Through some experiments, we found out that increasing the default
RX buffers count from 512 to 1024, gives slightly better throughput
and significantly reduces the no_wqe_rx errs on the receiver side.
Along with these, other parameters like cpu usage, retrans seg etc
also show some improvement with 1024 value.
Following are some snippets from the experiments
ntttcp tests with 512 Rx buffers
---------------------------------------
connections| throughput| no_wqe errs|
---------------------------------------
1 | 40.93Gbps | 123,211 |
16 | 180.15Gbps | 190,120 |
128 | 180.20Gbps | 173,508 |
256 | 180.27Gbps | 189,884 |
ntttcp tests with 1024 Rx buffers
---------------------------------------
connections| throughput| no_wqe errs|
---------------------------------------
1 | 44.22Gbps | 19,864 |
16 | 180.19Gbps | 4,430 |
128 | 180.21Gbps | 2,560 |
256 | 180.29Gbps | 1,529 |
So, increasing the default RX buffers per queue count to 1024
Signed-off-by: Shradha Gupta <shradhagupta@linux.microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: Pavan Chebbi <pavan.chebbi@broadcom.com>
Link: https://patch.msgid.link/1727667875-29908-1-git-send-email-shradhagupta@linux.microsoft.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Currently napi_disable() gets called during rxq and txq cleanup,
even before napi is enabled and hrtimer is initialized. It causes
kernel panic.
? page_fault_oops+0x136/0x2b0
? page_counter_cancel+0x2e/0x80
? do_user_addr_fault+0x2f2/0x640
? refill_obj_stock+0xc4/0x110
? exc_page_fault+0x71/0x160
? asm_exc_page_fault+0x27/0x30
? __mmdrop+0x10/0x180
? __mmdrop+0xec/0x180
? hrtimer_active+0xd/0x50
hrtimer_try_to_cancel+0x2c/0xf0
hrtimer_cancel+0x15/0x30
napi_disable+0x65/0x90
mana_destroy_rxq+0x4c/0x2f0
mana_create_rxq.isra.0+0x56c/0x6d0
? mana_uncfg_vport+0x50/0x50
mana_alloc_queues+0x21b/0x320
? skb_dequeue+0x5f/0x80
Cc: stable@vger.kernel.org
Fixes: e1b5683ff6 ("net: mana: Move NAPI from EQ to CQ")
Signed-off-by: Souradeep Chakrabarti <schakrabarti@linux.microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: Shradha Gupta <shradhagupta@linux.microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
After napi_complete_done() is called when NAPI is polling in the current
process context, another NAPI may be scheduled and start running in
softirq on another CPU and may ring the doorbell before the current CPU
does. When combined with unnecessary rings when there is no need to arm
the CQ, it triggers error paths in the hardware.
This patch fixes this by calling napi_complete_done() after doorbell
rings. It limits the number of unnecessary rings when there is
no need to arm. MANA hardware specifies that there must be one doorbell
ring every 8 CQ wraparounds. This driver guarantees one doorbell ring as
soon as the number of consumed CQEs exceeds 4 CQ wraparounds. In practical
workloads, the 4 CQ wraparounds proves to be big enough that it rarely
exceeds this limit before all the napi weight is consumed.
To implement this, add a per-CQ counter cq->work_done_since_doorbell,
and make sure the CQ is armed as soon as passing 4 wraparounds of the CQ.
Cc: stable@vger.kernel.org
Fixes: e1b5683ff6 ("net: mana: Move NAPI from EQ to CQ")
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Long Li <longli@microsoft.com>
Link: https://patch.msgid.link/1723219138-29887-1-git-send-email-longli@linuxonhyperv.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Pull rdma updates from Jason Gunthorpe:
"Usual collection of small improvements and fixes:
- Bug fixes and minor improvments in efa, irdma, mlx4, mlx5, rxe,
hf1, qib, ocrdma
- bnxt_re support for MSN, which is a new retransmit logic
- Initial mana support for RC qps
- Use after free bug and cleanups in iwcm
- Reduce resource usage in mlx5 when RDMA verbs features are not used
- New verb to drain shared recieve queues, similar to normal recieve
queues. This is necessary to allow ULPs a clean shutdown. Used in
the iscsi rdma target
- mlx5 support for more than 16 bits of doorbell indexes
- Doorbell moderation support for bnxt_re
- IB multi-plane support for mlx5
- New EFA adaptor PCI IDs
- RDMA_NAME_ASSIGN_TYPE_USER to hint to userspace that it shouldn't
rename the device
- A collection of hns bugs
- Fix long standing bug in bnxt_re with incorrect endian handling of
immediate data"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (65 commits)
IB/hfi1: Constify struct flag_table
RDMA/mana_ib: Set correct device into ib
bnxt_re: Fix imm_data endianness
RDMA: Fix netdev tracker in ib_device_set_netdev
RDMA/hns: Fix mbx timing out before CMD execution is completed
RDMA/hns: Fix insufficient extend DB for VFs.
