This patch verifies that the number of Rx queues configured for RSS
is supported by the device hardware.
RSS support in mlx4 requires contiguous chunk of QPs to be reserved,
there is a hardware limitation on the amount of contiguous QPs which
is reported by the hardware. Ignoring this value will cause Rx queues
creation to fail.
Cc: stable@dpdk.org
Signed-off-by: Moti Haimovsky <motih@mellanox.com>
Acked-by: Matan Azrad <matan@mellanox.com>
A constructor is usually declared with RTE_INIT* macros.
As it is a static function, no need to declare before its definition.
The macro is used directly in the function definition.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Temporary IB device context and list are not freed in case of a successful
initialization of the device.
This issue is caused by the two following commits, the first of which
causes initialization to return early, while the second one goes a bit
overboard while switching to negative errno values; an internal variable
(err) is needed to tell success from failure at the end of the function
since rte_errno is not reliable enough.
Fixes: f2318196c7 ("net/mlx4: remove limitation on number of instances")
Fixes: 9d14b27308 ("net/mlx4: standardize on negative errno values")
Cc: stable@dpdk.org
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Although uncommon, applications may destroy a device immediately after
probing it without going through dev_configure() first.
This patch addresses a crash which occurs when mlx4_dev_close() calls
mlx4_mr_release() due to an uninitialized entry in the private structure.
In addition MR cache init takes place on the device configuration.
When the device is re-configured multiple times, for example when
changing the number of queue on the flight, deadlock can happen.
This patch moved MR cache init from device configuration function to
probe function to make sure init only once.
Fixes: 9797bfcce1 ("net/mlx4: add new memory region support")
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Xueming Li <xuemingl@mellanox.com>
Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
A new hook function is added and called inside the PMDs at the end
of the device probing:
- in primary process, after allocating, init and config
- in secondary process, after attaching and local init
This new function is almost empty for now.
It will be used later to add some post-initialization processing.
For the PMDs calling the helpers rte_eth_dev_create() or
rte_eth_dev_pci_generic_probe(), the hook rte_eth_dev_probing_finish()
is called from here, and not in the PMD itself.
Note that the helper rte_eth_dev_create() could be used more,
especially for vdevs, avoiding some code duplication in PMDs.
Cc: stable@dpdk.org
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Reviewed-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
This is the new design of Memory Region (MR) for mlx PMD, in order to:
- Accommodate the new memory hotplug model.
- Support non-contiguous Mempool.
There are multiple layers for MR search.
L0 is to look up the last-hit entry which is pointed by mr_ctrl->mru (Most
Recently Used). If L0 misses, L1 is to look up the address in a fixed-sized
array by linear search. L0/L1 is in an inline function -
mlx4_mr_lookup_cache().
If L1 misses, the bottom-half function is called to look up the address
from the bigger local cache of the queue. This is L2 - mlx4_mr_addr2mr_bh()
and it is not an inline function. Data structure for L2 is the Binary Tree.
If L2 misses, the search falls into the slowest path which takes locks in
order to access global device cache (priv->mr.cache) which is also a B-tree
and caches the original MR list (priv->mr.mr_list) of the device. Unless
the global cache is overflowed, it is all-inclusive of the MR list. This is
L3 - mlx4_mr_lookup_dev(). The size of the L3 cache table is limited and
can't be expanded on the fly due to deadlock. Refer to the comments in the
code for the details - mr_lookup_dev(). If L3 is overflowed, the list will
have to be searched directly bypassing the cache although it is slower.
If L3 misses, a new MR for the address should be created -
mlx4_mr_create(). When it creates a new MR, it tries to register adjacent
memsegs as much as possible which are virtually contiguous around the
address. This must take two locks - memory_hotplug_lock and
priv->mr.rwlock. Due to memory_hotplug_lock, there can't be any
allocation/free of memory inside.
In the free callback of the memory hotplug event, freed space is searched
from the MR list and corresponding bits are cleared from the bitmap of MRs.
This can fragment a MR and the MR will have multiple search entries in the
caches. Once there's a change by the event, the global cache must be
rebuilt and all the per-queue caches will be flushed as well. If memory is
frequently freed in run-time, that may cause jitter on dataplane processing
in the worst case by incurring MR cache flush and rebuild. But, it would be
the least probable scenario.
To guarantee the most optimal performance, it is highly recommended to use
an EAL option - '--socket-mem'. Then, the reserved memory will be pinned
and won't be freed dynamically. And it is also recommended to configure
per-lcore cache of Mempool. Even though there're many MRs for a device or
MRs are highly fragmented, the cache of Mempool will be much helpful to
reduce misses on per-queue caches anyway.
'--legacy-mem' is also supported.
Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
Linux 4.15 and 4.16 may report inner RSS as a supported capability of the
device, however it can't be used due to missing code in the kernel.
