linux-2.6-microblaze.git
5 years agoMerge branch '100GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/jkirsher/next...
David S. Miller [Thu, 1 Aug 2019 17:41:43 +0000 (13:41 -0400)]
Merge branch '100GbE' of git://git./linux/kernel/git/jkirsher/next-queue

Jeff Kirsher says:

====================
100GbE Intel Wired LAN Driver Updates 2019-07-31

This series contains updates to ice driver only.

Paul adds support for reporting what the link partner is advertising for
flow control settings.

Jake fixes the hardware statistics register which is prone to rollover
since the statistic registers are either 32 or 40 bits wide, depending
on which register is being read.  So use a 64 bit software statistic to
store off the hardware statistics to track past when it rolls over.
Fixes an issue with the locking of the control queue, where locks were
being destroyed at run time.

Tony fixes an issue that was created when interrupt tracking was
refactored and the call to ice_vsi_setup_vector_base() was removed from
the PF VSI instead of the VF VSI.  Adds a check before trying to
configure a port to ensure that media is attached.

Brett fixes an issue in the receive queue configuration where prefena
(Prefetch Enable) was being set to 0 which caused the hardware to only
fetch descriptors when there are none free in the cache for a received
packet.  Updates the driver to only bump the receive tail once per
napi_poll call, instead of the current model of bumping the tail up to 4
times per napi_poll call.  Adds statistics for receive drops at the port
level to ethtool/netlink.  Cleans up duplicate code in the allocation of
receive buffer code.

Akeem updates the driver to ensure that VFs stay disabled until the
setup or reset is completed.  Modifies the driver to use the allocated
number of transmit queues per VSI to set up the scheduling tree versus
using the total number of available transmit queues.  Also fix the
driver to update the total number of configured queues, after a
successful VF request to change its number of queues before updating the
corresponding VSI for that VF.  Cleaned up unnecessary flags that are no
longer needed.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agoMerge branch 'net-hns3-some-code-optimizations-bugfixes-features'
David S. Miller [Thu, 1 Aug 2019 17:32:13 +0000 (13:32 -0400)]
Merge branch 'net-hns3-some-code-optimizations-bugfixes-features'

Huazhong Tan says:

====================
net: hns3: some code optimizations & bugfixes & features

This patch-set includes code optimizations, bugfixes and features for
the HNS3 ethernet controller driver.

[patch 01/12] adds support for reporting link change event.

[patch 02/12] adds handler for NCSI error.

[patch 03/12] fixes bug related to debugfs.

[patch 04/12] adds a code optimization for setting ring parameters.

[patch 05/12 - 09/12] adds some cleanups.

[patch 10/12 - 12/12] adds some patches related to reset issue.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: activate reset timer when calling reset_event
Huazhong Tan [Thu, 1 Aug 2019 03:55:45 +0000 (11:55 +0800)]
net: hns3: activate reset timer when calling reset_event

When calling hclge_reset_event() within HCLGE_RESET_INTERVAL,
it returns directly now. If no one call it again, then the
error which needs a reset to fix it can not be fixed.

So this patch activates the reset timer for this case, and
adds checking in the end of the reset procedure to make this
error fixed earlier.

Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Reviewed-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: clear reset interrupt status in hclge_irq_handle()
Huazhong Tan [Thu, 1 Aug 2019 03:55:44 +0000 (11:55 +0800)]
net: hns3: clear reset interrupt status in hclge_irq_handle()

Currently, the reset interrupt is cleared in the reset task, which
is too late. Since, when the hardware finish the previous reset,
it can begin to do a new global/IMP reset, if this new coming reset
type is same as the previous one, the driver will clear them together,
then driver can not get that there is another reset, but the hardware
still wait for the driver to deal with the second one.

So this patch clears PF's reset interrupt status in the
hclge_irq_handle(), the hardware waits for handshaking from
driver before doing reset, so the driver and hardware deal with reset
one by one.

BTW, when VF doing global/IMP reset, it reads PF's reset interrupt
register to get that whether PF driver's re-initialization is done,
since VF's re-initialization should be done after PF's. So we add
a new command and a register bit to do that. When VF receive reset
interrupt, it sets up this bit, and PF finishes re-initialization
send command to clear this bit, then VF do re-initialization.

Fixes: 4ed340ab8f49 ("net: hns3: Add reset process in hclge_main")
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Reviewed-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: fix some reset handshake issue
Huazhong Tan [Thu, 1 Aug 2019 03:55:43 +0000 (11:55 +0800)]
net: hns3: fix some reset handshake issue

Currently, the driver sets handshake status to tell the hardware
that the driver have downed the netdev and it can continue with
reset process. The driver will clear the handshake status when
re-initializing the CMDQ, and does not recover this status
when reset fail, which may cause the hardware to wait for
the handshake status to be set and not being able to continue
with reset process.

So this patch delays clearing handshake status just before UP,
and recovers this status when reset fail.

BTW, this patch adds a new function hclge(vf)_reset_handshake() to
deal with the reset handshake issue, and renames
HCLGE(VF)_NIC_CMQ_ENABLE to HCLGE(VF)_NIC_SW_RST_RDY which
represents this register bit more accurately.

Fixes: ada13ee3db7b ("net: hns3: add handshake with hardware while doing reset")
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Reviewed-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: rename a member in struct hclge_mac_ethertype_idx_rd_cmd
Guojia Liao [Thu, 1 Aug 2019 03:55:42 +0000 (11:55 +0800)]
net: hns3: rename a member in struct hclge_mac_ethertype_idx_rd_cmd

The member 'mac_add' defined in hclge_mac_ethertype_idx_rd_cmd
means MAC address, so 'mac_addr' is a better name for it.

Signed-off-by: Guojia Liao <liaoguojia@huawei.com>
Signed-off-by: Guangbin Huang <huangguangbin2@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: simplify hclge_cmd_query_error()
Weihang Li [Thu, 1 Aug 2019 03:55:41 +0000 (11:55 +0800)]
net: hns3: simplify hclge_cmd_query_error()

The 4th and 5th parameter of hclge_cmd_query_error is useless, so this
patch removes them.

Signed-off-by: Weihang Li <liweihang@hisilicon.com>
Reviewed-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: minior error handling change for hclge_tm_schd_info_init
Yunsheng Lin [Thu, 1 Aug 2019 03:55:40 +0000 (11:55 +0800)]
net: hns3: minior error handling change for hclge_tm_schd_info_init

When hclge_tm_schd_info_update calls hclge_tm_schd_info_init to
initialize the schedule info, hdev->tm_info.num_pg and
hdev->tx_sch_mode is not changed, which makes the checking in
hclge_tm_schd_info_init unnecessary.

So this patch moves the hdev->tm_info.num_pg and hdev->tx_sch_mode
checking into hclge_tm_schd_init and changes the return type of
hclge_tm_schd_info_init from int to void.

Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Reviewed-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: minor cleanup in hns3_clean_rx_ring
Yunsheng Lin [Thu, 1 Aug 2019 03:55:39 +0000 (11:55 +0800)]
net: hns3: minor cleanup in hns3_clean_rx_ring

The unused_count variable is used to indicate how many
RX BD need attaching new buffer in hns3_clean_rx_ring,
and the clean_count variable has the similar meaning.

This patch removes the clean_count variable and use
unused_count to uniformly indicate the RX BD that need
attaching new buffer.

This patch also clean up some coding style related to
variable assignment in hns3_clean_rx_ring.

Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Reviewed-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: remove unnecessary variable in hclge_get_mac_vlan_cmd_status()
Jian Shen [Thu, 1 Aug 2019 03:55:38 +0000 (11:55 +0800)]
net: hns3: remove unnecessary variable in hclge_get_mac_vlan_cmd_status()

The local variable return_status in hclge_get_mac_val_cmd_status()
is useless. So this patch returns the error code directly, instead of
using this variable. Also, replace some '%d' with '%u' in
hclge_get_mac_val_cmd_status().

Signed-off-by: Jian Shen <shenjian15@huawei.com>
Reviewed-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: refine for set ring parameters
Jian Shen [Thu, 1 Aug 2019 03:55:37 +0000 (11:55 +0800)]
net: hns3: refine for set ring parameters

Previously, when changing the ring parameters, we free the old
ring resources firstly, and then setup the new ring resources.
In some case of an memory allocation fail, there will be no
resources to use. This patch refines it by setup new ring
resources and free the old ring resources in order.

