue libraries are explicitly skipped as they are not part of
the application ABI.
Signed-off-by: David Marchand <david.marchand@redhat.com> Acked-by: Luca Boccassi <bluca@debian.org>
David Marchand [Sun, 2 Feb 2020 21:08:33 +0000 (22:08 +0100)]
devtools: split meson build helper
No functional change intended, prepare for reusing this code.
The config and compilation parts are separated in helpers.
Unsetting CC is moved to the caller of the helper.
Signed-off-by: David Marchand <david.marchand@redhat.com> Acked-by: Luca Boccassi <bluca@debian.org>
Sunil Kumar Kori [Tue, 28 Jan 2020 05:34:56 +0000 (11:04 +0530)]
examples/l3fwd: split pipelines based on capability
Add infra to split eventdev framework based on event Tx adapter
capability.
If event Tx adapter has internal port capability then we use
`rte_event_eth_tx_adapter_enqueue` to transmitting packets else
we use a SINGLE_LINK event queue to enqueue packets to a service
core which is responsible for transmitting packets.
Signed-off-by: Sunil Kumar Kori <skori@marvell.com> Acked-by: Jerin Jacob <jerinj@marvell.com>
Sunil Kumar Kori [Tue, 28 Jan 2020 05:34:55 +0000 (11:04 +0530)]
examples/l3fwd: add framework for event device
Add framework to enable event device as a producer of packets.
To switch between event mode and poll mode the following options
have been added:
`--mode="eventdev"` or `--mode="poll"`
Allow the user to select the schedule type to be either
RTE_SCHED_TYPE_ORDERED, RTE_SCHED_TYPE_ATOMIC or RTE_SCHED_TYPE_PARALLEL
through:
`--eventq-sched="ordered"` or `--eventq-sched="atomic"` or
`--eventq-sched="parallel"`
Allow the user to specify the number of Rx queues to be connected to
event queue using:
`--event-eth-rxqs`
Poll mode is still the default operation mode.
Signed-off-by: Sunil Kumar Kori <skori@marvell.com> Signed-off-by: Nipun Gupta <nipun.gupta@nxp.com> Acked-by: Jerin Jacob <jerinj@marvell.com>
David Marchand [Mon, 20 Jan 2020 20:35:13 +0000 (21:35 +0100)]
maintainers: set git tree for EAL and core libraries
Without a git tree, the "guess" script from dpdk-ci will ignore all files
that are referenced in those sections and let a patchset land in any
subtree.
Changes in the EAL, mbuf, mempool... go through the main repository.
Signed-off-by: David Marchand <david.marchand@redhat.com> Acked-by: Thomas Monjalon <thomas@monjalon.net>
David Marchand [Tue, 14 Jan 2020 09:18:15 +0000 (10:18 +0100)]
maintainers: set git tree for test applications
Changes on the main test applications should be considered against
subtrees where most of the API changes and development happen:
- testpmd goes through dpdk-next-net as it is mainly about ethdev,
- dpdk-test-compress-perf goes through dpdk-next-crypto since the
compress API goes through this tree,
- dpdk-test-crypto-perf through dpdk-next-crypto,
- dpdk-test-eventdev through dpdk-next-eventdev,
Signed-off-by: David Marchand <david.marchand@redhat.com> Acked-by: Jerin Jacob <jerinj@marvell.com> Acked-by: Thomas Monjalon <thomas@monjalon.net> Acked-by: Ferruh Yigit <ferruh.yigit@intel.com> Acked-by: Andrew Rybchenko <arybchenko@solarflare.com> Acked-by: Akhil Goyal <akhil.goyal@nxp.com>
Unfortunately due to lack of time, I've been unable to even participate to
flow API discussions for several months. Better make it official since this
is not going to improve anytime soon.
This doesn't mean I won't contribute to rte_flow in the future!
Cc: stable@dpdk.org Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com> Acked-by: Ori Kam <orika@mellanox.com>
Ferruh Yigit [Tue, 21 Jan 2020 09:58:01 +0000 (09:58 +0000)]
net/ionic: ignore missing field initializers warning
The compiler warning is:
from .../drivers/net/ionic/ionic_dev.c:7:
.../drivers/net/ionic/ionic_if.h:202:5: note: ‘rsvd’ declared here
u8 rsvd[62];
^
This has been observed with gcc 4.8.5, newer 9+ compiler are not giving
this warning.
Warning is a reminder to the user that there are some fields in the
struct not initialized with the default value.
But the C standard clarifies that in that case the field value will be
zero and code is aware of this behavior, so no initializing to a default
value is intentional and it is safe to ignore this compiler warning.
Adding '-Wno-missing-field-initializers' compiler flag to disable the
warning.
