Currently, externally created heaps are supposed to be automatically
mapped for VFIO DMA by EAL, however they only do so if, at the time of
heap creation, VFIO is initialized and has at least one device
available. If no devices are available at the time of heap creation (or
if devices were available, but were since hot-unplugged, thus dropping
all VFIO container mappings), then VFIO mapping code would have skipped
over externally allocated heaps.
The fix is two-fold. First, we allow externally allocated memory
segments to be marked as "heap" segments. This allows us to distinguish
between external memory segments that were created via heap API, from
those that were created via rte_extmem_register() API.
Then, we fix the VFIO code to only skip non-heap external segments.
Also, since external heaps are not guaranteed to have valid IOVA
addresses, we will skip those which have invalid IOVA addresses as well.
Fixes: 0f526d674f ("malloc: separate creating memseg list and malloc heap")
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Rajesh Ravi <rajesh.ravi@broadcom.com>
Acked-by: David Marchand <david.marchand@redhat.com>
The rte_vfio_dma_map/unmap API's have been marked as deprecated in
release 19.05. Remove them.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: David Marchand <david.marchand@redhat.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
When requesting DMA mapping to default container, we are meant to
supply the RTE_VFIO_DEFAULT_CONTAINER_FD value, however this is
not handled correctly by get_vfio_cfg_by_container_fd(), because
it only looks at actual fd values and does not check for this
special case.
Fix it to return default container if the fd requested is the
special RTE_VFIO_DEFAULT_CONTAINER_FD value.
Fixes: 4106d89a18 ("vfio: allow DMA map to the default container")
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Acked-by: Shahaf Shuler <shahafs@mellanox.com>
Some of the internal header files have 'rte_' prefix
and some don't.
Remove 'rte_' prefix from all internal header files.
Suggested-by: Thomas Monjalon <thomas@monjalon.net>
Signed-off-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Reviewed-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
Any file with ABI versioned functions needs different macros for shared and
static builds, so we need to accommodate that. Rather than building
everything twice, we just flag to the build system which libraries need
that handling, by setting use_function_versioning in the meson.build files.
To ensure we don't get silent errors at build time due to this meson flag
being missed, we add an explicit error to the function versioning header
file if a known C macro is not defined. Since "make" builds always only
build one of shared or static libraries, this define can be always set, and
so is added to the global CFLAGS. For meson, the build flag - and therefore
the C define - is set for the three libraries that need the function
versioning: "distributor", "lpm" and "timer".
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Tested-by: Andrzej Ostruszka <amo@semihalf.com>
Reviewed-by: Andrzej Ostruszka <amo@semihalf.com>
The compat.h header file provided macros for two purposes:
1. it provided the macros for marking functions as rte_experimental
2. it provided the macros for doing function versioning
Although these were in the same file, #1 is something that is for use by
public header files, which #2 is for internal use only. Therefore, we can
split these into two headers, keeping #1 in rte_compat.h and #2 in a new
file rte_function_versioning.h. For "make" builds, since internal objects
pick up the headers from the "include/" folder, we need to add the new
header to the installation list, but for "meson" builds it does not need to
be installed as it's not for public use.
The rework also serves to allow the use of the function versioning macros
to files that actually need them, so the use of experimental functions does
not need including of the versioning code.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Reviewed-by: Andrzej Ostruszka <amo@semihalf.com>
Starting with kernel version 4.10, there are new min/max MTU values in
net_device structure, which are set to ETH_MIN_MTU and ETH_DATA_LEN by
default. We should be able to change these values to allow MTU more than
1500 to be set on KNI.
Signed-off-by: Igor Ryzhov <iryzhov@nfware.com>
Acked-by: Ferruh Yigit <ferruh.yigit@intel.com>
Now that all elements of the rte_config structure have (deinlined)
accessors, we can hide it.
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
This code belongs to the lcore API, move the prototype to the right
header, then factorize the code into the common code.
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Those functions are used to setup or take control decisions.
Move them into the EAL common code and put them directly in the stable
ABI.
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Define an accessor so that users can write their debug message to the
same stream than the rte_log infrastructure.
Use it in the qat infrastructure.
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Kevin Traynor <ktraynor@redhat.com>
Let's avoid exporting structures without an identified usecase.
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
Remove rte_malloc_virt2phy as announced previously.
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
Remove rte_cpu_check_supported as announced previously.
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
The internal structure of lcore_config does not need to be part of
visible API/ABI. Make it private to EAL.
Rearrange the structure so it takes less memory (and cache footprint).