Also reduce the max ring BD number to 32760 according to UM.

Signed-off-by: Jian Shen <shenjian15@huawei.com>
Reviewed-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: do not query unsupported commands in debugfs
Yufeng Mo [Thu, 1 Aug 2019 03:55:36 +0000 (11:55 +0800)]
net: hns3: do not query unsupported commands in debugfs

Some commands are not supported on DCB-unsupported ports.
This patch distinguishes these commands and does not query
unsupported commands in debugfs.

This patch also fix an error in the dump "qos buf cfg"
command in debugfs.

Fixes: 2849d4e7a1be ("net: hns3: Add "tc config" info query function")
Fixes: 7d9d7f8864ba ("net: hns3: Add "qos buffer" config info query function")
Signed-off-by: Yufeng Mo <moyufeng@huawei.com>
Reviewed-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: add handler for NCSI error mailbox
Huazhong Tan [Thu, 1 Aug 2019 03:55:35 +0000 (11:55 +0800)]
net: hns3: add handler for NCSI error mailbox

When NCSI has HW error, the IMP will report this error to the driver
by sending a mailbox. After received this message, the driver should
assert a global reset to fix this kind of HW error.

Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Reviewed-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: hns3: add link change event report
Jian Shen [Thu, 1 Aug 2019 03:55:34 +0000 (11:55 +0800)]
net: hns3: add link change event report

Previously, PF updates link status per second. For some scenario,
it requires link down event being reported more quickly.
To solve it, firmware pushes the link change event to PF with
CMDQ message, and driver updates the link status directly.

Signed-off-by: Jian Shen <shenjian15@huawei.com>
Reviewed-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: phy: xgene: use devm_platform_ioremap_resource() to simplify code
YueHaibing [Thu, 1 Aug 2019 12:46:30 +0000 (20:46 +0800)]
net: phy: xgene: use devm_platform_ioremap_resource() to simplify code

Use devm_platform_ioremap_resource() to simplify the code a bit.
This is detected by coccinelle.

Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agobcm63xx_enet: use devm_platform_ioremap_resource() to simplify code
YueHaibing [Thu, 1 Aug 2019 12:39:08 +0000 (20:39 +0800)]
bcm63xx_enet: use devm_platform_ioremap_resource() to simplify code

Use devm_platform_ioremap_resource() to simplify the code a bit.
This is detected by coccinelle.

Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: qcom/emac: use devm_platform_ioremap_resource() to simplify code
YueHaibing [Thu, 1 Aug 2019 12:34:30 +0000 (20:34 +0800)]
net: qcom/emac: use devm_platform_ioremap_resource() to simplify code

Use devm_platform_ioremap_resource() to simplify the code a bit.
This is detected by coccinelle.

Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: mediatek: use devm_platform_ioremap_resource() to simplify code
YueHaibing [Thu, 1 Aug 2019 12:33:08 +0000 (20:33 +0800)]
net: mediatek: use devm_platform_ioremap_resource() to simplify code

Use devm_platform_ioremap_resource() to simplify the code a bit.
This is detected by coccinelle.

Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: dsa: bcm_sf2: use devm_platform_ioremap_resource() to simplify code
YueHaibing [Thu, 1 Aug 2019 12:29:11 +0000 (20:29 +0800)]
net: dsa: bcm_sf2: use devm_platform_ioremap_resource() to simplify code

Use devm_platform_ioremap_resource() to simplify the code a bit.
This is detected by coccinelle.

Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: dsa: b53: use devm_platform_ioremap_resource() to simplify code
YueHaibing [Thu, 1 Aug 2019 12:27:32 +0000 (20:27 +0800)]
net: dsa: b53: use devm_platform_ioremap_resource() to simplify code

Use devm_platform_ioremap_resource() to simplify the code a bit.
This is detected by coccinelle.

Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: dsa: lantiq: use devm_platform_ioremap_resource() to simplify code
YueHaibing [Thu, 1 Aug 2019 12:25:46 +0000 (20:25 +0800)]
net: dsa: lantiq: use devm_platform_ioremap_resource() to simplify code

Use devm_platform_ioremap_resource() to simplify the code a bit.
This is detected by coccinelle.

Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agomvpp2: use devm_platform_ioremap_resource() to simplify code
YueHaibing [Thu, 1 Aug 2019 12:22:02 +0000 (20:22 +0800)]
mvpp2: use devm_platform_ioremap_resource() to simplify code

Use devm_platform_ioremap_resource() to simplify the code a bit.
This is detected by coccinelle.

Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: bridge: mcast: add delete due to fast-leave mdb flag
Nikolay Aleksandrov [Tue, 30 Jul 2019 12:20:41 +0000 (15:20 +0300)]
net: bridge: mcast: add delete due to fast-leave mdb flag

In user-space there's no way to distinguish why an mdb entry was deleted
and that is a problem for daemons which would like to keep the mdb in
sync with remote ends (e.g. mlag) but would also like to converge faster.
In almost all cases we'd like to age-out the remote entry for performance
and convergence reasons except when fast-leave is enabled. In that case we
want explicit immediate remote delete, thus add mdb flag which is set only
when the entry is being deleted due to fast-leave.

Signed-off-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agotc-testing: Clarify the use of tdc's -d option
Lucas Bates [Mon, 29 Jul 2019 23:18:12 +0000 (19:18 -0400)]
tc-testing: Clarify the use of tdc's -d option

The -d command line argument to tdc requires the name of a physical device
on the system where the tests will be run. If -d has not been used, tdc
will skip tests that require a physical device.

This patch is intended to better document what the -d option does and how
it is used.

Signed-off-by: Lucas Bates <lucasb@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agoMerge tag 'mlx5-updates-2019-07-29' of git://git.kernel.org/pub/scm/linux/kernel...
David S. Miller [Wed, 31 Jul 2019 22:48:01 +0000 (18:48 -0400)]
Merge tag 'mlx5-updates-2019-07-29' of git://git./linux/kernel/git/saeed/linux

Saeed Mahameed says:

====================
mlx5-updates-2019-07-29

This series includes updates to mlx5 driver,
1) Simplifications, cleanup and warning prints improvements

2) From Vlad Buslov:
Refactor mlx5 tc flow handling for unlocked execution (Part 1)

Currently, all cls API hardware offloads driver callbacks require caller
to hold rtnl lock when calling them. Cls API has already been updated to
update software filters in parallel (on classifiers that support
unlocked execution), however hardware offloads code still obtains rtnl
lock before calling driver tc callbacks. This set implements partial
support for unlocked execution that is leveraged by follow up
refactorings in specific mlx5 tc subsystems and patch to cls API that
implements API that allows drivers to register their callbacks as
rtnl-unlocked.

In mlx5 tc code mlx5e_tc_flow is the main structure that is used to
represent tc filter. Currently, code the structure itself and its
handlers in both tc and eswitch layers do not implement any kind of
synchronizations and assume external global synchronization provided by
rtnl lock instead. Implement following changes to remove dependency on
rtnl lock in flow handling code that are intended to be used a
groundwork for following changes to provide fully rtnl-independent mlx5
tc:

- Extend struct mlx5e_tc_flow with atomic reference counter and rcu to
  allow concurrent access from multiple tc and neigh update workqueue
  instances without introducing any additional locks specific to the
  structure. Its 'flags' field type is changed to atomic bitmask ops which
  is necessary for tc to interact with other concurrent tc instances or
  concurrent neigh update that need to skip flows that are not fully
  initialized (new INIT_DONE flow flag) and can change the flags
  according to neighbor state (flipping OFFLOADED flag).

- Protect unready flows list by new uplink_priv->unready_flows_lock
  mutex.

- Convert calls to netdev APIs that require rtnl lock in flow handling
  code to their rcu counterparts.

- Modify eswitch code that is called from tc layer and assume implicit
  external synchronization to be concurrency safe: change
  esw->offloads.num_flows type to atomic integer and re-arrange
  esw->state_lock usage to protect additional data.