This triggers an error when creating the default hash QP and prevents this
PMD from starting up without a prior call to rte_flow_isolate().
Fixes: 55e8991e31 ("net/mlx4: restore inner VXLAN RSS support")
Cc: stable@dpdk.org
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Using special types value -1 with mlx4_conv_rss_types() is supposed to
return a supported set of Verbs RSS hash fields, that is, priv->hw_rss_sup
unmodified.
Due to the way this function is written and because it is also used to
initially populate priv->hw_rss_sup however, this special value works
properly only once and fails with ENOTSUP errors afterward.
This problem can be seen when re-creating default flows (e.g. by entering
and leaving isolated mode).
Fixes: 024e87bef4 ("net/mlx4: restore UDP RSS by probing capabilities")
Cc: stable@dpdk.org
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
When creation of a flow rule fails during dev_start(), the usage count of
the common RSS context is not decremented, which triggers an assertion
failure in debug mode during dev_close().
This is addressed by tracking the initialization status of the common RSS
context in order to add missing cleanup code.
A similar issue exists in mlx4_rxq_attach(), where usage count is
incremented on a Rx queue but not released in case of error. This may lead
to the above issue since RSS contexts created by flow rules attach
themselves to Rx queues, incrementing their usage count.
Fixes: 5697a41421 ("net/mlx4: relax Rx queue configuration order")
Cc: stable@dpdk.org
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Since its inception, the rte_flow RSS action has been relying in part on
external struct rte_eth_rss_conf for compatibility with the legacy RSS API.
This structure lacks parameters such as the hash algorithm to use, and more
recently, a method to tell which layer RSS should be performed on [1].
Given struct rte_eth_rss_conf will never be flexible enough to represent a
complete RSS configuration (e.g. RETA table), this patch supersedes it by
extending the rte_flow RSS action directly.
A subsequent patch will add a field to use a non-default RSS hash
algorithm. To that end, a field named "types" replaces the field formerly
known as "rss_hf" and standing for "RSS hash functions" as it was
confusing. Actual RSS hash function types are defined by enum
rte_eth_hash_function.
This patch updates all PMDs and example applications accordingly.
It breaks ABI compatibility for the following public functions:
- rte_flow_copy()
- rte_flow_create()
- rte_flow_query()
- rte_flow_validate()
[1] commit 676b605182 ("doc: announce ethdev API change for RSS
configuration")
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Andrew Rybchenko <arybchenko@solarflare.com>
Previous to this commit mlx4 CRC stripping was executed by default and
there was no verbs API to disable it.
Signed-off-by: Ophir Munk <ophirmu@mellanox.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Aligning Mellanox SPDX copyrights to a single format.
In addition replace to SPDX licence files which were missed.
Signed-off-by: Shahaf Shuler <shahafs@mellanox.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Glue object files are looked up in RTE_EAL_PMD_PATH by default when set and
should be installed in this directory.
During startup, EAL attempts to load them automatically like other plug-ins
found there. While normally harmless, dlopen() fails when rdma-core is not
installed, EAL interprets this as a fatal error and terminates the
application.
This patch requests glue objects to be installed in a different directory
to prevent their automatic loading by EAL since they are PMD helpers, not
actual DPDK plug-ins.
Fixes: f6242d0655 ("net/mlx: make rdma-core glue path configurable")
Cc: stable@dpdk.org
Reported-by: Timothy Redaelli <tredaelli@redhat.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Tested-by: Timothy Redaelli <tredaelli@redhat.com>
Since rdma-core glue libraries are intrinsically tied to their respective
PMDs and used as internal plug-ins, their presence in the default search
path among other system libraries for the dynamic linker is not necessarily
desired.
This commit enables their installation and subsequent look-up at run time
in RTE_EAL_PMD_PATH if configured to a nonempty string. This path can also
be overridden by environment variables MLX[45]_GLUE_PATH.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
When built as separate objects, these libraries do not have unique names.
Since they do not maintain a stable ABI, loading an incompatible library
may result in a crash (e.g. in case multiple versions are installed).
This patch addresses the above by versioning glue libraries, both on the
file system (version suffix) and by comparing a dedicated version field
member in glue structures.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
This code should catch mistakes early if a glue structure member is added
without a corresponding implementation in the library.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
When mlx4 is not compiled directly as an independent shared object (e.g.
CONFIG_RTE_BUILD_SHARED_LIB not enabled for performance reasons), DPDK
applications inherit its dependencies on libibverbs and libmlx4 through
rte.app.mk.
This is an issue both when DPDK is delivered as a binary package (Linux
distributions) and for end users because rdma-core then propagates as a
mandatory dependency for everything.
Application writers relying on binary DPDK packages are not necessarily
aware of this fact and may end up delivering packages with broken
dependencies.
