IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
As stated in [1], dma_set_mask() with a 64-bit mask never fails if
dev->dma_mask is non-NULL.
So, if it fails, the 32 bits case will also fail for the same reason.
Moreover, dma_set_mask_and_coherent() returns 0 or -EIO, so the return
code of the function can be used directly.
Finally, inline bnx2x_set_coherency_mask() because it is now only a wrapper
for a single dma_set_mask_and_coherent() call.
Simplify code and remove some dead code accordingly.
[1]: https://lkml.org/lkml/2021/6/7/398
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Link: https://lore.kernel.org/r/29608a525876afddceabf8f11b2ba606da8748fc.1641730747.git.christophe.jaillet@wanadoo.fr
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
As stated in [1], dma_set_mask() with a 64-bit mask never fails if
dev->dma_mask is non-NULL.
So, if it fails, the 32 bits case will also fail for the same reason.
Moreover, dma_set_mask_and_coherent() returns 0 or -EIO, so the return
code of the function can be used directly. There is no need to 'rc = -EIO'
explicitly.
Simplify code and remove some dead code accordingly.
[1]: https://lkml.org/lkml/2021/6/7/398
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Link: https://lore.kernel.org/r/b9aa46e7e5a5aa61f56aac5ea439930f41ad9946.1641726804.git.christophe.jaillet@wanadoo.fr
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
As stated in [1], dma_set_mask() with a 64-bit mask never fails if
dev->dma_mask is non-NULL.
So, if it fails, the 32 bits case will also fail for the same reason.
So if dma_set_mask_and_coherent() succeeds, 'netdev->features' will have
NETIF_F_HIGHDMA in all cases. Move the assignment of this feature in
be_netdev_init() instead be_probe() which is a much logical place.
Simplify code and remove some dead code accordingly.
[1]: https://lkml.org/lkml/2021/6/7/398
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Link: https://lore.kernel.org/r/637696d7141faa68c29fc34b70f9aa67d5e605f0.1641718999.git.christophe.jaillet@wanadoo.fr
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
As stated in [1], dma_set_mask() with a 64-bit mask will never fail if
dev->dma_mask is non-NULL.
So, if it fails, the 32 bits case will also fail for the same reason.
So, if dma_set_mask_and_coherent() succeeds, 'using_dac' is known to be
'true'. This variable can be removed.
Simplify code and remove some dead code accordingly.
[1]: https://lkml.org/lkml/2021/6/7/398
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Link: https://lore.kernel.org/r/1d5a7b3f4fa735f1233c3eb3fa07e71df95fad75.1641658516.git.christophe.jaillet@wanadoo.fr
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
As stated in [1], dma_set_mask() with a 64-bit mask will never fail if
dev->dma_mask is non-NULL.
So, if it fails, the 32 bits case will also fail for the same reason.
If dma_set_mask_and_coherent() succeeds, 'ap->pci_using_dac' is known to be
1. So 'pci_using_dac' can be removed from the 'struct ace_private'.
Simplify code and remove some dead code accordingly.
[1]: https://lkml.org/lkml/2021/6/7/398
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Link: https://lore.kernel.org/r/1a414c05c27b21c661aef61dffe1adcd1578b1f5.1641651917.git.christophe.jaillet@wanadoo.fr
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
As stated in [1], dma_set_mask() with a 64-bit mask will never fail if
dev->dma_mask is non-NULL.
So, if it fails, the 32 bits case will also fail for the same reason.
So qlcnic_set_dma_mask(), (in qlcnic_main.c) can be simplified a lot and
inlined directly in its only caller.
If dma_set_mask_and_coherent() succeeds, 'pci_using_dac' is known to be 1.
So it can be removed from all the calling chain.
qlcnic_setup_netdev() can finally be simplified as-well.
[1]: https://lkml.org/lkml/2021/6/7/398
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Link: https://lore.kernel.org/r/4996ab0337d62ec6a54b2edf234cd5ced4b4d7ad.1641649611.git.christophe.jaillet@wanadoo.fr
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Kees reports quoted commit introduced the following warning on arm64:
drivers/net/ethernet/allwinner/sun4i-emac.c:922:60: error: format '%x' expects argument of type 'unsigned int', but argument 3 has type 'resource_size_t' {aka 'long long unsigned int'} [-Werror=format=]
922 | netdev_info(ndev, "get io resource from device: 0x%x, size = %u\n",
| ~^
| | | unsigned int
| %llx
923 | regs->start, resource_size(regs));
| ~~~~~~~~~~~
| |
| resource_size_t {aka long long unsigned int}
.. and another one like that for resource_size().
