Rather than have two files that keeps getting out of sync, let's
annotate the version.map to generate the Windows export file.
Some mlx5 symbols (haswell_broadwell_cpu, mlx5_glue, mlx5_os_*) were
only exported for Windows.
All of them are available and used by Linux too, so this patch adds
them in version.map.
Note: Existing version.map annotation achieved with:
$ for dir in lib/librte_eal drivers/common/mlx5; do
./buildtools/map-list-symbol.sh $dir/*.map |
while read file version sym; do
! git grep -qw $sym $dir/*.def || continue;
sed -i -e "s/$sym;/$sym; # WINDOWS_NO_EXPORT/" $dir/*.map;
done;
done
Signed-off-by: David Marchand <david.marchand@redhat.com>
Tested-by: Tal Shnaiderman <talshn@nvidia.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
rte_service_get_id() was removed in v17.11 but the API description
still referenced it and a version node was still present in EAL map.
Fixes: 8edc9aaaf2 ("service: use id in get by name function")
Cc: stable@dpdk.org
Signed-off-by: David Marchand <david.marchand@redhat.com>
Reviewed-by: Harry van Haaren <harry.van.haaren@intel.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Avoid race with unregister interrupt handler if interrupt
source has some active callbacks at the moment, use wrapper
around rte_intr_callback_unregister() to check for -EAGAIN
return value and to loop until rte_intr_callback_unregister()
succeeds.
Signed-off-by: Renata Saiakhova <renata.saiakhova@ekinops.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Harman Kalra <hkalra@marvell.com>
When using huge_unlink we unlink the segment right
after allocation. Although we unlink the file we keep
the fd in fd_list so file still exist just the path deleted.
When freeing the hugepage we need to close the fd and assign
it with (-1) in fd_list for the page to be released.
The current flow fails rte_malloc in the following flow when working
with --huge-unlink option:
1. alloc_seg() for segment A -
We allocate a segment, unlink the path to the segment
and keep the file descriptor in fd_list.
2. free_seg() for segment A -
We clear the segment metadata and return - without closing fd
or assigning (-1) in fd list.
3. alloc_seg() for segment A again -
We find segment A as available, try to allocate it,
find the old fd in fd_list try to unlink it
as part of alloc_seg() but failed because path doesn't exist.
The impact of such error is falsely failing rte_malloc()
although we have hugepages available.
Fixes: d435aad37d ("mem: support --huge-unlink mode")
Cc: stable@dpdk.org
Signed-off-by: Roy Shterman <roy.shterman@vastdata.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
The names of the prototypes pci_probe_t and pci_remove_t
are missing a prefix rte_.
These function types are simply renamed.
No compatibility break is expected for the applications
because it is considered as an internal name in the driver interface.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Reviewed-by: Xueming Li <xuemingl@nvidia.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
The name of the constant PCI_ANY_ID was missing RTE_ prefix.
It is renamed, and the old name becomes a deprecated alias.
While renaming, the duplicate definitions in rte_bus_pci.h
are removed to keep only those in rte_pci.h.
Note: rte_pci.h is included in rte_bus_pci.h
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Reviewed-by: Parav Pandit <parav@nvidia.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Currently, when we set the pstate governor to "performance", we check if
it is already set to this value, and if it is, we skip setting it.
However, we never save this value anywhere, so that next time we come
back and request the governor to be set to its original value, the
original value is empty.
Fix it by saving the original pstate governor first. While we're at it,
replace `strlcpy` with `rte_strscpy`.
Fixes: e6c6dc0f96 ("power: add p-state driver compatibility")
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Reshma Pattan <reshma.pattan@intel.com>
Previous fix for base frequency handling in pstate mode introduced a
couple of issues:
- When base_frequency file does not exist, it simply bails out because
of what appears to be accidental addition of FOPEN_OR_ERR_RET. This is
incorrect, as absence of this file is not fatal and is in fact
expected on kernel versions earlier than 5.3
- When base_frequency file does exist, it gets opened, but never gets
closed, resulting in a resource leak
Both issues also manifest themselves as Coverity defects (dead code, and
a resource leak), so this fix addresses both.