This patch therefore introduces an intermediate internal plug-in
hard-linked with rdma-core (to preserve symbol versioning) loaded by the
PMD through dlopen(), so that a missing rdma-core does not cause unresolved
symbols, allowing applications to start normally.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
This lays the groundwork for externalizing rdma-core as an optional
run-time dependency instead of a mandatory one.
No functional change.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
In failsafe device start can be called for ports/devices that
had been plugged out.
The mlx4 PMD detects device removal by listening to the device RMV
events, when the mlx4 port is being stopped, the PMD no longer
listens to these events causing the PMD to stop detecting device
removals.
This patch fixes this issue by moving installation of the interrupt
handler to device configuration, and toggle only the Rx-queue
interrupts on start/stop.
Fixes: a6e8b01c3c ("net/mlx4: compact interrupt functions")
Cc: stable@dpdk.org
Signed-off-by: Moti Haimovsky <motih@mellanox.com>
Acked-by: Shahaf Shuler <shahafs@mellanox.com>
The number of mlx4 present ports is calculated as follows:
conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
That is - all ones sequence (due to -1 subtraction)
When retrieving the number of ports, 1 must be added in order to obtain
the correct number of ports to the power of 2, as follows:
uint32_t ports = rte_log2_u32(conf->ports.present + 1);
If 1 was not added, in the case of one port, the number of ports would
be falsely calculated as 0.
Fixes: 8264279967 ("net/mlx4: check max number of ports dynamically")
Cc: stable@dpdk.org
Signed-off-by: Ophir Munk <ophirmu@mellanox.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Create a rte_ethdev_driver.h file and move PMD specific APIs here.
Drivers updated to include this new header file.
There is no update in header content and since ethdev.h included by
ethdev_driver.h, nothing changed from driver point of view, only
logically grouping of APIs. From applications point of view they can't
access to driver specific APIs anymore and they shouldn't.
More PMD specific data structures still remain in ethdev.h because of
inline functions in header use them. Those will be handled separately.
Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
Acked-by: Shreyansh Jain <shreyansh.jain@nxp.com>
Acked-by: Andrew Rybchenko <arybchenko@solarflare.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Max number of Tx scatter-gather entries is a property of the device
and is queried at init. This value was not changed in a while and
most probably will not be changed in the future, Therefore and
in order to enhance Tx performance, the Tx max-sge value is hardcoded
in mlx4 PRM code.
This patch adds a verification that the above assumption still holds
and that the hardcoded value is still supported by the mlx4 hardware.
Signed-off-by: Moti Haimovsky <motih@mellanox.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
This workaround was needed to properly handle device removal with old
Mellanox OFED releases that are not supported by this PMD anymore.
Starting from rdma-core v16 this removal issue shouldn't happen when
setting MLX4_DEVICE_FATAL_CLEANUP environment variable to 1.
Set the aforementioned variable to 1.
Reverts: 5f4677c6ad ("net/mlx4: workaround verbs error after plug-out")
Cc: stable@dpdk.org
Signed-off-by: Matan Azrad <matan@mellanox.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Inner VXLAN RSS was supported and performed by default prior to the entire
mlx4 refactoring that occurred in DPDK 17.11, however so far the new Verbs
RSS API did not provide means to enable it. This will be addressed in
Linux 4.15 and in RDMA core.
Thanks to RSS capabilities, the PMD can now probe for its support and
enable it again by default.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Until now, UDP RSS support could not be relied on due to a problem in the
Linux kernel implementation and mlx4 RSS capabilities were not reported at
all, hence the PMD had to make assumptions.
Since both issues will be addressed simultaneously in Linux 4.15 (related
patches already upstream) and likely backported afterward, UDP RSS support
can be enabled by probing RSS capabilities.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
This patch enhances the Rx packet type offload to also report the L4
protocol information in the hw ptype filled by the PMD for each received
packet.
Signed-off-by: Moti Haimovsky <motih@mellanox.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
This flag is not necessary at the ether layer anymore.
Buses are able to advertise their hotplug support. The ether layer can
rely upon this capability instead of a special flag.
Signed-off-by: Gaetan Rivet <gaetan.rivet@6wind.com>
Acked-by: Andrew Rybchenko <arybchenko@solarflare.com>
Acked-by: John McNamara <john.mcnamara@intel.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Various hardware limitations apply to RSS indirection tables, one of
them being they must be an exact 1:1 mapping of the configured Rx queue
indices.
While this restriction is enforced when creating RSS flow rules, it is
not the case when Rx queues themselves are created; underlying WQ
numbers are assigned in turn, not according to queue index.
Applications such as l3fwd-power that create Rx queues from highest to
lowest index (or any other non-sequential order) thus fail to get a
working RSS context.
This commit postpones WQ initialization to dev_start(), once all Rx
queues are configured in order to address this issue.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
This patch adds hardware offloading support for IPv4, UDP and TCP checksum
calculation, including inner/outer checksums on supported tunnel types.
