if_bnxt: fix powerpc build of the MPC ring code

powerpc has no 16-bit acquire/release atomics, so the
atomic_load_acq_16() and atomic_store_rel_16() calls on the MPC TX
ring indices fail to build th

if_bnxt: fix powerpc build of the MPC ring code

powerpc has no 16-bit acquire/release atomics, so the
atomic_load_acq_16() and atomic_store_rel_16() calls on the MPC TX
ring indices fail to build there. Make the producer and consumer
indices 32 bits wide and use the 32-bit primitives.

Reviewed by: gallatin, chandrakanth.patil_broadcom.com
Sponsored by: Broadcom Inc
MFC after: 1 week
Fixes: 088a42c0ff76 ("if_bnxt: add MPC (Mid-Path Channel) ring infrastructure")
Differential Revision: https://reviews.freebsd.org/D60698

show more ...

bnxt_en: dcb: stop zeroing ETS and PFC config in bnxt_dcb_init()

bnxt_dcb_init() currently pushes all-zero ETS and PFC settings to
the firmware at every attach. This reserves bandwidth allocation
fo

bnxt_en: dcb: stop zeroing ETS and PFC config in bnxt_dcb_init()

bnxt_dcb_init() currently pushes all-zero ETS and PFC settings to
the firmware at every attach. This reserves bandwidth allocation
for RoCE traffic, which prevents L2 traffic from reaching line rate.

These settings should only be updated with valid values when the
RoCE driver is loaded. Therefore, we should stop programming them
during bnxt_dcb_init().

Signed-off-by: Andy Gospodarek <gospo@broadcom.com>
Reviewed by: chandrakanth.patil_broadcom.com, gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D60516

show more ...

if_bnxt: dcb: serialize HWRM sends and avoid freeing DMA bufs on timeout

Every _hwrm_send_message() call in bnxt_dcb.c ran without BNXT_HWRM_LOCK,
unlike the rest of the driver, letting concurrent H

if_bnxt: dcb: serialize HWRM sends and avoid freeing DMA bufs on timeout

Every _hwrm_send_message() call in bnxt_dcb.c ran without BNXT_HWRM_LOCK,
unlike the rest of the driver, letting concurrent HWRM commands race on
the shared MMIO doorbell and response buffer. Route single-shot sends
through the already-locked hwrm_send_message() wrapper, and wrap the
remaining send+response-read sequences in explicit BNXT_HWRM_LOCK/UNLOCK.

Also stop freeing the transient DMA buffers used for structured-data
get/set on ETIMEDOUT: a host-side timeout doesn't guarantee firmware
actually gave up, so a late completion could still DMA into memory
that's since been freed and reused. Leak the buffer instead in that case.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D60179

show more ...

if_bnxt: add support for SIOCGI2CPB page/bank i2c requests

Advertise IFLIB_I2C_PAGE_BANK and forward the requested page/bank to
firmware, so CMIS optics can be read correctly instead of falling
back

if_bnxt: add support for SIOCGI2CPB page/bank i2c requests

Advertise IFLIB_I2C_PAGE_BANK and forward the requested page/bank to
firmware, so CMIS optics can be read correctly instead of falling
back to bogus legacy SFF-8472 decoding. Also fix bnxt_i2c_req() to
return positive errno values.

Signed-off-by: gallatin
Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D59573

show more ...

if_bnxt: don't reserve RoCE MSI-x on chips that don't support RoCE

RoCE is only supported on Thor and onwards chips.
bnxt_init_sctx_variants() unconditionally set isc_admin_intrcnt to
BNXT_ROCE_IRQ_

if_bnxt: don't reserve RoCE MSI-x on chips that don't support RoCE

RoCE is only supported on Thor and onwards chips.
bnxt_init_sctx_variants() unconditionally set isc_admin_intrcnt to
BNXT_ROCE_IRQ_COUNT for every PF device, stealing 9 MSI-x vectors from
older, non-RoCE-capable chips that will never use them.