Some of approaches to synchronizations presented in this patch set are
quite complicated (lockless concurrent usage of data structures with rcu
and reference counting, using fine-grained locking when necessary, retry
mechanisms to handle concurrent insertion of another instance of data
structure with same key, etc.). This is necessary to allow calling the
firmware in parallel in most cases, which is the main motivation of this
change since firmware calls are mach heavier operation than atomic
operations, multitude of locks and potential multiple retries during
concurrent accesses to same elements.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agoice: Bump version number
Tony Nguyen [Wed, 26 Jun 2019 09:20:27 +0000 (02:20 -0700)]
ice: Bump version number

Update driver version to 0.7.5

Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Remove flag to track VF interrupt status
Akeem G Abodunrin [Wed, 26 Jun 2019 09:20:26 +0000 (02:20 -0700)]
ice: Remove flag to track VF interrupt status

As a result of refactoring of VF VSIs interrupts code, there is no
need to track its configuration status again with ICE_VF_STATE_CFG_INTR
flag - In fact, it is not being checked anywhere in the code right now, so
this patch removes the dead code as applicable to the flag.

Signed-off-by: Akeem G Abodunrin <akeem.g.abodunrin@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Remove unnecessary flag ICE_FLAG_MSIX_ENA
Brett Creeley [Wed, 26 Jun 2019 09:20:25 +0000 (02:20 -0700)]
ice: Remove unnecessary flag ICE_FLAG_MSIX_ENA

This flag is not needed and is called every time we re-enable interrupts
in the hotpath so remove it. Also remove ice_vsi_req_irq() because it
was a wrapper function for ice_vsi_req_irq_msix() whose sole purpose was
checking the ICE_FLAG_MSIX_ENA flag.

Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Don't return error for disabling LAN Tx queue that does exist
Akeem G Abodunrin [Wed, 26 Jun 2019 09:20:24 +0000 (02:20 -0700)]
ice: Don't return error for disabling LAN Tx queue that does exist

Since Tx rings are being managed by FW/NVM, Tx rings might have not been
set up or driver had already wiped them off - In that case, call to
disable LAN Tx queue is being returned as not in existence. This patch
makes sure we don't return unnecessary error for such scenario.

Signed-off-by: Akeem G Abodunrin <akeem.g.abodunrin@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Remove duplicate code in ice_alloc_rx_bufs
Brett Creeley [Wed, 26 Jun 2019 09:20:23 +0000 (02:20 -0700)]
ice: Remove duplicate code in ice_alloc_rx_bufs

Currently if the call to ice_alloc_mapped_page() fails we jump to the
no_buf label, possibly call ice_release_rx_desc(), and return true
indicating that there is more work to do. In the success case we just
fall out of the while loop, possibly call ice_alloc_mapped_page(), and
return false saying we exhausted cleaned_count. This flow can be
improved by breaking if ice_alloc_mapped_page() fails and then the flow
outside of the while loop is the same for the failure and success case.

Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Add stats for Rx drops at the port level
Brett Creeley [Wed, 26 Jun 2019 09:20:22 +0000 (02:20 -0700)]
ice: Add stats for Rx drops at the port level

Currently we are not reporting dropped counts at the port level to
ethtool or netlink. This was found when debugging Rx dropped issues
and the total packets sent did not equal the total packets received
minus the rx_dropped, which was very confusing. To determine dropped
counts at the port level we need to read the PRTRPB_RDPC register.
To fix reporting we will store the dropped counts in the PF's
rx_discards. This will be reported to netlink by storing it in the
PF VSI's rx_missed_errors signaling that the receiver missed the
packet. Also, we will report this to ethtool in the rx_dropped.nic
field.

Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Update number of VF queue before setting VSI resources
Akeem G Abodunrin [Wed, 26 Jun 2019 09:20:21 +0000 (02:20 -0700)]
ice: Update number of VF queue before setting VSI resources

In case there is a request from a VF to change its number of queues, and
the request was successful, we need to update number of queues
configured on the VF before updating corresponding VSI for that VF,
especially LAN Tx queue tree and TC update, otherwise, we would continued
to use old value of vf->num_vf_qs for allocated Tx/Rx queues...

Signed-off-by: Akeem G Abodunrin <akeem.g.abodunrin@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Set up Tx scheduling tree based on alloc VSI Tx queues
Akeem G Abodunrin [Wed, 26 Jun 2019 09:20:20 +0000 (02:20 -0700)]
ice: Set up Tx scheduling tree based on alloc VSI Tx queues

This patch uses allocated number of Tx queues per VSI to set up its
scheduling tree instead of using total number of available Tx queues.
Only PF VSIs have total number of allocated Tx queues equal to number
of available Tx queues, other VSIs have different number of queues
configured.

Signed-off-by: Akeem G Abodunrin <akeem.g.abodunrin@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Only bump Rx tail and release buffers once per napi_poll
Brett Creeley [Wed, 26 Jun 2019 09:20:19 +0000 (02:20 -0700)]
ice: Only bump Rx tail and release buffers once per napi_poll

Currently we bump the Rx tail and release/give buffers to hardware every
16 descriptors. This causes us to bump Rx tail up to 4 times per
napi_poll call. Also we are always bumping tail on an odd index and this
is a problem because hardware ignores the lower 3 bits in the QRX_TAIL
register. This is making it so hardware sees tail bumps only every 8
descriptors. Instead lets only bump Rx tail once per napi_poll if
the value aligns with hardware's expectations of the lower 3 bits being
cleared. Also only release/give Rx buffers once per napi_poll call.

Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Disable VFs until reset is completed
Akeem G Abodunrin [Wed, 26 Jun 2019 09:20:18 +0000 (02:20 -0700)]
ice: Disable VFs until reset is completed

This patch adds code to clear VFs enable status until reset is completed,
and Tx/Rx rings are setup. Without this patch, the code flow request Tx
queues to be disabled after reset, especially PFR - where VF VSI Tx rings
have already been wiped off in the NVM and result to adminq error based on
the call to disable Tx LAN queue in ice_reset_all_vfs function call.

Signed-off-by: Akeem G Abodunrin <akeem.g.abodunrin@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Do not configure port with no media
Tony Nguyen [Wed, 26 Jun 2019 09:20:17 +0000 (02:20 -0700)]
ice: Do not configure port with no media

The firmware reports an error when trying to configure a port with no
media. Instead of always configuring the port, check for media before
attempting to configure it. In the absence of media, turn off link and
poll for media to become available before re-enabling link.

Move ice_force_phys_link_state() up to avoid forward declaration.

Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: separate out control queue lock creation
Jacob Keller [Wed, 26 Jun 2019 09:20:16 +0000 (02:20 -0700)]
ice: separate out control queue lock creation

The ice_init_all_ctrlq and ice_shutdown_all_ctrlq functions create and
destroy the locks used to protect the send and receive process of each
control queue.

This is problematic, as the driver may use these functions to shutdown
and re-initialize the control queues at run time. For example, it may do
this in response to a device reset.

If the driver failed to recover from a reset, it might leave the control
queues offline. In this case, the locks will no longer be initialized.
A later call to ice_sq_send_cmd will then attempt to acquire a lock that
has been destroyed.

It is incorrect behavior to access a lock that has been destroyed.

Indeed, ice_aq_send_cmd already tries to avoid accessing an offline
control queue, but the check occurs inside the lock.

The root of the problem is that the locks are destroyed at run time.

Modify ice_init_all_ctrlq and ice_shutdown_all_ctrlq such that they no
longer create or destroy the locks.

Introduce new functions, ice_create_all_ctrlq and ice_destroy_all_ctrlq.
Call these functions in ice_init_hw and ice_deinit_hw.

Now, the control queue locks will remain valid for the life of the
driver, and will not be destroyed until the driver unloads.

This also allows removing a duplicate check of the sq.count and
rq.count values when shutting down the controlqs. The ice_shutdown_ctrlq
function already checks this value under the lock. Previously
commit dec64ff10ed9 ("ice: use [sr]q.count when checking if queue is
initialized") needed this check to happen outside the lock, because it
prevented duplicate attempts at destroying the locks.

The driver may now safely use ice_init_all_ctrlq and
ice_shutdown_all_ctrlq while handling reset events, without causing the
locks to be invalid.

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Always set prefena when configuring an Rx queue
Brett Creeley [Wed, 26 Jun 2019 09:20:15 +0000 (02:20 -0700)]
ice: Always set prefena when configuring an Rx queue

Currently we are always setting prefena to 0. This is causing the
hardware to only fetch descriptors when there are none free in the cache
for a received packet instead of prefetching when it has used the last
descriptor regardless of incoming packets. Fix this by allowing the
hardware to prefetch Rx descriptors.

Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: Move vector base setup to PF VSI
Tony Nguyen [Wed, 26 Jun 2019 09:20:14 +0000 (02:20 -0700)]
ice: Move vector base setup to PF VSI

When interrupt tracking was refactored, during rebuild, the call to
ice_vsi_setup_vector_base() was inadvertently removed from the PF VSI
instead of being removed from the VF VSI. During reset, the failure to
properly setup the vector base generates a call trace. Correct this so
that resets/rebuilds properly complete.

Fixes: cbe66bfee6a0 ("ice: Refactor interrupt tracking")
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: track hardware stat registers past rollover
Jacob Keller [Wed, 26 Jun 2019 09:20:13 +0000 (02:20 -0700)]
ice: track hardware stat registers past rollover

Currently, ice_stat_update32 and ice_stat_update40 will limit the
value of the software statistic to 32 or 40 bits wide, depending on
which register is being read.

This means that if a driver is running for a long time, the displayed
software register values will roll over to zero at 40 bits or 32 bits.

This occurs because the functions directly assign the difference between
the previous value and current value of the hardware statistic.

Instead, add this value to the current software statistic, and then
update the previous value.

In this way, each time ice_stat_update40 or ice_stat_update32 are
called, they will increment the software tracking value by the
difference of the hardware register from its last read. The software
tracking value will correctly count up until it overflows a u64.

The only requirement is that the ice_stat_update functions be called at
least once each time the hardware register overflows.

While we're fixing ice_stat_update40, modify it to use rd64 instead of
two calls to rd32. Additionally, drop the now unnecessary hireg
function parameter.

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agoice: add lp_advertising flow control support
Paul Greenwalt [Wed, 26 Jun 2019 09:20:12 +0000 (02:20 -0700)]
ice: add lp_advertising flow control support

Add support for reporting link partner advertising when
ETHTOOL_GLINKSETTINGS defined. Get pause param reports the Tx/Rx
pause configured, and then ethtool issues ETHTOOL_GSET ioctl and
ice_get_settings_link_up reports the negotiated Tx/Rx pause. Negotiated
pause frame report per IEEE 802.3-2005 table 288-3.

$ ethtool --show-pause ens6f0
Pause parameters for ens6f0:
Autonegotiate:  on
RX:             on
TX:             on
RX negotiated:  on
TX negotiated:  on

$ ethtool ens6f0
Settings for ens6f0:
        Supported ports: [ FIBRE ]
        Supported link modes:   25000baseCR/Full
        Supported pause frame use: Symmetric
        Supports auto-negotiation: Yes
        Supported FEC modes: None BaseR RS
        Advertised link modes:  25000baseCR/Full
        Advertised pause frame use: Symmetric Receive-only
        Advertised auto-negotiation: Yes
        Advertised FEC modes: None BaseR RS
        Link partner advertised link modes:  Not reported
        Link partner advertised pause frame use: Symmetric
        Link partner advertised auto-negotiation: Yes
        Link partner advertised FEC modes: Not reported
        Speed: 25000Mb/s
        Duplex: Full
        Port: Direct Attach Copper
        PHYAD: 0
        Transceiver: internal
        Auto-negotiation: on
        Supports Wake-on: g
        Wake-on: g
        Current message level: 0x00000007 (7)
                               drv probe link
        Link detected: yes

When ETHTOOL_GLINKSETTINGS is not defined, get pause param reports the
negotiated Tx/Rx pause.

Signed-off-by: Paul Greenwalt <paul.greenwalt@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
5 years agostaging/octeon: Fix build error without CONFIG_NETDEVICES
YueHaibing [Wed, 31 Jul 2019 16:02:19 +0000 (18:02 +0200)]
staging/octeon: Fix build error without CONFIG_NETDEVICES

While do COMPILE_TEST build without CONFIG_NETDEVICES,
we get Kconfig warning:

WARNING: unmet direct dependencies detected for PHYLIB
  Depends on [n]: NETDEVICES [=n]
  Selected by [y]:
  - OCTEON_ETHERNET [=y] && STAGING [=y] && (CAVIUM_OCTEON_SOC && NETDEVICES [=n] || COMPILE_TEST [=y])

Reported-by: Hulk Robot <hulkci@huawei.com>
Reported-by: Mark Brown <broonie@kernel.org>
Fixes: 171a9bae68c7 ("staging/octeon: Allow test build on !MIPS")
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agoMerge tag 'mac80211-next-for-davem-2019-07-31' of git://git.kernel.org/pub/scm/linux...
David S. Miller [Wed, 31 Jul 2019 15:59:41 +0000 (08:59 -0700)]
Merge tag 'mac80211-next-for-davem-2019-07-31' of git://git./linux/kernel/git/jberg/mac80211-next

Johannes Berg says:

====================
We have a reasonably large number of changes:
 * lots more HE (802.11ax) support, particularly things
   relevant for the the AP side, but also mesh support
 * debugfs cleanups from Greg
 * some more work on extended key ID
 * start using genl parallel_ops, as preparation for
   weaning ourselves off RTNL and getting parallelism
 * various other changes all over
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agoMerge branch 'mlxsw-Test-coverage-for-DSCP-leftover-fix'
David S. Miller [Wed, 31 Jul 2019 15:47:14 +0000 (08:47 -0700)]
Merge branch 'mlxsw-Test-coverage-for-DSCP-leftover-fix'

Petr Machata says:

====================
mlxsw: Test coverage for DSCP leftover fix

This patch set fixes some global scope pollution issues in the DSCP tests
(in patch #1), and then proceeds (in patch #2) to add a new test for
checking whether, after DSCP prioritization rules are removed from a port,
DSCP rewrites consistently to zero, instead of the last removed rule still
staying in effect.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agoselftests: mlxsw: Add a test for leftover DSCP rule
Petr Machata [Wed, 31 Jul 2019 10:30:27 +0000 (10:30 +0000)]
selftests: mlxsw: Add a test for leftover DSCP rule

Commit dedfde2fe1c4 ("mlxsw: spectrum_dcb: Configure DSCP map as the last
rule is removed") fixed a problem in mlxsw where last DSCP rule to be
removed remained in effect when DSCP rewrite was applied.

Add a selftest that covers this problem.

Signed-off-by: Petr Machata <petrm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agoselftests: mlxsw: Fix local variable declarations in DSCP tests
Petr Machata [Wed, 31 Jul 2019 10:30:26 +0000 (10:30 +0000)]
selftests: mlxsw: Fix local variable declarations in DSCP tests

These two tests have some problems in the global scope pollution and on
contrary, contain unnecessary local declarations. Fix them.

Signed-off-by: Petr Machata <petrm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agomyri10ge: remove unneeded variable
Ding Xiang [Wed, 31 Jul 2019 08:53:46 +0000 (16:53 +0800)]
myri10ge: remove unneeded variable

"error" is unneeded,just return 0

Signed-off-by: Ding Xiang <dingxiang@cmss.chinamobile.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: ag71xx: Slighly simplify code in 'ag71xx_rings_init()'
Christophe JAILLET [Wed, 31 Jul 2019 08:06:38 +0000 (10:06 +0200)]
net: ag71xx: Slighly simplify code in 'ag71xx_rings_init()'

A few lines above, we have:
   tx_size = BIT(tx->order);

So use 'tx_size' directly to be consistent with the way 'rx->descs_cpu' and
'rx->descs_dma' are computed below.

Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agomac80211: HE STA disassoc due to QOS NULL not sent
Shay Bar [Wed, 3 Jul 2019 13:18:48 +0000 (16:18 +0300)]
mac80211: HE STA disassoc due to QOS NULL not sent

In case of HE AP-STA link, ieee80211_send_nullfunc() will not
send the QOS NULL packet to check if AP is still associated.

In this case, probe_send_count will be non-zero and
ieee80211_sta_work() will later disassociate the AP, even
though no packet was ever sent.

Fix this by decrementing probe_send_count and not calling
ieee80211_send_nullfunc() in case of HE link, so that we
still wait for some time for the AP beacon to reappear and
don't disconnect right away.