RDMA/hns: Fix undifined behavior caused by invalid max_sge
RDMA/hns: Fix shift-out-bounds when max_inline_data is 0
RDMA/hns: Fix missing pagesize and alignment check in FRMR
RDMA/hns: Fix unmatch exception handling when init eq table fails
RDMA/hns: Fix soft lockup under heavy CEQE load
RDMA/hns: Check atomic wr length
RDMA/ocrdma: Don't inline statistics functions
RDMA/core: Introduce "name_assign_type" for an IB device
RDMA/qib: Fix truncation compilation warnings in qib_verbs.c
RDMA/qib: Fix truncation compilation warnings in qib_init.c
RDMA/efa: Add EFA 0xefa3 PCI ID
RDMA/mlx5: Support per-plane port IB counters by querying PPCNT register
net/mlx5: mlx5_ifc update for accessing ppcnt register of plane ports
RDMA/mlx5: Add plane index support when querying PTYS registers
...
As defined by the MANA Hardware spec, the queue size for DMA is 4KB
minimal, and power of 2. And, the HWC queue size has to be exactly
4KB.
To support page sizes other than 4KB on ARM64, define the minimal
queue size as a macro separately from the PAGE_SIZE, which we always
assumed it to be 4KB before supporting ARM64.
Also, add MANA specific macros and update code related to size
alignment, DMA region calculations, etc.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: Michael Kelley <mhklinux@outlook.com>
Link: https://lore.kernel.org/r/1718655446-6576-1-git-send-email-haiyangz@microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Erick Archer says:
====================
mana: Add flex array to struct mana_cfg_rx_steer_req_v2 (part)
The "struct mana_cfg_rx_steer_req_v2" uses a dynamically sized set of
trailing elements. Specifically, it uses a "mana_handle_t" array. So,
use the preferred way in the kernel declaring a flexible array [1].
At the same time, prepare for the coming implementation by GCC and Clang
of the __counted_by attribute. Flexible array members annotated with
__counted_by can have their accesses bounds-checked at run-time via
CONFIG_UBSAN_BOUNDS (for array indexing) and CONFIG_FORTIFY_SOURCE (for
strcpy/memcpy-family functions).
Also, avoid the open-coded arithmetic in the memory allocator functions
[2] using the "struct_size" macro.
Moreover, use the "offsetof" helper to get the indirect table offset
instead of the "sizeof" operator and avoid the open-coded arithmetic in
pointers using the new flex member. This new structure member also allow
us to remove the "req_indir_tab" variable since it is no longer needed.
Now, it is also possible to use the "flex_array_size" helper to compute
the size of these trailing elements in the "memcpy" function.
Specifically, the first commit adds the flex member and the patches 2 and
3 refactor the consumers of the "struct mana_cfg_rx_steer_req_v2".
This code was detected with the help of Coccinelle, and audited and
modified manually. The Coccinelle script used to detect this code pattern
is the following:
virtual report
@rule1@
type t1;
type t2;
identifier i0;
identifier i1;
identifier i2;
identifier ALLOC =~ "kmalloc|kzalloc|kmalloc_node|kzalloc_node|vmalloc|vzalloc|kvmalloc|kvzalloc";
position p1;
@@
i0 = sizeof(t1) + sizeof(t2) * i1;
...
i2 = ALLOC@p1(..., i0, ...);
@script:python depends on report@
p1 << rule1.p1;
@@
msg = "WARNING: verify allocation on line %s" % (p1[0].line)
coccilib.report.print_report(p1[0],msg)
Link: https://www.kernel.org/doc/html/next/process/deprecated.html#zero-length-and-one-element-arrays [1]
Link: https://www.kernel.org/doc/html/next/process/deprecated.html#open-coded-arithmetic-in-allocator-arguments [2]
v1: https://lore.kernel.org/linux-hardening/AS8PR02MB7237974EF1B9BAFA618166C38B382@AS8PR02MB7237.eurprd02.prod.outlook.com/
v2: https://lore.kernel.org/linux-hardening/AS8PR02MB723729C5A63F24C312FC9CD18B3F2@AS8PR02MB7237.eurprd02.prod.outlook.com/
====================
Link: https://lore.kernel.org/r/AS8PR02MB72374BD1B23728F2E3C3B1A18B022@AS8PR02MB7237.eurprd02.prod.outlook.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
mana_get_rxbuf_cfg() aligns the RX buffer's DMA datasize to be
multiple of 64. So a packet slightly bigger than mtu+14, say 1536,
can be received and cause skb_over_panic.