Coverity issue: 369693, 369694
Bugzilla ID: 668
Fixes: 4db9587bbf ("power: check sysfs base frequency")
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Reshma Pattan <reshma.pattan@intel.com>
If the rtd theme is available, passing it by name is enough to select
it. Sphinx itself recognises the "sphinx_rtd_theme" name as a special
case and tries to find its path automatically.
On the other hand, passing a html_theme_path makes sphinx parse all
themes availables in this path, which in some environment (like GHA) is
/usr/share and makes sphinx error on the first zipfile it finds (in GHA,
some Azure CLI thingy) that has no sphinx theme in it.
Fixes: 46562be650 ("doc: import sphinx rtd theme when available")
Cc: stable@dpdk.org
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Aaron Conole <aconole@redhat.com>
The HW virtq object can be destroyed either when the device is closed or
when the state of the virtq becomes disabled.
Some parameters of the virtq should continue to be managed when the
virtq state is changed but all of them must be initialized when the
device is closed.
Wrongly, the enable parameter stayed on when the device is closed what
might cause creation of invalid virtq in the next time a device is
assigned to the driver.
Clean all the virtqs memory when the device is closed.
Fixes: c47d6e8333 ("vdpa/mlx5: support queue update")
Cc: stable@dpdk.org
Signed-off-by: Matan Azrad <matan@nvidia.com>
Acked-by: Xueming Li <xuemingl@nvidia.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
The private virtio_user_dev structure embeds a virtio_hw which itself
contains the ethdev port_id.
Make use of it and remove the duplicate port_id field.
Signed-off-by: David Marchand <david.marchand@redhat.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Short option "s" was passed to getopt_long function, while there was
no condition on this option.
Fixes: f5188211c7 ("examples/vhost_crypto: add sample application")
Cc: stable@dpdk.org
Signed-off-by: Ibtisam Tariq <ibtisam.tariq@emumba.com>
Acked-by: Fan Zhang <roy.fan.zhang@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Similar as single dequeue, the multiple accesses of descriptor length
will lead to potential risk. One-time access of descriptor length can
eliminate this risk.
Fixes: 75ed516978 ("vhost: add packed ring batch dequeue")
Cc: stable@dpdk.org
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Similar as split ring, the multiple accesses of descriptor length will
lead to potential risk. One-time access of descriptor length can
eliminate this risk.
Fixes: 2f3225a7d6 ("vhost: add vector filling support for packed ring")
Cc: stable@dpdk.org
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
In vhost datapath, descriptor's length are mostly used in two coherent
operations. First step is used for address translation, second step is
used for memory transaction from guest to host. But the interval between
two steps will give a window for malicious guest, in which can change
descriptor length after vhost calculated buffer size. Thus may lead to
buffer overflow in vhost side. This potential risk can be eliminated by
accessing the descriptor length once.
Fixes: 1be4ebb1c4 ("vhost: support indirect descriptor in mergeable Rx")
Cc: stable@dpdk.org
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
This patch fixes unchecked return value for rte_vhost_get_mem_table(),
which is reported by coverity.
Coverity issue: 364233
Fixes: ca059fa5e2 ("examples/vhost: demonstrate the new generic APIs")
Cc: stable@dpdk.org
Signed-off-by: Chenbo Xia <chenbo.xia@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
The return value of rte_pci_read_config should be checked.
Coverity issue: 302860
Fixes: a3f8150eac ("net/ifcvf: add ifcvf vDPA driver")
Cc: stable@dpdk.org
Signed-off-by: Xiao Wang <xiao.w.wang@intel.com>
Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Avoid calling rte_vhost_get_vhost_ring_inflight() and
rte_vhost_get_vring_base_from_inflight() when
VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD is not set.