David Marchand [Mon, 2 Dec 2019 15:42:20 +0000 (16:42 +0100)]
eal: remove limitation on cpuset with --lcores
Contrary to the -c/-l options, where a logical core runs on the same
physical core in a 1:1 fashion (example: lcore 0 runs on core 0, lcore
16 runs on core 16), the --lcores option makes it possible to select the
physical cores on which runs a logical core.
However the current parsing code still limits the cpuset to the
[0, RTE_MAX_LCORE] range.
Example, before the patch, on a 24 cores system with RTE_MAX_LCORE == 16:
$ ./master/app/testpmd --no-huge --no-pci -m 512 --log-level *:debug \
--lcores 0@16,1@17 -- -i --total-num-mbufs 2048
EAL: Detected lcore 0 as core 0 on socket 0
EAL: Detected lcore 1 as core 1 on socket 0
EAL: Detected lcore 2 as core 2 on socket 0
EAL: Detected lcore 3 as core 3 on socket 0
EAL: Detected lcore 4 as core 4 on socket 0
EAL: Detected lcore 5 as core 5 on socket 0
EAL: Detected lcore 6 as core 6 on socket 0
EAL: Detected lcore 7 as core 8 on socket 0
EAL: Detected lcore 8 as core 9 on socket 0
EAL: Detected lcore 9 as core 10 on socket 0
EAL: Detected lcore 10 as core 11 on socket 0
EAL: Detected lcore 11 as core 12 on socket 0
EAL: Detected lcore 12 as core 13 on socket 0
EAL: Detected lcore 13 as core 14 on socket 0
EAL: Detected lcore 14 as core 0 on socket 0
EAL: Detected lcore 15 as core 1 on socket 0
EAL: Skipped lcore 16 as core 2 on socket 0
EAL: Skipped lcore 17 as core 3 on socket 0
EAL: Skipped lcore 18 as core 4 on socket 0
EAL: Skipped lcore 19 as core 5 on socket 0
EAL: Skipped lcore 20 as core 6 on socket 0
EAL: Skipped lcore 21 as core 8 on socket 0
EAL: Skipped lcore 22 as core 9 on socket 0
EAL: Skipped lcore 23 as core 10 on socket 0
EAL: Skipped lcore 24 as core 11 on socket 0
EAL: Skipped lcore 25 as core 12 on socket 0
EAL: Skipped lcore 26 as core 13 on socket 0
EAL: Skipped lcore 27 as core 14 on socket 0
EAL: Support maximum 16 logical core(s) by configuration.
EAL: Detected 16 lcore(s)
EAL: Detected 1 NUMA nodes
EAL: invalid parameter for --lcores
We can remove this limitation by using a cpuset_t (which is a more
natural type since this is what gets passed to pthread_setaffinity*
in the end).
After the patch:
$ ./master/app/testpmd --no-huge --no-pci -m 512 --log-level *:debug \
--lcores 0@16,1@17 -- -i --total-num-mbufs 2048
[...]
EAL: Master lcore 0 is ready (tid=7f94217bbc00;cpuset=[16])
EAL: lcore 1 is ready (tid=7f941f491700;cpuset=[17])
Signed-off-by: David Marchand <david.marchand@redhat.com>
Bing Zhao [Mon, 20 Jan 2020 09:43:07 +0000 (11:43 +0200)]
net/mlx5: fix modify actions support limitation
In the root table, there is some limitation of total number of header
modify actions, 16 or 8 for each. But in other tables, there is no
such strict limitation. In an IPv6 case, the IP fields modifying
will occupy more actions than that in IPv4, so the total support
number should be increased in order to support as many actions as
possible for an IPv6 + TCP packet.
And in the meanwhile, the memory consumption should also be taken
into consideration because sometimes only several actions are needed.
The root table checking could also be done in low layer driver and
the error code will be returned if the actions number is over the
maximal supported value.
Fixes: 0e9d00027686 ("net/mlx5: check maximum modify actions number") Cc: stable@dpdk.org Signed-off-by: Bing Zhao <bingz@mellanox.com> Acked-by: Ori Kam <orika@mellanox.com> Acked-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
Suanming Mou [Wed, 8 Jan 2020 10:41:09 +0000 (12:41 +0200)]
net/mlx5: fix meter suffix flow
The meter suffix flow item pointer restore is not correct to decrease
a fixed value. The incorrect operation will cause incorrect match to
the meter suffix flow, the flow create will fail once the magic number
in the wrong offset memory start with RTE_FLOW_ITEM_TYPE_END.
The pointer should decrease the real offset it increases.
Set the decrease value to the real offset the pointer increases to fix
the issue.