Sample dmesg:
[ 5325.237162] skbuff: skb_over_panic: text:ffffffffc043277a len:1536 put:1536 head:ff1100018b517000 data:ff1100018b517100 tail:0x700 end:0x6ea dev:<NULL>
[ 5325.243689] ------------[ cut here ]------------
[ 5325.245748] kernel BUG at net/core/skbuff.c:192!
[ 5325.247838] invalid opcode: 0000 [#1] PREEMPT SMP NOPTI
[ 5325.258374] RIP: 0010:skb_panic+0x4f/0x60
[ 5325.302941] Call Trace:
[ 5325.304389] <IRQ>
[ 5325.315794] ? skb_panic+0x4f/0x60
[ 5325.317457] ? asm_exc_invalid_op+0x1f/0x30
[ 5325.319490] ? skb_panic+0x4f/0x60
[ 5325.321161] skb_put+0x4e/0x50
[ 5325.322670] mana_poll+0x6fa/0xb50 [mana]
[ 5325.324578] __napi_poll+0x33/0x1e0
[ 5325.326328] net_rx_action+0x12e/0x280
As discussed internally, this alignment is not necessary. To fix
this bug, remove it from the code. So oversized packets will be
marked as CQE_RX_TRUNCATED by NIC, and dropped.
Cc: stable@vger.kernel.org
Fixes: 2fbbd712ba ("net: mana: Enable RX path to handle various MTU sizes")
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: Dexuan Cui <decui@microsoft.com>
Link: https://lore.kernel.org/r/1712087316-20886-1-git-send-email-haiyangz@microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Pull rdma updates from Jason Gunthorpe:
"Small cycle, with some typical driver updates:
- General code tidying in siw, hfi1, idrdma, usnic, hns rtrs and
bnxt_re
- Many small siw cleanups without an overeaching theme
- Debugfs stats for hns
- Fix a TX queue timeout in IPoIB and missed locking of the mcast
list
- Support more features of P7 devices in bnxt_re including a new work
submission protocol
- CQ interrupts for MANA
- netlink stats for erdma
- EFA multipath PCI support
- Fix Incorrect MR invalidation in iser"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (66 commits)
RDMA/bnxt_re: Fix error code in bnxt_re_create_cq()
RDMA/efa: Add EFA query MR support
IB/iser: Prevent invalidating wrong MR
RDMA/erdma: Add hardware statistics support
RDMA/erdma: Introduce dma pool for hardware responses of CMDQ requests
IB/iser: iscsi_iser.h: fix kernel-doc warning and spellos
RDMA/mana_ib: Add CQ interrupt support for RAW QP
RDMA/mana_ib: query device capabilities
RDMA/mana_ib: register RDMA device with GDMA
RDMA/bnxt_re: Fix the sparse warnings
RDMA/bnxt_re: Fix the offset for GenP7 adapters for user applications
RDMA/bnxt_re: Share a page to expose per CQ info with userspace
RDMA/bnxt_re: Add UAPI to share a page with user space
IB/ipoib: Fix mcast list locking
RDMA/mlx5: Expose register c0 for RDMA device
net/mlx5: E-Switch, expose eswitch manager vport
net/mlx5: Manage ICM type of SW encap
RDMA/mlx5: Support handling of SW encap ICM area
net/mlx5: Introduce indirect-sw-encap ICM properties
RDMA/bnxt_re: Adds MSN table capability for Gen P7 adapters
...
This patch allows to assign and poll more than one EQ on the same
msix index.
It is achieved by introducing a list of attached EQs in each IRQ context.
It also removes the existing msix_index map that tried to ensure that there
is only one EQ at each msix_index.
This patch exports symbols for creating EQs from other MANA kernel modules.
Signed-off-by: Konstantin Taranov <kotaranov@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Handle the case when GSO SKB linear length is too large.
MANA NIC requires GSO packets to put only the header part to SGE0,
otherwise the TX queue may stop at the HW level.
So, use 2 SGEs for the skb linear part which contains more than the
packet header.
Fixes: ca9c54d2d6 ("net: mana: Add a driver for Microsoft Azure Network Adapter (MANA)")
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Shradha Gupta <shradhagupta@linux.microsoft.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