Since we change the ABI, bump the library version.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
Many features require to store data inside the mbuf. As the room in mbuf
structure is limited, it is not possible to have a field for each
feature. Also, changing fields in the mbuf structure can break the API
or ABI.
This commit addresses these issues, by enabling the dynamic registration
of fields or flags:
- a dynamic field is a named area in the rte_mbuf structure, with a
given size (>= 1 byte) and alignment constraint.
- a dynamic flag is a named bit in the rte_mbuf structure.
The typical use case is a PMD that registers space for an offload
feature, when the application requests to enable this feature. As
the space in mbuf is limited, the space should only be reserved if it
is going to be used (i.e when the application explicitly asks for it).
The registration can be done at any moment, but it is not possible
to unregister fields or flags.
Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Currently, mem config will be mapped without using the virtual
area reservation infrastructure, which means it will be mapped
at an arbitrary location. This may cause failures to map the
shared config in secondary process due to things like PCI
whitelist arguments allocating memory in a space where the
primary has allocated the shared mem config.
Fix this by using virtual area reservation to reserve space for
the mem config, thereby avoiding the problem and reserving the
shared config (hopefully) far away from any normal memory
allocations.
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Not all OS's follow Linux's memory layout, which may lead to
problems following the suggested common address hint absent
of a base-virtaddr flag. Make this address hint OS-specific.
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Moving RTE_CPU* definitions from the common code to the Linux and
FreeBSD rte_os.h file to avoid #ifdef clutter.
Signed-off-by: Pallavi Kadam <pallavi.kadam@intel.com>
Signed-off-by: Antara Ganesh Kolar <antara.ganesh.kolar@intel.com>
Reviewed-by: Ranjit Menon <ranjit.menon@intel.com>
Reviewed-by: Jerin Jacob <jerinj@marvell.com>
Signed-off-by: David Marchand <david.marchand@redhat.com>
Fix off-by-one error in 64bit reciprocal division when divisor is 32bit.
Caught with the unit test:
RTE>>reciprocal_division
Validating unsigned 32bit division.
Validating unsigned 64bit division.
Validating unsigned 64bit division with 32bit divisor.
Division failed, 16983222950483802557/819 = expected 20736535959076681
result 20736535959076682
Validating division by power of 2.
Test Failed
Fixes: 6d45659eac ("eal: add u64-bit variant for reciprocal divide")
Cc: stable@dpdk.org
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Right now RTE_CACHE_ and IOVA definitions are located inside rte_memory.h
That might cause an unwanted inclusions of arch/os specific header files.
See [1] for particular problem example.
Probably the simplest way to deal with such problems -
move these definitions into rte_commmon.h
Note that this move doesn't introduce any change in functionality.
[1] https://bugs.dpdk.org/show_bug.cgi?id=321
Suggested-by: Vipin Varghese <vipin.varghese@intel.com>
Signed-off-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Acked-by: Michel Machado <michel@digirati.com.br>
We are using '--base-virtaddr' in a few places. We have a define for that,
so use it instead.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
According to our docs, only Linuxapp supports base-virtaddr option.
That is, strictly speaking, not true because most of the things
that are attempting to respect base-virtaddr are in common files,
so FreeBSD already *mostly* supports this option in practice.
This commit fixes the remaining bits to explicitly support
base-virtaddr option, and moves the arg parsing from EAL to common
options parsing code. Documentation is also updated to reflect
that all platforms now support base-virtaddr.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
RTE_EAL_ALLOW_INV_SOCKET_ID had been introduced and documented as used
with xen dom0 support (dropped for some time now).
Closely looking at this, the code was changed later and ensures that the
socket id is in the [0..RTE_MAX_NUMA_NODES] range anyway.
Let's drop this dead code and the build option with it.
Fixes: 94ef296414 ("eal/linux: fix numa node detection")
Cc: stable@dpdk.org
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
An ifdef present in eal_memory.c references "RTE_ARCH_PPC64" when
it should actually use "RTE_ARCH_PPC_64". Simple testing revealed
that both the PPC_64 and non-PPC_64 versions of the code involved
work, but the PPC_64 version of the code is retained to be
consistent with other instances in the same file where mmapped
memory is accessed in reverse order on Power platforms.
Fixes: 66cc45e293 ("mem: replace memseg with memseg lists")
Cc: stable@dpdk.org
Signed-off-by: David Christensen <drc@linux.vnet.ibm.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
This patch adds the implementation of the 128-bit atomic compare
exchange API on aarch64. Using 64-bit 'ldxp/stxp' instructions
can perform this operation. Moreover, on the LSE atomic extension
accelerated platforms, it is implemented by 'casp' instructions for
better performance.