Switch to %pa and a cast.
Reported-by: Kees Cook <keescook@chromium.org>
Fixes: 47869e82c8b8 ("sun4i-emac.c: add dma support")
Link: https://lore.kernel.org/r/20220108034438.2227343-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Build bot reports:
drivers/net/ethernet/mellanox/mlx5/core/en_stats.c: In function 'fec_set_block_stats':
drivers/net/ethernet/mellanox/mlx5/core/en_stats.c:1235:48: error: 'outl' undeclared (first use in this function); did you mean 'out'?
1235 | if (mlx5_core_access_reg(mdev, in, sz, outl, sz, MLX5_REG_PPCNT, 0, 0))
| ^~~~
| out
Reported-by: kernel test robot <lkp@intel.com>
Suggested-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/r/20220109213321.2292830-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
While it has always been possible to infer that an HWRM command was
abandoned due to an unhealthy firmware status by the shortened timeout
reported, this change improves the log messaging to account for this
case explicitly. In the interests of further clarity, the firmware
status is now also reported in these new messages.
v2: Remove inline keyword for hwrm_wait_must_abort() in .c file.
Reviewed-by: Andy Gospodarek <gospo@broadcom.com>
Signed-off-by: Edwin Peer <edwin.peer@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Some older devices cannot accommodate the 40 seconds timeout
cap for long running commands (such as NVRAM commands) due to
hardware limitations. Allow these devices to request more time for
these long running commands, but print a warning, since the longer
timeout may cause the hung task watchdog to trigger. In the case of a
firmware update operation, this is preferable to failing outright.
v2: Use bp->hwrm_cmd_max_timeout directly without the constants.
Fixes: 881d8353b05e ("bnxt_en: Add an upper bound for all firmware command timeouts.")
Signed-off-by: Edwin Peer <edwin.peer@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The current driver design relies on the PF netdev being open in order
to intercept the following HWRM commands from a VF:
- HWRM_FUNC_VF_CFG
- HWRM_CFA_L2_FILTER_ALLOC
- HWRM_PORT_PHY_QCFG (only if FW_CAP_LINK_ADMIN is not supported)
If the PF is closed, then VFs are subjected to rather inscrutable error
messages in response to any configuration requests involving the above
command types. Recent firmware distinguishes this problem case from
other errors by returning HWRM_ERR_CODE_PF_UNAVAILABLE. In most cases,
the appropriate course of action is still to fail, but this can now be
accomplished with the aid of more user informative log messages. For L2
filter allocations that are already asynchronous, an automatic retry
seems more appropriate.
v2: Delete extra newline.
Signed-off-by: Edwin Peer <edwin.peer@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add logging of firmware messages. These can be useful for diagnosing
issues in the field, but due to their verbosity are only appropriate
at a debug message level.
Signed-off-by: Edwin Peer <edwin.peer@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Create an inline function for resetting the driver
to reduce code duplication.
Signed-off-by: Nati Koler <nkoler@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Changed bad_csum to csum_bad to align with csum_unchecked & csum_good
Signed-off-by: Nati Koler <nkoler@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add qid and req_id to error prints when ENA_REGS_RESET_INV_TX_REQ_ID
reset occurs.
Switch from %hu to %u, since u16 should be printed with %u, as
explained in [1].
[1] - https://www.kernel.org/doc/html/latest/core-api/printk-formats.html
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This struct was used to pass data from callee function to its caller.
Its usage can be avoided.
Removing it results in less code without any damage to code readability.
Also it allows to consolidate ring size calculation into a single
function (ena_calc_io_queue_size()).
Signed-off-by: Shay Agroskin <shayagr@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The print that indicates that device reset has finished is
currently called from ena_restore_device().
Move it to ena_fw_reset_device() as it is the more natural
location for it.
Signed-off-by: Shay Agroskin <shayagr@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The ena_com_indirect_table_fill_entry() function only returns -EINVAL
or 0, no need to check for -EOPNOTSUPP.
Signed-off-by: Shay Agroskin <shayagr@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Use the capabilities field to query the device for ENI stats
support.