Signed-off-by: Keiichi Watanabe <keiichiw@chromium.org>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Add rte_vhost_get_negotiated_protocol_features, which returns a set of
enabled protocol features.
Signed-off-by: Keiichi Watanabe <keiichiw@chromium.org>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
This patch moves vhost_virtqueue struct fields in order
to both optimize packing and move hot fields on the first
cachelines.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Tested-by: Balazs Nemeth <bnemeth@redhat.com>
This patch moves the per-virtqueue's dirty logging cache
out of the virtqueue struct, by allocating it dynamically
only when live-migration is enabled.
It saves 8 cachelines in vhost_virtqueue struct.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Tested-by: Balazs Nemeth <bnemeth@redhat.com>
This patch removes the "backend" field of the
vhost_virtqueue struct, which is not used by the
library.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Tested-by: Balazs Nemeth <bnemeth@redhat.com>
This patch optimizes packing of the virtqueue
struct by moving fields around to fill holes.
Offset field is not used and so can be removed.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Tested-by: Balazs Nemeth <bnemeth@redhat.com>
While it is worth clarifying whether the fake mbuf
in virtnet_rx struct is really necessary, it is sure
that it heavily impacts cache usage by being part of
the struct. Indeed, it uses two cachelines, and
requires alignment on a cacheline.
Before this series, it means it took 120 bytes in
virtnet_rx struct:
struct virtnet_rx {
struct virtqueue *vq; /*0 8*/
/* XXX 56 bytes hole, try to pack */
/* --- cacheline 1 boundary (64 bytes) --- */
struct rte_mbuf fake_mbuf __attribute__((__aligned__(64))); /*64 128*/
/* --- cacheline 3 boundary (192 bytes) --- */
This patch allocates it using malloc in order to optimize
virtnet_rx cache usage and so virtqueue cache usage.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Tested-by: Balazs Nemeth <bnemeth@redhat.com>
This patch improves the error path of virtio_init_queue(),
by cleaning in reversing order all resources that have
been allocated.
Suggested-by: Chenbo Xia <chenbo.xia@intel.com>
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Tested-by: Balazs Nemeth <bnemeth@redhat.com>
Vrings are part of the virtqueues, so we don't need
to have a pointer to it in Vrings descriptions.
Instead, let's just subtract from its offset to
calculate virtqueue address.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Tested-by: Balazs Nemeth <bnemeth@redhat.com>
The member page_offset is always zero. Having this in the qede_rx_entry
makes it larger than it needs to be and this has cache performance
implications so remove that field. In addition, since qede_rx_entry only
has an rte_mbuf*, remove the definition of qede_rx_entry.
Signed-off-by: Balazs Nemeth <bnemeth@redhat.com>
Reviewed-by: Igor Russkikh <irusskikh@marvell.com>
While handling the current mbuf, pull the next mbuf into the cache. Note
that the last four mbufs pulled into the cache are not handled, but that
doesn't matter.
Signed-off-by: Balazs Nemeth <bnemeth@redhat.com>
Reviewed-by: Igor Russkikh <irusskikh@marvell.com>
Ensure that, while ecore_chain_get_cons_idx is running, txq->hw_cons_ptr
is prefetched. This shows a slight performance improvement.
Signed-off-by: Balazs Nemeth <bnemeth@redhat.com>
Reviewed-by: Igor Russkikh <irusskikh@marvell.com>
rte_pktmbuf_free_bulk calls rte_mempool_put_bulk with the number of
pending packets to return to the mempool. In contrast, rte_pktmbuf_free
calls rte_mempool_put that calls rte_mempool_put_bulk with one object.
An important performance related downside of adding one packet at a time
to the mempool is that on each call, the per-core cache pointer needs to
be read from tls while a single rte_mempool_put_bulk only reads from the
tls once.