Since the '__ARM_FEATURE_ATOMICS' flag only supports GCC-9, this
patch adds a new config flag 'RTE_ARM_FEATURE_ATOMICS' to enable
the 'cas' version on older version compilers.
For octeontx2, we make sure that the lse (and other) extensions are
enabled even if the compiler does not know of the octeontx2 target
cpu.
Since direct x0 register used in the code and cas_op_name() and
rte_atomic128_cmp_exchange() is inline function, based on parent
function load, it may corrupt x0 register aka break aarch64 ABI.
Define CAS operations as rte_noinline functions to avoid an ABI
break [1].
1: https://git.dpdk.org/dpdk/commit/?id=5b40ec6b9662
Suggested-by: Jerin Jacob <jerinj@marvell.com>
Signed-off-by: Phil Yang <phil.yang@arm.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Tested-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
This ensures secondary processes never have to calculate the TSC rate
themselves, which can be noticeable in VMs that don't have access to
arch-specific detection mechanism (such as CPUID leaf 0x15 or MSR 0xCE
on x86).
Since rte_mem_config is now internal to the EAL library, we can add
tsc_hz without ABI breakage concerns.
Reduces rte_eal_init() execution time in a secondary process from 165ms
to 66ms on my test system.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
For a valid service, the core mask of the service
is checked against the current core and the corresponding
entry in the active_on_lcore array is set or reset.
Upto 8 cores share the same cache line for their
service active_on_lcore array entries since each entry is a uint8_t.
Some number of these entries also share the cache line with
the internal_flags member of struct rte_service_spec_impl,
hence this false sharing also makes the service_valid() check
expensive.
Eliminate false sharing by moving the active_on_lcore array to
a per-core data structure. The array is now indexed by service id.
Signed-off-by: Nikhil Rao <nikhil.rao@intel.com>
Acked-by: Gage Eads <gage.eads@intel.com>
Acked-by: Harry van Haaren <harry.van.haaren@intel.com>
This code was added 7+ years ago in
commit fb022b85ba ("timer: check TSC reliability")
presumably when variant TSCs were still somewhat common.
But this code doesn't do anything except print a warning,
and the warning doesn't give any kind of advice to the user,
so let's just remove it.
While the warning has no functional meaning, the /proc/cpuinfo
parsing consumes a non-trivial amount of time which is especially
noticeable in secondary processes.
On my test system, it consumes 21ms out of the 66ms total execution
time for rte_eal_init() in a secondary process.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
The rte_atomic64_exchange operation for ppc_64 incorrectly linked
back to a 32 bit generic operation (__atomic_exchange_4) rather than
the 64 bit generic operation (__atomic_exchange_8). As a result,
applications that used rte_eth_link_get_nowait() would only receive
the link speed, they would not receive the link state, link duplex,
or link autoneg properties.
Fixes: ff2863570f ("eal: introduce atomic exchange operation")
Cc: stable@dpdk.org
Signed-off-by: David Christensen <drc@linux.vnet.ibm.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
This patch adds support to allow users enable/disable allmulticast mode for
kni interface.
This requirement comes from bugzilla 312, more details can refer to:
https://bugs.dpdk.org/show_bug.cgi?id=312
Bugzilla ID: 312
Signed-off-by: Xiaolong Ye <xiaolong.ye@intel.com>
Acked-by: Ferruh Yigit <ferruh.yigit@intel.com>
EAL should always use rte_log instead of putting errors to
stderr (which maybe redirected to /dev/null in a daemon).
Also checks for null before rte_free are unnecessary.
Minor code consistency improvements.
Fixes: 21698354c8 ("service: introduce service cores concept")
Cc: stable@dpdk.org
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Harry van Haaren <harry.van.haaren@intel.com>
Acked-by: Harry van Haaren <harry.van.haaren@intel.com>
Have rte_eal_config_reattach clean up the mapped address which is a valid
address but not the one intended.
Coverity issue: 343439
Fixes: 4e8854ae89 ("eal: do not panic on shared memory init")
Fixes: b149a70642 ("eal/freebsd: add config reattach in secondary process")
Cc: stable@dpdk.org
Signed-off-by: Arnon Warshavsky <arnon@qwilt.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
The code checks both rte_mp_request_sync() return code and that the number
of messages in the reply equals 1. If rte_mp_request_sync() succeeds but
there was more than one message, those messages would get leaked.