This replaces the previous method that tried to get the ENI stats
during ena_probe() and used the success or failure as an indication
for support by the device.
Remove eni_stats_supported field from struct ena_adapter. This field
was used for the previous method of queriying for ENI stats support.
Change the severity level of the print in case of
ena_com_get_eni_stats() failure from info to error.
With the previous method of querying form ENI stats support, failure
to get ENI stats was normal for devices that don't support it.
With the use of the capabilities field such a failure is unexpected,
as it is called only if the device reported that it supports ENI
stats.
Signed-off-by: Shay Agroskin <shayagr@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This bitmask field indicates what capabilities are supported by the
device.
The capabilities field differs from the 'supported_features' field which
indicates what sub-commands for the set/get feature commands are
supported. The sub-commands are specified in the 'feature_id' field of
the 'ena_admin_set_feat_cmd' struct in the following way:
struct ena_admin_set_feat_cmd cmd;
cmd.aq_common_descriptor.opcode = ENA_ADMIN_SET_FEATURE;
cmd.feat_common.feature_
The 'capabilities' field, on the other hand, specifies different
capabilities of the device. For example, whether the device supports
querying of ENI stats.
Also add an enumerator which contains all the capabilities. The
first added capability macro is for ENI stats feature.
Capabilities are queried along with the other device attributes (in
ena_com_get_dev_attr_feat()) during device initialization and are stored
in the ena_com_dev struct. They can be later queried using the
ena_com_get_cap() helper function.
Signed-off-by: Shay Agroskin <shayagr@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
ena_calc_io_queue_size() always returns 0, therefore make it a
void function and update the calling function to stop checking
the return value.
Signed-off-by: Shay Agroskin <shayagr@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add support for flushing the MAC table on a given port in the ocelot
switch library, and use this functionality in the felix DSA driver.
This operation is needed when a port leaves a bridge to become
standalone, and when the learning is disabled, and when the STP state
changes to a state where no FDB entry should be present.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Link: https://lore.kernel.org/r/20220107144229.244584-1-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Assuming the test setup described here:
https://patchwork.kernel.org/project/netdevbpf/cover/20210205130240.4072854-1-vladimir.oltean@nxp.com/
(swp1 and swp2 are in bond0, and bond0 is in a bridge with swp0)
it can be seen that when swp1 goes down (on either board A or B), then
traffic that should go through that port isn't forwarded anywhere.
A dump of the PGID table shows the following:
PGID_DST[0] = ports 0
PGID_DST[1] = ports 1
PGID_DST[2] = ports 2
PGID_DST[3] = ports 3
PGID_DST[4] = ports 4
PGID_DST[5] = ports 5
PGID_DST[6] = no ports
PGID_AGGR[0] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[1] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[2] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[3] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[4] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[5] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[6] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[7] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[8] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[9] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[10] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[11] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[12] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[13] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[14] = ports 0, 1, 2, 3, 4, 5
PGID_AGGR[15] = ports 0, 1, 2, 3, 4, 5
PGID_SRC[0] = ports 1, 2
PGID_SRC[1] = ports 0
PGID_SRC[2] = ports 0
PGID_SRC[3] = no ports
PGID_SRC[4] = no ports
PGID_SRC[5] = no ports
PGID_SRC[6] = ports 0, 1, 2, 3, 4, 5
Whereas a "good" PGID configuration for that setup should have looked
like this:
PGID_DST[0] = ports 0
PGID_DST[1] = ports 1, 2
PGID_DST[2] = ports 1, 2
PGID_DST[3] = ports 3
PGID_DST[4] = ports 4
PGID_DST[5] = ports 5
PGID_DST[6] = no ports
PGID_AGGR[0] = ports 0, 2, 3, 4, 5
PGID_AGGR[1] = ports 0, 2, 3, 4, 5
PGID_AGGR[2] = ports 0, 2, 3, 4, 5
PGID_AGGR[3] = ports 0, 2, 3, 4, 5
PGID_AGGR[4] = ports 0, 2, 3, 4, 5
PGID_AGGR[5] = ports 0, 2, 3, 4, 5
PGID_AGGR[6] = ports 0, 2, 3, 4, 5
PGID_AGGR[7] = ports 0, 2, 3, 4, 5
PGID_AGGR[8] = ports 0, 2, 3, 4, 5
PGID_AGGR[9] = ports 0, 2, 3, 4, 5
PGID_AGGR[10] = ports 0, 2, 3, 4, 5
PGID_AGGR[11] = ports 0, 2, 3, 4, 5
PGID_AGGR[12] = ports 0, 2, 3, 4, 5
PGID_AGGR[13] = ports 0, 2, 3, 4, 5
PGID_AGGR[14] = ports 0, 2, 3, 4, 5
PGID_AGGR[15] = ports 0, 2, 3, 4, 5
PGID_SRC[0] = ports 1, 2
PGID_SRC[1] = ports 0
PGID_SRC[2] = ports 0
PGID_SRC[3] = no ports
PGID_SRC[4] = no ports
PGID_SRC[5] = no ports
PGID_SRC[6] = ports 0, 1, 2, 3, 4, 5
In other words, in the "bad" configuration, the attempt is to remove the
inactive swp1 from the destination ports via PGID_DST. But when a MAC
table entry is learned, it is learned towards PGID_DST 1, because that
is the logical port id of the LAG itself (it is equal to the lowest
numbered member port). So when swp1 becomes inactive, if we set
PGID_DST[1] to contain just swp1 and not swp2, the packet will not have
any chance to reach the destination via swp2.