Signed-off-by: Shay Bar <shay.bar@celeno.com>
Link: https://lore.kernel.org/r/20190703131848.22879-1-shay.bar@celeno.com
[clarify commit message]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
5 years agomac80211: allow setting spatial reuse parameters from bss_conf
John Crispin [Tue, 30 Jul 2019 16:37:01 +0000 (18:37 +0200)]
mac80211: allow setting spatial reuse parameters from bss_conf

Store the OBSS PD parameters inside bss_conf when bringing up an AP and/or
when a station connects to an AP. This allows the driver to configure the
HW accordingly.

Signed-off-by: John Crispin <john@phrozen.org>
Link: https://lore.kernel.org/r/20190730163701.18836-3-john@phrozen.org
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
5 years agonl80211: add strict start type
Johannes Berg [Wed, 31 Jul 2019 08:58:20 +0000 (10:58 +0200)]
nl80211: add strict start type

Add a strict start type so all new attributes starting from
NL80211_ATTR_HE_OBSS_PD are validated strictly.

Signed-off-by: Johannes Berg <johannes.berg@intel.com>
5 years agocfg80211: add support for parsing OBBS_PD attributes
John Crispin [Tue, 30 Jul 2019 16:37:00 +0000 (18:37 +0200)]
cfg80211: add support for parsing OBBS_PD attributes

Add the data structure, policy and parsing code allowing userland to send
the OBSS PD information into the kernel.

Signed-off-by: John Crispin <john@phrozen.org>
Link: https://lore.kernel.org/r/20190730163701.18836-2-john@phrozen.org
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
5 years agomac80211: reject zero MAC address in add station
Karthikeyan Periyasamy [Wed, 24 Jul 2019 09:16:10 +0000 (14:46 +0530)]
mac80211: reject zero MAC address in add station

This came up in fuzz testing, and really we don't consider
all-zeroes to be a valid MAC address in most places, so
also reject it here to avoid confusion later on.

Signed-off-by: Karthikeyan Periyasamy <periyasa@codeaurora.org>
Link: https://lore.kernel.org/r/1563959770-21570-1-git-send-email-periyasa@codeaurora.org
[rewrite commit message]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
5 years agocfg80211: use parallel_ops for genl
Johannes Berg [Mon, 29 Jul 2019 14:31:09 +0000 (16:31 +0200)]
cfg80211: use parallel_ops for genl

Over time, we really need to get rid of all of our global locking.
One of the things needed is to use parallel_ops. This isn't really
the most important (RTNL is much more important) but OTOH we just
keep adding uses of genl_family_attrbuf() now. Use .parallel_ops to
disallow this.

Reviewed-By: Denis Kenzior <denkenz@gmail.com>
Link: https://lore.kernel.org/r/20190729143109.18683-1-johannes@sipsolutions.net
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
5 years agomac80211_hwsim: fill boottime_ns in netlink RX path
Johannes Berg [Mon, 29 Jul 2019 16:06:05 +0000 (18:06 +0200)]
mac80211_hwsim: fill boottime_ns in netlink RX path

Give a proper boottime_ns value for netlink RX to avoid scan
issues here.

Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Link: https://lore.kernel.org/r/20190729160605.1074-1-johannes@sipsolutions.net
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
5 years agomac80211: add missing null return check from call to ieee80211_get_sband
Colin Ian King [Tue, 30 Jul 2019 14:32:05 +0000 (15:32 +0100)]
mac80211: add missing null return check from call to ieee80211_get_sband

The return from ieee80211_get_sband can potentially be a null pointer, so
it seems prudent to add a null check to avoid a null pointer dereference
on sband.

Addresses-Coverity: ("Dereference null return")
Fixes: 2ab45876756f ("mac80211: add support for the ADDBA extension element")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Link: https://lore.kernel.org/r/20190730143205.14261-1-colin.king@canonical.com
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
5 years agoMerge branch 'net-dsa-ksz-Add-Microchip-KSZ87xx-support'
David S. Miller [Tue, 30 Jul 2019 22:12:50 +0000 (15:12 -0700)]
Merge branch 'net-dsa-ksz-Add-Microchip-KSZ87xx-support'

Marek Vasut says:

====================
net: dsa: ksz: Add Microchip KSZ87xx support

This series adds support for Microchip KSZ87xx switches, which are
slightly simpler compared to KSZ9xxx .
====================

Signed-off-by: Marek Vasut <marex@denx.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: dsa: ksz: Add Microchip KSZ8795 DSA driver
Tristram Ha [Mon, 29 Jul 2019 17:49:47 +0000 (19:49 +0200)]
net: dsa: ksz: Add Microchip KSZ8795 DSA driver

Add Microchip KSZ8795 DSA driver.

Signed-off-by: Tristram Ha <Tristram.Ha@microchip.com>
Signed-off-by: Marek Vasut <marex@denx.de>
Cc: Andrew Lunn <andrew@lunn.ch>
Cc: David S. Miller <davem@davemloft.net>
Cc: Florian Fainelli <f.fainelli@gmail.com>
Cc: Tristram Ha <Tristram.Ha@microchip.com>
Cc: Vivien Didelot <vivien.didelot@gmail.com>
Cc: Woojung Huh <woojung.huh@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: dsa: ksz: Add KSZ8795 tag code
Tristram Ha [Mon, 29 Jul 2019 17:49:46 +0000 (19:49 +0200)]
net: dsa: ksz: Add KSZ8795 tag code

Add DSA tag code for Microchip KSZ8795 switch. The switch is simpler
and the tag is only 1 byte, instead of 2 as is the case with KSZ9477.

Signed-off-by: Tristram Ha <Tristram.Ha@microchip.com>
Signed-off-by: Marek Vasut <marex@denx.de>
Cc: Andrew Lunn <andrew@lunn.ch>
Cc: David S. Miller <davem@davemloft.net>
Cc: Florian Fainelli <f.fainelli@gmail.com>
Cc: Tristram Ha <Tristram.Ha@microchip.com>
Cc: Vivien Didelot <vivien.didelot@gmail.com>
Cc: Woojung Huh <woojung.huh@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agodt-bindings: net: dsa: ksz: document Microchip KSZ87xx family switches
Marek Vasut [Mon, 29 Jul 2019 17:49:45 +0000 (19:49 +0200)]
dt-bindings: net: dsa: ksz: document Microchip KSZ87xx family switches

Document Microchip KSZ87xx family switches. These include
KSZ8765 - 5 port switch
KSZ8794 - 4 port switch
KSZ8795 - 5 port switch

Signed-off-by: Marek Vasut <marex@denx.de>
Cc: Andrew Lunn <andrew@lunn.ch>
Cc: David S. Miller <davem@davemloft.net>
Cc: Florian Fainelli <f.fainelli@gmail.com>
Cc: Rob Herring <robh+dt@kernel.org>
Cc: Tristram Ha <Tristram.Ha@microchip.com>
Cc: Vivien Didelot <vivien.didelot@gmail.com>
Cc: Woojung Huh <woojung.huh@microchip.com>
Cc: devicetree@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agoMerge branch 'vsock-virtio-optimizations-to-increase-the-throughput'
David S. Miller [Tue, 30 Jul 2019 22:00:00 +0000 (15:00 -0700)]
Merge branch 'vsock-virtio-optimizations-to-increase-the-throughput'

Stefano Garzarella says:

====================
vsock/virtio: optimizations to increase the throughput

This series tries to increase the throughput of virtio-vsock with slight
changes.
While I was testing the v2 of this series I discovered an huge use of memory,
so I added patch 1 to mitigate this issue. I put it in this series in order
to better track the performance trends.

v5:
- rebased all patches on net-next
- added Stefan's R-b and Michael's A-b

v4: https://patchwork.kernel.org/cover/11047717
v3: https://patchwork.kernel.org/cover/10970145
v2: https://patchwork.kernel.org/cover/10938743
v1: https://patchwork.kernel.org/cover/10885431

Below are the benchmarks step by step. I used iperf3 [1] modified with VSOCK
support. As Michael suggested in the v1, I booted host and guest with 'nosmap'.