Found via code review by Anatoly Burakov of patches that used the vhost
code as a template for using rte_mp_request_sync().
Fixes: 83a73c5fef ("vfio: use generic multi-process channel")
Cc: stable@dpdk.org
Reported-by: Anatoly Burakov <anatoly.burakov@intel.com>
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
Restrict this header inclusion to its real users.
Fixes: 028669bc9f ("eal: hide shared memory config")
Cc: stable@dpdk.org
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
ARM is supporting maximum 4 hugepage sizes (64K, 2M, 32M
and 1G) when granule is 4KB since very long and DPDK
support maximum 3 hugepage sizes.
With all 4 hugepage sizes enabled, applications and some
stacks like VPP which are working over DPDK and using
"in-memory" eal option, or using separate mount points
on ARM based platform, fails at huge page initialization,
reporting error messages from eal:
EAL: FATAL: Cannot get hugepage information.
EAL: Cannot get hugepage information.
EAL: Error - exiting with code: 1
This issue is originated from Linux 5.0
(a21b0b78eaf7 "arm64: hugetlb: Register hugepages during arch init")
where kernel is by default creating directories for each supported
hugepage size in /sys/kernel/mm/hugepages/
On earlier Stable Kernel LTR's, the directories visible in
/sys/kernel/mm/hugepages/ were dependent upon what hugepage
sizes are configured at boot time.
This change increases the maximum supported hugepage sizes
to 4 for ARM based platforms.
Cc: stable@dpdk.org
Signed-off-by: Gagandeep Singh <g.singh@nxp.com>
Signed-off-by: Nipun Gupta <nipun.gupta@nxp.com>
This function has never been used outside of this code unit.
Mark it static and remove it from the eal internal header.
Fixes: 9e29251b2a ("eal: thread affinity API")
Cc: stable@dpdk.org
Signed-off-by: David Marchand <david.marchand@redhat.com>
When using --no-huge mode, dynamic allocation is not supported.
Because of this limitation, the option --legacy-mem is implied
and -m may be needed to specify the amount of memory to allocate.
Otherwise the default amount MEMSIZE_IF_NO_HUGE_PAGE will be allocated.
The option --socket-mem can also be used with --legacy-mem
when hugepages are supported.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
The ctrl thread cpu affinity setting has been broken when using --lcores.
Using -l/-c options makes each lcore associated to a physical cpu in a 1:1
fashion.
On the contrary, when using --lcores, each lcore cpu affinity can be set
to a list of any online cpu on the system.
To handle both cases, each lcore cpu affinity is considered and removed
from the process startup cpu affinity.
Introduced macros to manipulate dpdk cpu sets in both Linux and FreeBSD.
Examples on a 8 cores Linux system:
$ cd /sys/fs/cgroup/cpuset/
$ mkdir dpdk
$ cd dpdk
$ echo 4-7 > cpuset.cpus
$ echo 0 > cpuset.mems
$ echo $$ > tasks
Before the fix:
$ ./master/app/testpmd --master-lcore 0 --lcores '(0,7)@(7,4,5)' \
--no-huge --no-pci -m 512 -- -i --total-num-mbufs=2048
8427 cpu_list=4-5,7 testpmd
8428 cpu_list=4-6 eal-intr-thread
8429 cpu_list=4-6 rte_mp_handle
8430 cpu_list=4-5,7 lcore-slave-7
$ taskset -c 7 \
./master/app/testpmd --master-lcore 0 --lcores '(0,7)@(7,4,5)' \
--no-huge --no-pci -m 512 -- -i --total-num-mbufs=2048
EAL: Detected 8 lcore(s)
EAL: Detected 1 NUMA nodes
EAL: Failed to create thread for interrupt handling
EAL: FATAL: Cannot init interrupt-handling thread
EAL: Cannot init interrupt-handling thread
PANIC in main():
Cannot init EAL
After the fix:
$ ./master/app/testpmd --master-lcore 0 --lcores '(0,7)@(7,4,5)' \
--no-huge --no-pci -m 512 -- -i --total-num-mbufs=2048
15214 cpu_list=4-5,7 testpmd
15215 cpu_list=6 eal-intr-thread
15216 cpu_list=6 rte_mp_handle
15217 cpu_list=4-5,7 lcore-slave-7
$ taskset -c 7 \
./master/app/testpmd --master-lcore 0 --lcores '(0,7)@(7,4,5)' \
--no-huge --no-pci -m 512 -- -i --total-num-mbufs=2048
15297 cpu_list=4-5,7 testpmd
15298 cpu_list=4-5,7 eal-intr-thread
15299 cpu_list=4-5,7 rte_mp_handle
15300 cpu_list=4-5,7 lcore-slave-7
Bugzilla ID: 322
Fixes: c3568ea376 ("eal: restrict control threads to startup CPU affinity")
Cc: stable@dpdk.org
Reported-by: Johan Källström <johan.kallstrom@ericsson.com>
Signed-off-by: David Marchand <david.marchand@redhat.com>
When IOMMU is not available, /sys/kernel/iommu_groups will not be
populated. This is happening since at least 3.6 when VFIO support
was added. If the directory is empty, EAL should not pick IOVA as
VA as the default IOVA mode.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Tested-by: Jerin Jacob <jerinj@marvell.com>
Reviewed-by: Jerin Jacob <jerinj@marvell.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
This was missed when promoting this API to stable.