Signed-off-by: Moti Haimovsky <motih@mellanox.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
This patch dissociates single-queue indirection tables and hash QP objects
from Rx queue structures to relinquish their control to users through the
RSS flow rule action, while simultaneously allowing multiple queues to be
associated with RSS contexts.
Flow rules share identical RSS contexts (hashed fields, hash key, target
queues) to save on memory and other resources. The trade-off is some added
complexity due to reference counters management on RSS contexts.
The QUEUE action is re-implemented on top of an automatically-generated
single-queue RSS context.
The following hardware limitations apply to RSS contexts:
- The number of queues in a group must be a power of two.
- Queue indices must be consecutive, for instance the [0 1 2 3] set is
allowed, however [3 2 1 0], [0 2 1 3] and [0 0 1 1 2 3 3 3] are not.
- The first queue of a group must be aligned to a multiple of the context
size, e.g. if queues [0 1 2 3 4] are defined globally, allowed group
combinations are [0 1] and [2 3]; groups [1 2] and [3 4] are not
supported.
- RSS hash key, while configurable per context, must be exactly 40 bytes
long.
- The only supported hash algorithm is Toeplitz.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Although their "removed" version acts as a safety against unexpected bursts
while queues are being modified by the control path, these callbacks are
set per device instead of per queue. It makes sense to update them during
start/stop/close cycles instead of queue setup.
As a side effect, this commit addresses a bug left over from a prior
commit: bringing the link down causes the "removed" Tx callback to be used,
however the normal callback is not restored when bringing it back up,
preventing the application from sending traffic at all.
Updating callbacks for a link change is not necessary as bringing the
netdevice down is normally enough to prevent traffic from flowing in.
Fixes: 3f75a02719 ("net/mlx4: drop scatter/gather support")
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Implement promiscuous and all multicast through internal flow rules
automatically generated according to the configured mode.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
This commit brings back VLAN filter configuration support without any
artificial limitation on the number of simultaneous VLANs that can be
configured (previously 127).
Also thanks to the fact it does not rely on fixed per-queue arrays for
potential Verbs flow handle storage anymore, this version wastes a lot less
memory (previously 128 * 127 * pointer size, i.e. 130 kiB per Rx queue,
only one of which actually had any use for this room: the RSS parent
queue).
The number of internal flow rules generated still depends on the number of
configured MAC addresses times that of configured VLAN filters though.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
This commit brings back support for configuring up to 128 MAC addresses on
a port through internal flow rules automatically generated on demand.
Unlike its previous incarnation, the necessary extra flow rule for
broadcast traffic does not consume an entry from the MAC array anymore.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Since flow rules synchronization function mlx4_flow_sync() takes into
account the state of the device (whether it is started), trigger functions
mlx4_flow_start() and mlx4_flow_stop() are redundant. Standardize on
mlx4_flow_sync().
Use this opportunity to enhance this function with better error reporting
as the inability to start the device due to a problem with a flow rule
otherwise results in a nondescript error code.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
When not in isolated mode, a flow rule is automatically configured by the
PMD to receive traffic addressed to the MAC address of the device. This
somewhat duplicates flow API functionality.
Remove legacy support for internal flow rules to instead handle them
through the flow API implementation.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Allocation and management of Tx/Rx queue arrays is done by wrappers at the
ethdev level. The resulting information is copied to the private structure
while configuring the device, where it is managed separately by the PMD.
This is redundant and consumes space in the private structure.
Relying more on ethdev also means there is no need to protect the PMD
against burst function calls while closing the device anymore.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Add missing includes and sort them, then update/remove comments around them
for consistency.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
While internal static functions do not cause link time conflicts, this
differentiates them from their mlx5 PMD counterparts while debugging.
No impact on functionality.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Only the common filter control operation callback needs to be exposed.
No impact on functionality.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Private functions are now prefixed with "mlx4_" to prevent them from
conflicting with their mlx5 PMD counterparts at link time.
No impact on functionality.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Private functions are now prefixed with "mlx4_" to prevent them from
conflicting with their mlx5 PMD counterparts at link time.
No impact on functionality.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Private functions are now prefixed with "mlx4_" to prevent them from
conflicting with their mlx5 PMD counterparts at link time.
No impact on functionality.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
This commit groups all data plane functions (Rx/Tx) into a separate file
and adjusts header files accordingly.
Private functions are now prefixed with "mlx4_" to prevent them from
conflicting with their mlx5 PMD counterparts at link time.
No impact on functionality.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Except for a minor documentation update on internal structure definitions
to make them more Doxygen-friendly, there is no impact on functionality.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Private functions are now prefixed with "mlx4_" to prevent them from
conflicting with their mlx5 PMD counterparts at link time.
No impact on functionality.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