The "correct" way to remove swp1 as a destination is via PGID_AGGR
(remove swp1 from the aggregation port groups for all aggregation
codes). This means that PGID_DST[1] and PGID_DST[2] must still contain
both swp1 and swp2. This makes the MAC table still treat packets
destined towards the single-port LAG as "multicast", and the inactive
ports are removed via the aggregation code tables.
The change presented here is a design one: the ocelot_get_bond_mask()
function used to take an "only_active_ports" argument. We don't need
that. The only call site that specifies only_active_ports=true,
ocelot_set_aggr_pgids(), must retrieve the entire bonding mask, because
it must program that into PGID_DST. Additionally, it must also clear the
inactive ports from the bond mask here, which it can't do if bond_mask
just contains the active ports:
ac = ocelot_read_rix(ocelot, ANA_PGID_PGID, i);
ac &= ~bond_mask; <---- here
/* Don't do division by zero if there was no active
* port. Just make all aggregation codes zero.
*/
if (num_active_ports)
ac |= BIT(aggr_idx[i % num_active_ports]);
ocelot_write_rix(ocelot, ac, ANA_PGID_PGID, i);
So it becomes the responsibility of ocelot_set_aggr_pgids() to take
ocelot_port->lag_tx_active into consideration when populating the
aggr_idx array.
Fixes: 23ca3b727ee6 ("net: mscc: ocelot: rebalance LAGs on link up/down events")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://lore.kernel.org/r/20220107164332.402133-1-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The variable `ret_code' used for returning is never changed in function
`iavf_shutdown_adminq'. So that it can be removed and just return its
initial value 0 at the end of `iavf_shutdown_adminq' function.
Signed-off-by: Jason Wang <wangborong@cdjrlc.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
The code that uses variables pe_cntx_size and pe_filt_size
has been removed, so they should be removed as well.
Eliminate the following clang warnings:
drivers/net/ethernet/intel/i40e/i40e_common.c:4139:20:
warning: variable 'pe_filt_size' set but not used.
drivers/net/ethernet/intel/i40e/i40e_common.c:4139:6:
warning: variable 'pe_cntx_size' set but not used.
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Signed-off-by: Yang Li <yang.lee@linux.alibaba.com>
Tested-by: Gurucharan G <gurucharanx.g@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Remove non-inclusive language from the driver.
Signed-off-by: Aleksandr Loktionov <aleksandr.loktionov@intel.com>
Signed-off-by: Mateusz Palczewski <mateusz.palczewski@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
The i40e_asq_send_command will now use a non blocking usleep_range if
possible (non-atomic context), instead of busy-waiting udelay. The
usleep_range function uses hrtimers to provide better performance and
removes the negative impact of busy-waiting in time-critical
environments.
1. Rename i40e_asq_send_command to i40e_asq_send_command_atomic
and add 5th parameter to inform if called from an atomic context.
Call inside usleep_range (if non-atomic) or udelay (if atomic).
2. Change i40e_asq_send_command to invoke
i40e_asq_send_command_atomic(..., false).
3. Change two functions:
- i40e_aq_set_vsi_uc_promisc_on_vlan
- i40e_aq_set_vsi_mc_promisc_on_vlan
to explicitly use i40e_asq_send_command_atomic(..., true)
instead of i40e_asq_send_command, as they use spinlocks and do some
work in an atomic context.