Signed-off-by: Balazs Nemeth <bnemeth@redhat.com>
Reviewed-by: Igor Russkikh <irusskikh@marvell.com>
The ring txq->sw_tx_ring is managed with txq->sw_tx_cons. As long as
txq->sw_tx_cons is correct, there is no need to check if
txq->sw_tx_ring[idx] is null explicitly.
Signed-off-by: Balazs Nemeth <bnemeth@redhat.com>
Reviewed-by: Igor Russkikh <irusskikh@marvell.com>
Calling ecore_chain_get_cons_idx repeatedly is slower than calling it
once and using the result for the remainder of qede_process_tx_compl.
Signed-off-by: Balazs Nemeth <bnemeth@redhat.com>
Reviewed-by: Igor Russkikh <irusskikh@marvell.com>
Each sw_tx_ring entry was of type struct qede_tx_entry:
struct qede_tx_entry {
struct rte_mbuf *mbuf;
uint8_t flags;
};
Leaving the unused flags member here has a few performance implications.
First, each qede_tx_entry takes up more memory which has caching
implications as less entries fit in a cache line while multiple entries
are frequently handled in batches. Second, an array of qede_tx_entry
entries is incompatible with existing APIs that expect an array of
rte_mbuf pointers. Consequently, an extra array need to be allocated
before calling such APIs and each entry needs to be copied over.
This patch omits the flags field and replaces the qede_tx_entry entry
by a simple rte_mbuf pointer.
Signed-off-by: Balazs Nemeth <bnemeth@redhat.com>
Reviewed-by: Igor Russkikh <irusskikh@marvell.com>
Modification of the 802.1Q Tag Identifier, VXLAN Network
Identifier or GENEVE Network Identifier is not supported.
Reject attempt to modify these fields via the MODIFY_FIELD
action and document this mlx5 driver limitation.
Fixes: 641dbe4fb0 ("net/mlx5: support modify field flow action")
Cc: stable@dpdk.org
Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
There is a limitation about copying one header field to another for
the Flow group 0. Such copy action is not allowed there. But setting
a header field with an immediate value is perfectly fine.
Allow the MODIFY_FIELD action on group 0 in case the source field
is an immediate value or a pointer to it.
Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
The MODIFY_FIELD action requires the extended metadata support
in order to manipulate on MARK register. Check if it is supported
and reject the MODIFY_FIELD action if it is not.
Fixes: 641dbe4fb0 ("net/mlx5: support modify field flow action")
Cc: stable@dpdk.org
Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Masks that used to modify a packet field must be in a big
endian format. Convert then to BE to ensure proper modification.
Fixes: 641dbe4fb0 ("net/mlx5: support modify field flow action")
Cc: stable@dpdk.org
Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Add a validation check to make sure that the specified width
for MODIFY_FIELD RTE action is not bigger than a field size.
Fixes: 641dbe4fb0 ("net/mlx5: support modify field flow action")
Cc: stable@dpdk.org
Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
To simplify BlueField HPF representor(vf[-1]) probe, this patch allows
probe it with "sf" syntax: "sf[-1]".
Signed-off-by: Xueming Li <xuemingl@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
In case of kernel bonding device, counter was read from first bonding PF
member.
This patch reads all member PFs and sums to get bond xstats.
Signed-off-by: Xueming Li <xuemingl@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
With kernel bonding, there was an error when setting VF MAC address
through representor. The Netlink API requires ifindex of owner PF, not
bonding device ifindex.
Uses owner PF ifindex to modify VF default MAC in case of bonding
device.
Fixes: c21e5facf7 ("net/mlx5: use bond index for netdev operations")
Cc: stable@dpdk.org
Signed-off-by: Xueming Li <xuemingl@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Since kernel bonding netdev doesn't provide statistics counter that
reflects all member ports, PMD has to manually summarize counters from
each member ports.