A brief description of patches:
- Patches 1:   limit the memory usage with an extra copy for small packets
- Patches 2+3: reduce the number of credit update messages sent to the
               transmitter
- Patches 4+5: allow the host to split packets on multiple buffers and use
               VIRTIO_VSOCK_MAX_PKT_BUF_SIZE as the max packet size allowed

                    host -> guest [Gbps]
pkt_size before opt   p 1     p 2+3    p 4+5

32         0.032     0.030    0.048    0.051
64         0.061     0.059    0.108    0.117
128        0.122     0.112    0.227    0.234
256        0.244     0.241    0.418    0.415
512        0.459     0.466    0.847    0.865
1K         0.927     0.919    1.657    1.641
2K         1.884     1.813    3.262    3.269
4K         3.378     3.326    6.044    6.195
8K         5.637     5.676   10.141   11.287
16K        8.250     8.402   15.976   16.736
32K       13.327    13.204   19.013   20.515
64K       21.241    21.341   20.973   21.879
128K      21.851    22.354   21.816   23.203
256K      21.408    21.693   21.846   24.088
512K      21.600    21.899   21.921   24.106

                    guest -> host [Gbps]
pkt_size before opt   p 1     p 2+3    p 4+5

32         0.045     0.046    0.057    0.057
64         0.089     0.091    0.103    0.104
128        0.170     0.179    0.192    0.200
256        0.364     0.351    0.361    0.379
512        0.709     0.699    0.731    0.790
1K         1.399     1.407    1.395    1.427
2K         2.670     2.684    2.745    2.835
4K         5.171     5.199    5.305    5.451
8K         8.442     8.500   10.083    9.941
16K       12.305    12.259   13.519   15.385
32K       11.418    11.150   11.988   24.680
64K       10.778    10.659   11.589   35.273
128K      10.421    10.339   10.939   40.338
256K      10.300     9.719   10.508   36.562
512K       9.833     9.808   10.612   35.979

As Stefan suggested in the v1, I measured also the efficiency in this way:
    efficiency = Mbps / (%CPU_Host + %CPU_Guest)

The '%CPU_Guest' is taken inside the VM. I know that it is not the best way,
but it's provided for free from iperf3 and could be an indication.

        host -> guest efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
pkt_size before opt   p 1     p 2+3    p 4+5

32         0.35      0.45     0.79     1.02
64         0.56      0.80     1.41     1.54
128        1.11      1.52     3.03     3.12
256        2.20      2.16     5.44     5.58
512        4.17      4.18    10.96    11.46
1K         8.30      8.26    20.99    20.89
2K        16.82     16.31    39.76    39.73
4K        30.89     30.79    74.07    75.73
8K        53.74     54.49   124.24   148.91
16K       80.68     83.63   200.21   232.79
32K      132.27    132.52   260.81   357.07
64K      229.82    230.40   300.19   444.18
128K     332.60    329.78   331.51   492.28
256K     331.06    337.22   339.59   511.59
512K     335.58    328.50   331.56   504.56

        guest -> host efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
pkt_size before opt   p 1     p 2+3    p 4+5

32         0.43      0.43     0.53     0.56
64         0.85      0.86     1.04     1.10
128        1.63      1.71     2.07     2.13
256        3.48      3.35     4.02     4.22
512        6.80      6.67     7.97     8.63
1K        13.32     13.31    15.72    15.94
2K        25.79     25.92    30.84    30.98
4K        50.37     50.48    58.79    59.69
8K        95.90     96.15   107.04   110.33
16K      145.80    145.43   143.97   174.70
32K      147.06    144.74   146.02   282.48
64K      145.25    143.99   141.62   406.40
128K     149.34    146.96   147.49   489.34
256K     156.35    149.81   152.21   536.37
512K     151.65    150.74   151.52   519.93

[1] https://github.com/stefano-garzarella/iperf/
====================

Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agovsock/virtio: change the maximum packet size allowed
Stefano Garzarella [Tue, 30 Jul 2019 15:43:34 +0000 (17:43 +0200)]
vsock/virtio: change the maximum packet size allowed

Since now we are able to split packets, we can avoid limiting
their sizes to VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE.
Instead, we can use VIRTIO_VSOCK_MAX_PKT_BUF_SIZE as the max
packet size.

Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agovhost/vsock: split packets to send using multiple buffers
Stefano Garzarella [Tue, 30 Jul 2019 15:43:33 +0000 (17:43 +0200)]
vhost/vsock: split packets to send using multiple buffers

If the packets to sent to the guest are bigger than the buffer
available, we can split them, using multiple buffers and fixing
the length in the packet header.
This is safe since virtio-vsock supports only stream sockets.

Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agovsock/virtio: fix locking in virtio_transport_inc_tx_pkt()
Stefano Garzarella [Tue, 30 Jul 2019 15:43:32 +0000 (17:43 +0200)]
vsock/virtio: fix locking in virtio_transport_inc_tx_pkt()

fwd_cnt and last_fwd_cnt are protected by rx_lock, so we should use
the same spinlock also if we are in the TX path.

Move also buf_alloc under the same lock.

Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agovsock/virtio: reduce credit update messages
Stefano Garzarella [Tue, 30 Jul 2019 15:43:31 +0000 (17:43 +0200)]
vsock/virtio: reduce credit update messages

In order to reduce the number of credit update messages,
we send them only when the space available seen by the
transmitter is less than VIRTIO_VSOCK_MAX_PKT_BUF_SIZE.

Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agovsock/virtio: limit the memory used per-socket
Stefano Garzarella [Tue, 30 Jul 2019 15:43:30 +0000 (17:43 +0200)]
vsock/virtio: limit the memory used per-socket

Since virtio-vsock was introduced, the buffers filled by the host
and pushed to the guest using the vring, are directly queued in
a per-socket list. These buffers are preallocated by the guest
with a fixed size (4 KB).

The maximum amount of memory used by each socket should be
controlled by the credit mechanism.
The default credit available per-socket is 256 KB, but if we use
only 1 byte per packet, the guest can queue up to 262144 of 4 KB
buffers, using up to 1 GB of memory per-socket. In addition, the
guest will continue to fill the vring with new 4 KB free buffers
to avoid starvation of other sockets.

This patch mitigates this issue copying the payload of small
packets (< 128 bytes) into the buffer of last packet queued, in
order to avoid wasting memory.

Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: Remove dev_err() usage after platform_get_irq()
Stephen Boyd [Tue, 30 Jul 2019 18:15:51 +0000 (11:15 -0700)]
net: Remove dev_err() usage after platform_get_irq()

We don't need dev_err() messages when platform_get_irq() fails now that
platform_get_irq() prints an error message itself when something goes
wrong. Let's remove these prints with a simple semantic patch.

// <smpl>
@@
expression ret;
struct platform_device *E;
@@

ret =
(
platform_get_irq(E, ...)
|
platform_get_irq_byname(E, ...)
);

if ( \( ret < 0 \| ret <= 0 \) )
{
(
-if (ret != -EPROBE_DEFER)
-{ ...
-dev_err(...);
-... }
|
...
-dev_err(...);
)
...
}
// </smpl>

While we're here, remove braces on if statements that only have one
statement (manually).

Cc: "David S. Miller" <davem@davemloft.net>
Cc: Kalle Valo <kvalo@codeaurora.org>
Cc: Saeed Mahameed <saeedm@mellanox.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Felix Fietkau <nbd@nbd.name>
Cc: Lorenzo Bianconi <lorenzo@kernel.org>
Cc: netdev@vger.kernel.org
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Stephen Boyd <swboyd@chromium.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agoMerge branch 'Finish-conversion-of-skb_frag_t-to-bio_vec'
David S. Miller [Tue, 30 Jul 2019 21:21:32 +0000 (14:21 -0700)]
Merge branch 'Finish-conversion-of-skb_frag_t-to-bio_vec'

Jonathan Lemon says:

====================
Finish conversion of skb_frag_t to bio_vec

The recent conversion of skb_frag_t to bio_vec did not include
skb_frag's page_offset.  Add accessor functions for this field,
utilize them, and remove the union, restoring the original structure.

v2:
  - rename accessors
  - follow kdoc conventions
====================

Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agolinux: Remove bvec page_offset, use bv_offset
Jonathan Lemon [Tue, 30 Jul 2019 14:40:34 +0000 (07:40 -0700)]
linux: Remove bvec page_offset, use bv_offset

Now that page_offset is referenced through accessors, remove
the union, and use bv_offset.

Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet: Use skb_frag_off accessors
Jonathan Lemon [Tue, 30 Jul 2019 14:40:33 +0000 (07:40 -0700)]
net: Use skb_frag_off accessors

Use accessor functions for skb fragment's page_offset instead
of direct references, in preparation for bvec conversion.

Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agolinux: Add skb_frag_t page_offset accessors
Jonathan Lemon [Tue, 30 Jul 2019 14:40:32 +0000 (07:40 -0700)]
linux: Add skb_frag_t page_offset accessors

Add skb_frag_off(), skb_frag_off_add(), skb_frag_off_set(),
and skb_frag_off_copy() accessors for page_offset.

Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agoMerge branch 'sctp-clean-up-sctp_connect-function'
David S. Miller [Tue, 30 Jul 2019 21:18:14 +0000 (14:18 -0700)]
Merge branch 'sctp-clean-up-sctp_connect-function'

Xin Long says:

====================
sctp: clean up __sctp_connect function

This patchset is to factor out some common code for
sctp_sendmsg_new_asoc() and __sctp_connect() into 2
new functioins.

v1->v2:
  - add the patch 1/5 to avoid a slab-out-of-bounds warning.
  - add some code comment for the check change in patch 2/5.
  - remove unused 'addrcnt' as Marcelo noticed in patch 3/5.
====================

Acked-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agosctp: factor out sctp_connect_add_peer
Xin Long [Tue, 30 Jul 2019 12:38:23 +0000 (20:38 +0800)]
sctp: factor out sctp_connect_add_peer

In this function factored out from sctp_sendmsg_new_asoc() and
__sctp_connect(), it adds a peer with the other addr into the
asoc after this asoc is created with the 1st addr.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agosctp: factor out sctp_connect_new_asoc
Xin Long [Tue, 30 Jul 2019 12:38:22 +0000 (20:38 +0800)]
sctp: factor out sctp_connect_new_asoc

In this function factored out from sctp_sendmsg_new_asoc() and
__sctp_connect(), it creates the asoc and adds a peer with the
1st addr.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agosctp: clean up __sctp_connect
Xin Long [Tue, 30 Jul 2019 12:38:21 +0000 (20:38 +0800)]
sctp: clean up __sctp_connect

__sctp_connect is doing quit similar things as sctp_sendmsg_new_asoc.
To factor out common functions, this patch is to clean up their code
to make them look more similar:

  1. create the asoc and add a peer with the 1st addr.
  2. add peers with the other addrs into this asoc one by one.

while at it, also remove the unused 'addrcnt'.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agosctp: check addr_size with sa_family_t size in __sctp_setsockopt_connectx
Xin Long [Tue, 30 Jul 2019 12:38:20 +0000 (20:38 +0800)]
sctp: check addr_size with sa_family_t size in __sctp_setsockopt_connectx

Now __sctp_connect() is called by __sctp_setsockopt_connectx() and
sctp_inet_connect(), the latter has done addr_size check with size
of sa_family_t.

In the next patch to clean up __sctp_connect(), we will remove
addr_size check with size of sa_family_t from __sctp_connect()
for the 1st address.

So before doing that, __sctp_setsockopt_connectx() should do
this check first, as sctp_inet_connect() does.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agosctp: only copy the available addr data in sctp_transport_init
Xin Long [Tue, 30 Jul 2019 12:38:19 +0000 (20:38 +0800)]
sctp: only copy the available addr data in sctp_transport_init

'addr' passed to sctp_transport_init is not always a whole size
of union sctp_addr, like the path:

  sctp_sendmsg() ->
  sctp_sendmsg_new_asoc() ->
  sctp_assoc_add_peer() ->
  sctp_transport_new() -> sctp_transport_init()

In the next patches, we will also pass the address length of data
only to sctp_assoc_add_peer().

So sctp_transport_init() should copy the only available data from
addr to peer->ipaddr, instead of 'peer->ipaddr = *addr' which may
cause slab-out-of-bounds.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agorxrpc: Fix -Wframe-larger-than= warnings from on-stack crypto
David Howells [Tue, 30 Jul 2019 14:56:57 +0000 (15:56 +0100)]
rxrpc: Fix -Wframe-larger-than= warnings from on-stack crypto

rxkad sometimes triggers a warning about oversized stack frames when
building with clang for a 32-bit architecture:

net/rxrpc/rxkad.c:243:12: error: stack frame size of 1088 bytes in function 'rxkad_secure_packet' [-Werror,-Wframe-larger-than=]
net/rxrpc/rxkad.c:501:12: error: stack frame size of 1088 bytes in function 'rxkad_verify_packet' [-Werror,-Wframe-larger-than=]

The problem is the combination of SYNC_SKCIPHER_REQUEST_ON_STACK() in
rxkad_verify_packet()/rxkad_secure_packet() with the relatively large
scatterlist in rxkad_verify_packet_1()/rxkad_secure_packet_encrypt().

The warning does not show up when using gcc, which does not inline the
functions as aggressively, but the problem is still the same.

Allocate the cipher buffers from the slab instead, caching the allocated
packet crypto request memory used for DATA packet crypto in the rxrpc_call
struct.

Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both")
Reported-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
cc: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agonet/mlx5e: Protect tc flow table with mutex
Vlad Buslov [Mon, 17 Sep 2018 11:01:58 +0000 (14:01 +0300)]
net/mlx5e: Protect tc flow table with mutex

TC flow table is created when first flow is added, and destroyed when last
flow is removed. This assumes that all accesses to the table are externally
synchronized with rtnl lock. To remove dependency on rtnl lock, add new
mutex mlx5e_tc_table->t_lock and use it to protect the flow table.

Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Reviewed-by: Jianbo Liu <jianbol@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Rely on rcu instead of rtnl lock when getting upper dev
Vlad Buslov [Tue, 12 Mar 2019 09:40:12 +0000 (11:40 +0200)]
net/mlx5e: Rely on rcu instead of rtnl lock when getting upper dev

Function netdev_master_upper_dev_get() generates warning if caller doesn't
hold rtnl lock. Modify rules update path to use rcu version of that
function.

Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Reviewed-by: Jianbo Liu <jianbol@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Eswitch, use state_lock to synchronize vlan change
Vlad Buslov [Tue, 11 Sep 2018 12:48:59 +0000 (15:48 +0300)]
net/mlx5e: Eswitch, use state_lock to synchronize vlan change

esw->state_lock is already used to protect vlan vport configuration change.
However, all preparation and correctness checks, and code that sets vport
data are not protected by this lock and assume external synchronization by
rtnl lock. In order to remove dependency on rtnl lock, extend
esw->state_lock protection to whole eswitch vlan add/del functions.

Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Reviewed-by: Jianbo Liu <jianbol@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Eswitch, change offloads num_flows type to atomic64
Vlad Buslov [Sun, 18 Nov 2018 11:12:59 +0000 (13:12 +0200)]
net/mlx5e: Eswitch, change offloads num_flows type to atomic64

Eswitch implements its own locking by means of state_lock mutex and
multiple fine-grained lock in containing data structures, and is supposed
to not rely on rtnl lock. However, eswitch offloads num_flows type is a
regular long long integer and cannot be modified concurrently. This is an
implicit assumptions that mlx5 tc is serialized (by rtnl lock or any other
means). In order to remove implicit dependency on rtnl lock, change
num_flows type to atomic64 to allow concurrent modifications.

Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Reviewed-by: Jianbo Liu <jianbol@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Protect unready flows with dedicated lock
Vlad Buslov [Wed, 13 Mar 2019 17:50:24 +0000 (19:50 +0200)]
net/mlx5e: Protect unready flows with dedicated lock

In order to remove dependency on rtnl lock for protecting unready_flows
list when reoffloading unready flows on workqueue, extend representor
uplink private structure with dedicated 'unready_flows_lock' mutex. Take
the lock in all users of unready_flows list before accessing it. Implement
helper functions to add and delete unready flow.

Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Reviewed-by: Jianbo Liu <jianbol@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Protect tc flows hashtable with rcu
Vlad Buslov [Wed, 10 Oct 2018 10:51:10 +0000 (13:51 +0300)]
net/mlx5e: Protect tc flows hashtable with rcu

In order to remove dependency on rtnl lock, access to tc flows hashtable
must be explicitly protected from concurrent flows removal.

Extend tc flow structure with rcu to allow concurrent parallel access. Use
rcu read lock to safely lookup flow in tc flows hash table, and take
reference to it. Use rcu free for flow deletion to accommodate concurrent
stats requests.

Add new DELETED flow flag. Imlement new flow_flag_test_and_set() helper
that is used to set a flag and return its previous value. Use it to
atomically set the flag in mlx5e_delete_flower() to guarantee that flow can
only be deleted once, even when same flow is deleted concurrently by
multiple tasks.

Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Reviewed-by: Jianbo Liu <jianbol@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Change flow flags type to unsigned long
Vlad Buslov [Thu, 8 Nov 2018 18:01:35 +0000 (20:01 +0200)]
net/mlx5e: Change flow flags type to unsigned long

To remove dependency on rtnl lock and allow concurrent modification of
'flags' field of tc flow structure, change flow flag type to unsigned long
and use atomic bit ops for reading and changing the flags. Implement
auxiliary functions for setting, resetting and getting specific flag, and
for checking most often used flag values.

Always set flags with smp_mb__before_atomic() to ensure that all
mlx5e_tc_flow are updated before concurrent readers can read new flags
value. Rearrange all code paths to actually set flow->rule[] pointers
before setting the OFFLOADED flag. On read side, use smp_mb__after_atomic()
when accessing flags to ensure that offload-related flow fields are only
read after the flags.

Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Reviewed-by: Jianbo Liu <jianbol@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Extend tc flow struct with reference counter
Vlad Buslov [Thu, 8 Nov 2018 15:46:06 +0000 (17:46 +0200)]
net/mlx5e: Extend tc flow struct with reference counter

With new classifier type that doesn't require rtnl lock, following
invariant holds:
 - Filter with specified cookie created only once.
 - Filter with specified cookie deleted only once.
 - Stats updates can be performed in parallel to each other.

Extend tc flow with rcu and reference counter. To protect from concurrent
delete, get reference to tc flow when:
 - Reading flow stats.
 - Accessing flow in neigh update handler.
 - Accessing flow in neigh update used value handler.

Only free flow when reference counter reached zero. Modify flow cleanup to
account for flows that could be not fully initialized by checking if flow
is actually in the list of corresponding mod_hdr, hairpin and encap
entries. Don't cleanup flow directly in case of error to allow concurrent
neigh update (neigh update will be modified to always take reference to
flow when using it).

Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Reviewed-by: Jianbo Liu <jianbol@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Simplify get_route_and_out_devs helper function
Eli Britstein [Mon, 3 Jun 2019 08:57:13 +0000 (08:57 +0000)]
net/mlx5e: Simplify get_route_and_out_devs helper function

The helper function has "if" branches that do the same. Merge them to
simplify the code.

Signed-off-by: Eli Britstein <elibr@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Fix unnecessary flow_block_cb_is_busy call
wenxu [Sat, 27 Jul 2019 14:59:55 +0000 (22:59 +0800)]
net/mlx5e: Fix unnecessary flow_block_cb_is_busy call

When call flow_block_cb_is_busy. The indr_priv is guaranteed to
NULL ptr. So there is no need to call flow_bock_cb_is_busy.

Fixes: 0d4fd02e7199 ("net: flow_offload: add flow_block_cb_is_busy() and use it")
Signed-off-by: wenxu <wenxu@ucloud.cn>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Improve ethtool rxnfc callback structure
Saeed Mahameed [Tue, 25 Jun 2019 18:17:18 +0000 (11:17 -0700)]
net/mlx5e: Improve ethtool rxnfc callback structure

Don't choose who implements the rxnfc "get/set" callbacks according to
CONFIG_MLX5_EN_RXNFC, instead have the callbacks always available and
delegate to a function of a different driver module when needed
(en_fs_ethtool.c), have stubs in en/fs.h to fallback to when
en_fs_ethtool.c is compiled out, to avoid complications and ifdefs in
en_main.c.

Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Avoid warning print when not required
Saeed Mahameed [Fri, 24 May 2019 19:24:43 +0000 (12:24 -0700)]
net/mlx5e: Avoid warning print when not required

When disabling CQE compression in favor of time-stamping, don't show a
warning when CQE compression is already disabled.

Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agonet/mlx5e: Print a warning when LRO feature is dropped or not allowed
Huy Nguyen [Thu, 27 Jun 2019 14:24:57 +0000 (09:24 -0500)]
net/mlx5e: Print a warning when LRO feature is dropped or not allowed

When user enables LRO via ethtool and if the RQ mode is legacy,
mlx5e_fix_features drops the request without any explanation.
Add netdev_warn to cover this case.

Fixes: 6c3a823e1e9c ("net/mlx5e: RX, Remove HW LRO support in legacy RQ")
Signed-off-by: Huy Nguyen <huyn@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
5 years agoMerge branch 'bnxt_en-TPA-57500'
David S. Miller [Mon, 29 Jul 2019 21:19:09 +0000 (14:19 -0700)]
Merge branch 'bnxt_en-TPA-57500'

Michael Chan says:

====================
bnxt_en: Add TPA (GRO_HW and LRO) on 57500 chips.

This patchset adds TPA v2 support on the 57500 chips.  TPA v2 is
different from the legacy TPA scheme on older chips and requires major
refactoring and restructuring of the existing TPA logic.  The main
difference is that the new TPA v2 has on-the-fly aggregation buffer
completions before a TPA packet is completed.  The larger aggregation
ID space also requires a new ID mapping logic to make it more
memory efficient.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agobnxt_en: Add PCI IDs for 57500 series NPAR devices.
Michael Chan [Mon, 29 Jul 2019 10:10:33 +0000 (06:10 -0400)]
bnxt_en: Add PCI IDs for 57500 series NPAR devices.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agobnxt_en: Support all variants of the 5750X chip family.
Michael Chan [Mon, 29 Jul 2019 10:10:32 +0000 (06:10 -0400)]
bnxt_en: Support all variants of the 5750X chip family.

Define the 57508, 57504, and 57502 chip IDs that are all part of the
BNXT_CHIP_P5 family of chips.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agobnxt_en: Refactor bnxt_init_one() and turn on TPA support on 57500 chips.
Michael Chan [Mon, 29 Jul 2019 10:10:31 +0000 (06:10 -0400)]
bnxt_en: Refactor bnxt_init_one() and turn on TPA support on 57500 chips.

With the new TPA feature in the 57500 chips, we need to discover the
feature first before setting up the netdev features.  Refactor the
the firmware probe and init logic more cleanly into 2 functions and
and make these calls before setting up the netdev features.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agobnxt_en: Support TPA counters on 57500 chips.
Michael Chan [Mon, 29 Jul 2019 10:10:30 +0000 (06:10 -0400)]
bnxt_en: Support TPA counters on 57500 chips.

Support the new expanded TPA v2 counters on 57500 B0 chips for
ethtool -S.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agobnxt_en: Allocate the larger per-ring statistics block for 57500 chips.
Michael Chan [Mon, 29 Jul 2019 10:10:29 +0000 (06:10 -0400)]
bnxt_en: Allocate the larger per-ring statistics block for 57500 chips.

The new TPA implemantation has additional TPA counters that extend the
per-ring statistics block.  Allocate the proper size accordingly.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agobnxt_en: Refactor ethtool ring statistics logic.
Michael Chan [Mon, 29 Jul 2019 10:10:28 +0000 (06:10 -0400)]
bnxt_en: Refactor ethtool ring statistics logic.

The current code assumes that the per ring statistics counters are
fixed.  In newer chips that support a newer version of TPA, the
TPA counters are also changed.  Refactor the code by defining these
counter names in arrays so that it is easy to add a new array for
a new set of counters supported by the newer chips.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agobnxt_en: Add hardware GRO setup function for 57500 chips.
Michael Chan [Mon, 29 Jul 2019 10:10:27 +0000 (06:10 -0400)]
bnxt_en: Add hardware GRO setup function for 57500 chips.

Add a more optimized hardware GRO function to setup the SKB on 57500
chips.  Some workaround code is no longer needed on 57500 chips and
the pseudo checksum is also calculated in hardware, so no need to
do the software pseudo checksum in the driver.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
5 years agobnxt_en: Add TPA ID mapping logic for 57500 chips.
Michael Chan [Mon, 29 Jul 2019 10:10:26 +0000 (06:10 -0400)]
bnxt_en: Add TPA ID mapping logic for 57500 chips.

The new TPA feature on 57500 supports a larger number of concurrent TPAs
(up to 1024) divided among the functions.  We need to add some logic to
map the hardware TPA ID to a software index that keeps track of each TPA
in progress.  A 1:1 direct mapping without translation would be too
wasteful as we would have to allocate 1024 TPA structures for each RX
ring on each PCI function.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>