All other calls to i40e_asq_send_command remain unchanged.
Signed-off-by: Dawid Lukwinski <dawid.lukwinski@intel.com>
Signed-off-by: Jedrzej Jagielski <jedrzej.jagielski@intel.com>
Tested-by: Tony Brelinski <tony.brelinski@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Trusted VF can use up every resource available, leaving nothing
to other trusted VFs.
Introduce define, which calculates MacVlan resources available based
on maximum available MacVlan resources, bare minimum for each VF and
number of currently allocated VFs.
Signed-off-by: Przemyslaw Patynowski <przemyslawx.patynowski@intel.com>
Signed-off-by: Karen Sornek <karen.sornek@intel.com>
Tested-by: Tony Brelinski <tony.brelinski@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
When a VF is removed respective PTP resources are not
being freed currently. This patch fixes it.
Fixes: 43510ef4ddad ("octeontx2-nicvf: Add PTP hardware clock support to NIX VF")
Signed-off-by: Rakesh Babu Saladi <rsaladi2@marvell.com>
Signed-off-by: Subbaraya Sundeep <sbhatta@marvell.com>
Signed-off-by: Sunil Goutham <sgoutham@marvell.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Before using the ptp pci device by AF driver increment
the reference count of it.
Fixes: a8b90c9d26d6 ("octeontx2-af: Add PTP device id for CN10K and 95O silcons")
Signed-off-by: Subbaraya Sundeep <sbhatta@marvell.com>
Signed-off-by: Sunil Goutham <sgoutham@marvell.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
1) Expose FEC per lane block counters via ethtool
2) Trivial fixes/updates/cleanup to mlx5e netdev driver
3) Fix htmldoc build warning
4) Spread mlx5 SFs (sub-functions) to all available CPU cores: Commits 1..5
Shay Drory Says:
================
Before this patchset, mlx5 subfunction shared the same IRQs (MSI-X) with
their peers subfunctions, causing them to use same CPU cores.
In large scale, this is very undesirable, SFs use small number of cpu
cores and all of them will be packed on the same CPU cores, not
utilizing all CPU cores in the system.
In this patchset we want to achieve two things.
a) Spread IRQs used by SFs to all cpu cores
b) Pack less SFs in the same IRQ, will result in multiple IRQs per core.
In this patchset, we spread SFs over all online cpus available to mlx5
irqs in Round-Robin manner. e.g.: Whenever a SF is created, pick the next
CPU core with least number of SF IRQs bound to it, SFs will share IRQs on
the same core until a certain limit, when such limit is reached, we
request a new IRQ and add it to that CPU core IRQ pool, when out of IRQs,
pick any IRQ with least number of SF users.
This enhancement is done in order to achieve a better distribution of
the SFs over all the available CPUs, which reduces application latency,
as shown bellow.
Machine details:
Intel(R) Xeon(R) CPU E5-2697 v3 @ 2.60GHz with 56 cores.
PCI Express 3 with BW of 126 Gb/s.
ConnectX-5 Ex; EDR IB (100Gb/s) and 100GbE; dual-port QSFP28; PCIe4.0
x16.
Base line test description:
Single SF on the system. One instance of netperf is running on-top the
SF.
Numbers: latency = 15.136 usec, CPU Util = 35%
Test description:
There are 250 SFs on the system. There are 3 instances of netperf
running, on-top three different SFs, in parallel.
Perf numbers:
# netperf SFs latency(usec) latency CPU utilization
affinity affinity (lower is better) increase %
1 cpu=0 cpu={0} ~23 (app 1-3) 35% 75%
2 cpu=0,2,4 cpu={0} app 1: 21.625 30% 68% (CPU 0)
app 2-3: 16.5 9% 15% (CPU 2,4)
3 cpu=0 cpu={0,2,4} app 1: ~16 7% 84% (CPU 0)
app 2-3: ~17.9 14% 22% (CPU 2,4)
4 cpu=0,2,4 cpu={0,2,4} 15.2 (app 1-3) 0% 33% (CPU 0,2,4)
- The first two entries (#1 and #2) show current state. e.g.: SFs are
using the same CPU. The last two entries (#3 and #4) shows the latency
reduction improvement of this patch. e.g.: SFs are on different CPUs.