Dekel Peled [Thu, 16 Jan 2020 18:36:23 +0000 (20:36 +0200)]
net/mlx5: support GTP
This patch adds to MLX5 PMD support of matching on GTP item,
fields msg_type and teid, according to RFC [1].
GTP item validation and translation functions are added and called.
GTP tunnel type is added to supported tunnels.
The free on completion queue keeps the indices of elts array,
all mbuf stored below this index should be freed on arrival
of normal send completion. In debug version it also contains
an index of completed transmitting descriptor (WQE) to check
queues synchronization.
The new software manged entity is introduced in Tx datapath
- free on completion queue. This queue keeps the information
how many buffers stored in elts array must freed on send
completion. Each element of the queue contains transmitting
descriptor index to be in synch with completion entries (in
debug build only) and the index in elts array to free buffers.
This is preparation step, we are going to store the index
of elts to free on completion in the dedicated free on
completion queue, this patch updates the elts freeing routine
and updates Tx error handling routine to be synced with
coming new queue.
The complete request flag is set once per Tx burst call,
the code of appropriate routine moved to the end of sending
loop. This is preparation step to remove WQE reserved field
usage to store index of elts to free.
Add support for managing RX filters based on MAC and VLAN.
Hardware cannot provide the list of filters, thus we keep
a local list.
Add support for promisc and allmulticast modes.
Signed-off-by: Alfredo Cardigliano <cardigliano@ntop.org> Reviewed-by: Shannon Nelson <snelson@pensando.io>
Initialize LIFs (Logical Interfaces) which represents
external connections. The NIC can multiplex many LIFs
to a single port, but in most setups, LIF0 is the
primary control for the port.
Create a device for each LIF.
Signed-off-by: Alfredo Cardigliano <cardigliano@ntop.org> Reviewed-by: Shannon Nelson <snelson@pensando.io>
Selwin Sebastian [Mon, 20 Jan 2020 16:42:58 +0000 (22:12 +0530)]
net/axgbe: add a HW quirk for register definitions
V1000/R1000 processors are using the same PCI ids for the network
device as SNOWYOWL processor but has altered register definitions
for determining the window settings for the indirect PCS access.
Add support to check for this hardware and if found use the new
register values.
Signed-off-by: Selwin Sebastian <selwin.sebastian@amd.com> Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Ali Alnubani [Mon, 20 Jan 2020 19:51:14 +0000 (19:51 +0000)]
mk: disable OCTEON TX2 EP with old arm64 compilers
OCTEON TX2 isn't built for gcc 4.8.5 as the compiler emits
"internal compiler error" on aarch64. This causes the following
build error when OCTEON TX2 EP is enabled:
/usr/bin/ld: cannot find -lrte_common_octeontx2
collect2: error: ld returned 1 exit status
Fixes: 56d46d13f736 ("raw/octeontx2_ep: add build infra and device probe") Signed-off-by: Ali Alnubani <alialnu@mellanox.com>
net/mlx5: allow allocated mbuf with external buffer
In the Rx datapath the flags in the newly allocated mbufs
are all explicitly cleared but the EXT_ATTACHED_MBUF must be
preserved. It would allow to use mbuf pools with pre-attached
external data buffers.
The vectorized rx_burst routines are updated in order to
inherit the EXT_ATTACHED_MBUF from mbuf pool private
RTE_PKTMBUF_POOL_F_PINNED_EXT_BUF flag.
app/testpmd: add mempool with external data buffers
The new mbuf pool type is added to testpmd. To engage the
mbuf pool with externally attached data buffers the parameter
"--mp-alloc=xbuf" should be specified in testpmd command line.
The objective of this patch is just to test whether mbuf pool
with externally attached data buffers works OK. The memory for
data buffers is allocated from DPDK memory, so this is not
"true" external memory from some physical device (this is
supposed the most common use case for such kind of mbuf pool).
The user should be aware that not all drivers support the mbuf
with EXT_ATTACHED_BUF flags set in newly allocated mbuf (many
PMDs just overwrite ol_flags field and flag value is getting
lost).
The dedicated routine rte_pktmbuf_pool_create_extbuf() is
provided to create mbuf pool with data buffers located in
the pinned external memory. The application provides the
external memory description and routine initializes each
mbuf with appropriate virtual and physical buffer address.
It is entirely application responsibility to register
external memory with rte_extmem_register() API, map this
memory, etc.
The new introduced flag RTE_PKTMBUF_POOL_F_PINNED_EXT_BUF
is set in private pool structure, specifying the new special
pool type. The allocated mbufs from pool of this kind will
have the EXT_ATTACHED_MBUF flag set and initialiazed shared
info structure, allowing cloning with regular mbufs (without
attached external buffers of any kind).