As a preparation, this patch collects bonding member port information
and saves to shared context data.
Signed-off-by: Xueming Li <xuemingl@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
To probe representors from different kernel bonding PFs, had to specify
2 separate devargs like this:
-a 03:00.0,representor=pf0vf[0-3] -a 03:00.0,representor=pf1vf[0-3]
This patch supports range or list of PF section in devargs, so the
alternative short devargs of above is:
-a 03:00.0,representor=pf[0-1]vf[0-3]
Signed-off-by: Xueming Li <xuemingl@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
To probe representor on 2nd PF of kernel bonding device, had to specify
PF1 BDF in devarg:
<PF1_BDF>,representor=0
When closing bonding device, all representors had to be closed together
and this implies all representors have to use primary PF of bonding
device. So after probing representor port on 2nd PF, when locating new
probed device using device argument, the filter used 2nd PF as PCI
address and failed to locate new device.
Conflict happened by using current representor devargs:
- Use PCI BDF to specify representor owner PF
- Use PCI BDF to locate probed representor device.
- PMD uses primary PCI BDF as PCI device.
To resolve such conflicts, new representor syntax is introduced here:
<primary BDF>,representor=pfXvfY
All representors must use primary PF as owner PCI device, PMD internally
locate owner PCI address by checking representor "pfX" part. To EAL, all
representors are registered to primary PCI device, the 2nd PF is hidden
to EAL, thus all search should be consistent.
Same to VF representor, HPF (host PF on BlueField) uses same syntax to
probe, example: representor=pf1vf[0-3,-1]
This patch also adds pf index into kernel bonding representor port name:
<BDF>_<ib_name>_representor_pf<X>vf<Y>
Signed-off-by: Xueming Li <xuemingl@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
With kernel bonding, representors on second PF are being probed by
devargs:
<primary_bdf>,representor=pf1vf<N>
No need to save primary PF port ID and lookup when probing sibling
ports, revert patch [1]
[1]:
commit e6818853c0 ("net/mlx5: set representor to first PF in bonding mode")
Signed-off-by: Xueming Li <xuemingl@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
This patch adds support for SF representor. Similar to VF representor,
switch port name of SF representor in phys_port_name sysfs key is
"pf<x>sf<y>".
Device representor argument is "representors=sf[list]", list member
could be mix of instance and range. Example:
representors=sf[0,2,4,8-12,-1]
To probe VF representor and SF representor, need to separate into 2
devices:
-a <BDF>,representor=vf[list] -a <BDF>,representor=sf[list]
Signed-off-by: Xueming Li <xuemingl@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
This patch supports representor name parsing for SF.
In sysfs, representor name stored under "phys_port_name" sysfs key,
similar to VF representor, switch port name of SF representor is
"pf<x>sf<y>".
For netlink message, net SF type is supported.
Examples:
pf0sf1
pf0sf[0-3]
Signed-off-by: Xueming Li <xuemingl@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
Coverity flags that 'ctx->tunnel' variable is used before
it's checked for NULL. This patch fixes this issue.
Coverity issue: 366201
Fixes: 868d2e342c ("net/mlx5: fix tunnel offload hub multi-thread protection")
Cc: stable@dpdk.org
Signed-off-by: Yunjian Wang <wangyunjian@huawei.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
The ICE_RSS_ANY_HEADERS will try to enable outer RSS for
non-tunnel case and inner RSS for tunnel case. This confuse
user.
As we already have ICE_RSS_INNER_HEADER for tunnel case,
So, replace ICE_RSS_ANY_HEADERS with ICE_RSS_OUTER_HEADERS
for all exist flow which only specified the outer pattern.
To enable inner RSS for any tunnel cases, a separated rule
should be enabled.
The patch also remove some unnecessary condition check for GTPU
in base code, as we already can support outer RSS for GTPU.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
Acked-by: Xuan Ding <xuan.ding@intel.com>