- Whenever we use several CPUs, in case there is a different CPU
utilization, write the utilization of each CPU separately.
- Whenever the latency result of the netperf instances were different,
write the latency of each netperf instances separately.
Commands:
- for netperf CPU=0:
$ for i in {1..3}; do taskset -c 0 netperf -H 1${i}.1.1.1 -t TCP_RR -- \
-o RT_LATENCY -r8 & done
- for netperf CPU=0,2,4
$ for i in {1..3}; do taskset -c $(( ($i - 1) * 2 )) netperf -H \
1${i}.1.1.1 -t TCP_RR -- -o RT_LATENCY -r8 & done
================
-----BEGIN PGP SIGNATURE-----
iQEzBAABCAAdFiEEGhZs6bAKwk/OTgTpSD+KveBX+j4FAmHXh+AACgkQSD+KveBX
+j68fQgAghUX4TFS2JSwa7+XSCtzz7GIu2Xrz8aWTAnydRLlNXuFuuHYcNed6I0l
7DaVOZwHp1tp3tnx3WMGPUU6ujDPEgasaDDblvG2UXix5LPVEHDXY44ittQX8mpC
SC8Yj9mNo6DSfOMUZklFDMbw57XuLJ+HEGnwnrOEEyLX7ruDXGEViUmVBd4IoC3B
F2fJHBkdTJfHWTJRB4pWbZD1dw7WbKd0RyPla3OkoHugEUCKnbjii8cMwNM64Bbp
Pjz/SiShVy+NTotqPzRNjcx7y4tHOXCYt33zt1VlGtdUxs5eCA5jkjHFz0jb12Lu
rvfHaBaU+elMKTw5G/WMGJxZQx0kEQ==
=VBWY
-----END PGP SIGNATURE-----
Merge tag 'mlx5-updates-2022-01-06' of git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux
Saeed Mahameed says:
====================
mlx5-updates-2022-01-06
1) Expose FEC per lane block counters via ethtool
2) Trivial fixes/updates/cleanup to mlx5e netdev driver
3) Fix htmldoc build warning
4) Spread mlx5 SFs (sub-functions) to all available CPU cores: Commits 1..5
Shay Drory Says:
================
Before this patchset, mlx5 subfunction shared the same IRQs (MSI-X) with
their peers subfunctions, causing them to use same CPU cores.
In large scale, this is very undesirable, SFs use small number of cpu
cores and all of them will be packed on the same CPU cores, not
utilizing all CPU cores in the system.
In this patchset we want to achieve two things.
a) Spread IRQs used by SFs to all cpu cores
b) Pack less SFs in the same IRQ, will result in multiple IRQs per core.
In this patchset, we spread SFs over all online cpus available to mlx5
irqs in Round-Robin manner. e.g.: Whenever a SF is created, pick the next
CPU core with least number of SF IRQs bound to it, SFs will share IRQs on
the same core until a certain limit, when such limit is reached, we
request a new IRQ and add it to that CPU core IRQ pool, when out of IRQs,
pick any IRQ with least number of SF users.
This enhancement is done in order to achieve a better distribution of
the SFs over all the available CPUs, which reduces application latency,
as shown bellow.
Machine details:
Intel(R) Xeon(R) CPU E5-2697 v3 @ 2.60GHz with 56 cores.
PCI Express 3 with BW of 126 Gb/s.
ConnectX-5 Ex; EDR IB (100Gb/s) and 100GbE; dual-port QSFP28; PCIe4.0
x16.
Base line test description:
Single SF on the system. One instance of netperf is running on-top the
SF.
Numbers: latency = 15.136 usec, CPU Util = 35%
Test description:
There are 250 SFs on the system. There are 3 instances of netperf
running, on-top three different SFs, in parallel.
Perf numbers:
# netperf SFs latency(usec) latency CPU utilization
affinity affinity (lower is better) increase %
1 cpu=0 cpu={0} ~23 (app 1-3) 35% 75%
2 cpu=0,2,4 cpu={0} app 1: 21.625 30% 68% (CPU 0)
app 2-3: 16.5 9% 15% (CPU 2,4)
3 cpu=0 cpu={0,2,4} app 1: ~16 7% 84% (CPU 0)
app 2-3: ~17.9 14% 22% (CPU 2,4)
4 cpu=0,2,4 cpu={0,2,4} 15.2 (app 1-3) 0% 33% (CPU 0,2,4)
- The first two entries (#1 and #2) show current state. e.g.: SFs are
using the same CPU. The last two entries (#3 and #4) shows the latency
reduction improvement of this patch. e.g.: SFs are on different CPUs.