Fixes: 7a0ac7cdb4 ("service: promote experimental functions to stable")
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Gage Eads <gage.eads@intel.com>
Sort the experimental symbols per release to make it easier/quicker to
check for how long we have them.
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Ferruh Yigit <ferruh.yigit@intel.com>
Acked-by: Michael Santana <msantana@redhat.com>
This reverts commit debacba029.
Reverting this patch as it currently breaks the initialization of
telemetry, more investigation is ongoing to fix the issue for the
printed error message for unrecognized argument.
Fixes: debacba029 ("eal: fix parsing option --telemetry")
Cc: stable@dpdk.org
Signed-off-by: Sean Morrissey <sean.morrissey@intel.com>
When using IOVA as VA mode, there is no need to map segments
page by page. This normally isn't a problem, but it becomes one
when attempting to use DPDK in no-huge mode, where VFIO subsystem
simply runs out of space to store mappings.
Fix this for x86 by triggering different callbacks based on whether
IOVA as VA mode is enabled.
Fixes: 73a6390859 ("vfio: allow to map other memory regions")
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Andrius Sirvys <andrius.sirvys@intel.com>
Add new ack interrupt API to avoid using
VFIO_IRQ_SET_ACTION_TRIGGER(rte_intr_enable()) for
acking interrupt purpose for VFIO based interrupt handlers.
This implementation is specific to Linux.
Using rte_intr_enable() for acking interrupt has below issues
* Time consuming to do for every interrupt received as it will
free_irq() followed by request_irq() and all other initializations
* A race condition because of a window between free_irq() and
request_irq() with packet reception still on and device still
enabled and would throw warning messages like below.
[158764.159833] do_IRQ: 9.34 No irq handler for vector
In this patch, rte_intr_ack() is a no-op for VFIO_MSIX/VFIO_MSI interrupts
as they are edge triggered and kernel would not mask the interrupt before
delivering the event to userspace and we don't need to ack.
Signed-off-by: Nithin Dabilpuram <ndabilpuram@marvell.com>
Signed-off-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Shahed Shaikh <shshaikh@marvell.com>
Signed-off-by: David Marchand <david.marchand@redhat.com>
This reverts commit 89aac60e0b.
"vfio: fix interrupts race condition"
The above mentioned commit moves the interrupt's eventfd setup
to probe time but only enables one interrupt for all types of
interrupt handles i.e VFIO_MSI, VFIO_LEGACY, VFIO_MSIX, UIO.
It works fine with default case but breaks below cases specifically
for MSIX based interrupt handles.
* Applications like l3fwd-power that request rxq interrupts
while ethdev setup.
* Drivers that need > 1 MSIx interrupts to be configured for
functionality to work.
VFIO PCI for MSIx expects all the possible vectors to be setup up
when using VFIO_IRQ_SET_ACTION_TRIGGER so that they can be
allocated from kernel pci subsystem. Only way to increase the number
of vectors later is first free all by using VFIO_IRQ_SET_DATA_NONE
with action trigger and then enable new vector count.
Above commit changes the behavior of rte_intr_[enable|disable] to
only mask and unmask unlike earlier behavior and thereby
breaking above two scenarios.
Fixes: 89aac60e0b ("vfio: fix interrupts race condition")
Cc: stable@dpdk.org
Signed-off-by: Nithin Dabilpuram <ndabilpuram@marvell.com>
Signed-off-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Stephen Hemminger <stephen@networkplumber.org>
Tested-by: Shahed Shaikh <shshaikh@marvell.com>
Tested-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Marchand <david.marchand@redhat.com>