Update detach routine to check the mbuf pool type.
Introduce the special internal version of detach routine to handle
the special case of pinned external bufferon mbuf freeing.
Artur Trybula [Fri, 20 Dec 2019 15:08:39 +0000 (16:08 +0100)]
mem: improve log message for too low memzone segments
In case of too low number of memzone segments user notification
was misleading. This patch improves the description by providing
better explanation about the cause.
Signed-off-by: Artur Trybula <arturx.trybula@intel.com> Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
FreeBSD 13 has changed the definition of vm_page_replace so we need
to have slightly different code paths around this function depending on
the BSD version.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
In (currently unreleased) FreeBSD 13, the CPU_NAND macro has been renamed
to CPU_ANDNOT, so we need to use different DPDK-specific macros depending
on what system-defined ones are present.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Bruce Richardson [Mon, 20 Jan 2020 12:22:18 +0000 (12:22 +0000)]
app/test: remove meson dependency on file in /sys
Meson versions 0.52 and 0.53 are being overly smart and detecting the path
"/sys/devices/system/cpu/present" in the call to cat in
app/test/meson.build and then adding it as a dependency to the build
configuration. This causes issues on systems where the timestamp of that
file always returns the current time, since it means that the build.ninja
file is always out of date, and therefore needs to be rebuilt.
We can fix this by just using a simple shell script to return the coremask
appropriately for BSD and Linux, and removing that code logic from meson -
thereby hiding the use of the /sys file.
Fixes: c70622ac6f72 ("test: detect number of cores with meson") Cc: stable@dpdk.org Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Jerin Jacob [Tue, 14 Jan 2020 21:06:03 +0000 (02:36 +0530)]
mempool: remove memory wastage on non-x86
The existing optimize_object_size() function address the memory object
alignment constraint on x86 for better performance.
Different (micro) architecture may have different memory alignment
constraint for better performance and it not the same as the existing
optimize_object_size().
Some use, XOR(kind of CRC) scheme to enable DRAM channel distribution
based on the address and some may have a different formula.
Introducing arch_mem_object_align() function to abstract
the difference between different (micro) architectures to avoid
wasting memory for mempool object alignment for the architecture
that it is not required to do so.
Details on the amount of memory saving:
Currently, arm64 based architectures use the default (nchan=4,
nrank=1). The worst case is for an object whose size (including mempool
header) is 2 cache lines, where it is optimized to 3 cache lines (+50%).
Olivier Matz [Fri, 17 Jan 2020 14:57:52 +0000 (15:57 +0100)]
mempool: fix populate with small virtual chunks
To populate a mempool with a virtual area, the mempool code calls
rte_mempool_populate_iova() for each iova-contiguous area. It happens
(rarely) that this area is too small to store one object. In this case,
rte_mempool_populate_iova() returns an error, which is forwarded by
rte_mempool_populate_virt().
This case should not throw an error in rte_mempool_populate_virt().
Instead, the area that is too small should just be ignored.
To fix this issue, change the return value of
rte_mempool_populate_iova() to 0 when no object can be populated,
so it can be ignored by the caller. As this would be an API/ABI change,
only do this modification internally for now.
Olivier Matz [Fri, 17 Jan 2020 09:51:49 +0000 (10:51 +0100)]
mempool: fix slow allocation of large pools
When allocating a mempool which is larger than the largest
available area, it can take a lot of time:
a- the mempool calculate the required memory size, and tries
to allocate it, it fails
b- then it tries to allocate the largest available area (this
does not request new huge pages)
c- add this zone to the mempool, this triggers the allocation
of a mem hdr, which request a new huge page
d- back to a- until mempool is populated or until there is no
more memory
This can take a lot of time to finally fail (several minutes): in step
a- it takes all available hugepages on the system, then release them
after it fails.
The problem appeared with commit eba11e364614 ("mempool: reduce wasted
space on populate"), because smaller chunks are now allowed. Previously,
it had to be at least one page size, which is not the case in step b-.
To fix this, implement our own way to allocate the largest available
area instead of using the feature from memzone: if an allocation fails,
try to divide the size by 2 and retry. When the requested size falls
below min_chunk_size, stop and return an error.
Fixes: eba11e364614 ("mempool: reduce wasted space on populate") Cc: stable@dpdk.org Signed-off-by: Olivier Matz <olivier.matz@6wind.com> Acked-by: Anatoly Burakov <anatoly.burakov@intel.com> Reviewed-by: Andrew Rybchenko <arybchenko@solarflare.com> Tested-by: Ali Alnubani <alialnu@mellanox.com>
git.droids-corp.org / dpdk.git / log