- Whenever we use several CPUs, in case there is a different CPU
utilization, write the utilization of each CPU separately.
- Whenever the latency result of the netperf instances were different,
write the latency of each netperf instances separately.
Commands:
- for netperf CPU=0:
$ for i in {1..3}; do taskset -c 0 netperf -H 1${i}.1.1.1 -t TCP_RR -- \
-o RT_LATENCY -r8 & done
- for netperf CPU=0,2,4
$ for i in {1..3}; do taskset -c $(( ($i - 1) * 2 )) netperf -H \
1${i}.1.1.1 -t TCP_RR -- -o RT_LATENCY -r8 & done
================
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Tony Nguyen says:
====================
100GbE Intel Wired LAN Driver Updates 2022-01-06
Victor adds restoring of advanced rules after reset.
Wojciech improves usage of switchdev control VSI by utilizing the
device's advanced rules for forwarding.
Christophe Jaillet removes some unneeded calls to zero bitmaps, changes
some bitmap operations that don't need to be atomic, and converts a
kfree() to a more appropriate bitmap_free().
* '100GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/next-queue:
ice: Use bitmap_free() to free bitmap
ice: Optimize a few bitmap operations
ice: Slightly simply ice_find_free_recp_res_idx
ice: improve switchdev's slow-path
ice: replay advanced rules after reset
====================
Link: https://lore.kernel.org/r/20220106183013.3777622-1-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Extend existing driver for Spectrum, Spectrum-2 and Spectrum-3 ASICs
to support Spectrum-4 ASIC as well.
Currently there is no released firmware version for Spectrum-4, so the
driver is not enforcing a minimum version.
Signed-off-by: Amit Cohen <amcohen@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Spectrum-4 will calculate hash function for bloom filter differently
from the existing ASICs.
First, two hash functions will be used to calculate 16 bits result.
The final result will be combination of the two results - 6 bits which
are result of CRC-6 will be used as MSB and 10 bits which are result of
CRC-10 will be used as LSB.
Second, while in Spectrum{2,3}, there is a padding in each chunk, so the
chunks use a sequence of whole bytes, in Spectrum-4 there is no padding,
so each chunk use 20 bytes minus 2 bits, so it is necessary to align the
chunks to be without holes.
Add dedicated 'mlxsw_sp_acl_bf_ops' for Spectrum-4 and add the required
tables for CRC calculations.
All the details are documented as part of the code for future use.
Signed-off-by: Amit Cohen <amcohen@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Spectrum-4 will calculate hash function for bloom filter differently from
the existing ASICs.
There are two changes:
1. Instead of using one hash function to calculate 16 bits output (CRC-16),
two functions will be used.
2. The chunks will be built differently, without padding.
As preparation for support of Spectrum-4 bloom filter, add 'ops'
structure to allow handling different calculation for different ASICs.
Signed-off-by: Amit Cohen <amcohen@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Spectrum-4 will calculate hash function for bloom filter differently from
the existing ASICs.
There are two changes:
1. Instead of using one hash function to calculate 16 bits output (CRC-16),
two functions will be used.
2. The chunks will be built differently, without padding.
As preparation for support of Spectrum-4 bloom filter, rename CRC table
to include "sp2" prefix and "crc16", as next patch will add two additional
tables. In addition, rename all the dedicated functions and defines for
Spectrum-{2,3} to include "sp2" prefix.
Signed-off-by: Amit Cohen <amcohen@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Spectrum-4 will calculate hash function for bloom filter differently from
the existing ASICs.
One of the changes is related to the way that the chunks will be build -
without padding.
As preparation for support of Spectrum-4 bloom filter, make
mlxsw_sp_acl_bf_key_encode() more flexible, so it will be able to use it
for Spectrum-4 as well.
Signed-off-by: Amit Cohen <amcohen@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, mlxsw_sp_acl_bf_rule_count_index_get() is implemented before
mlxsw_sp_acl_bf_index_get() but is used after it.
Adding a new function for Spectrum-4 would make them further apart still.
Fix by moving them around.
Signed-off-by: Amit Cohen <amcohen@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