Gate the reservation on the device ID: RoCE-capable chips keep
BNXT_ROCE_IRQ_COUNT, other chips fall back to 1, the baseline
isc_admin_intrcnt value used before RoCE support (commit 050d28e13cde)
repurposed it for RoCE MSI-x on every device.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D59091

show more ...

if_bnxt: port endian, memory-barrier, and bitops kAPIs to FreeBSD natives

There are few more Linux kernel APIs left in the code. So,
replaced them with corresponding FreeBSD native APIs.

- Replace

if_bnxt: port endian, memory-barrier, and bitops kAPIs to FreeBSD natives

There are few more Linux kernel APIs left in the code. So,
replaced them with corresponding FreeBSD native APIs.

- Replace cpu_to_le16/32/64/le16_to_cpu/le32_to_cpu/le64_to_cpu with
htole16/32/64/le16toh/le32toh/le64toh throughout bnxt_hwrm.c and
the vf-event producer path in if_bnxt.c.
- Replace rmb()/wmb() interprocessor-ordering uses in bnxt_ulp.c
(publishing/consuming ulp->async_events_bmap and max_async_event_id
across CPUs) with atomic_thread_fence_acq()/atomic_thread_fence_rel();
keep wmb() for the HWRM doorbell/DMA ordering barrier in
_hwrm_send_message() (bnxt_hwrm.c), since atomic_thread_fence_rel()
compiles to a bare compiler barrier on amd64/i386 and does not order
the request-buffer store against the doorbell MMIO write.
- Replace hweight32 with bitcount32, ARRAY_SIZE with nitems, and
DECLARE_BITMAP with a plain unsigned long array (sized via
howmany()) in bnxt_hwrm_func_drv_rgtr(); replace the __set_bit/
test_bit calls there with bit_set()/bit_test() from
<sys/bitstring.h>, which correctly index the multi-word bitmap
(event IDs up to 0x45 span more than one 64-bit word).
- Replace the remaining DIV_ROUND_UP/HZ with howmany/hz in
bnxt_handle_async_event()'s error-recovery timer setup.
- Replace bnxt_sriov.h's <linux/pci.h> include with the native
<dev/pci/pcireg.h> + generated "pci_if.h", now that nothing in
that header needs Linux's struct pci_dev.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D59090

show more ...

if_bnxt: port workqueue, task-queue, and synchronization kPIs to FreeBSD natives

- Replace Linux workqueue (struct workqueue_struct, bnxt_pf_wq) with a
FreeBSD taskqueue (bnxt_taskq) shared by eve

if_bnxt: port workqueue, task-queue, and synchronization kPIs to FreeBSD natives

- Replace Linux workqueue (struct workqueue_struct, bnxt_pf_wq) with a
FreeBSD taskqueue (bnxt_taskq) shared by every attached PF. bnxt_taskq
is created once at module load and destroyed once at module unload
(SYSINIT/SYSUNINIT on SI_SUB_KLD), matching the module-owned-taskqueue
idiom used elsewhere in the tree (e.g. sys/dev/cxgbe/t4_main.c's
reset_tq). This replaces an earlier lazy, per-attach/detach
create-or-reuse scheme guarded by a mutex and a "ready" flag, which
had a real race: bnxt_queue_sp_work()/bnxt_queue_fw_reset_work() and
bnxt_attach_pre() re-read the bnxt_taskq/bnxt_taskq_ready globals
outside the mutex right after the lazy-init call returned, so a
concurrent bnxt_detach() on another PF freeing the shared queue once
bnxt_num_pfs hit 0 could race a still-attached PF into enqueuing onto
a freed taskqueue. Tying the taskqueue's lifetime to the module
instead of to any one PF's attach/detach removes that race outright,
since the module can't unload while a PF is still attached.

- Replace struct work_struct sp_task with struct task sp_task;
replace struct delayed_work fw_reset_task with
struct timeout_task fw_reset_tmo.

- Replace INIT_WORK/INIT_DELAYED_WORK with TASK_INIT/TIMEOUT_TASK_INIT;
queue_work/schedule_work with taskqueue_enqueue;
queue_delayed_work/schedule_delayed_work with taskqueue_enqueue_timeout;
cancel_work_sync/cancel_delayed_work_sync with
taskqueue_drain/taskqueue_drain_timeout.

- Rename bnxt_fw_reset_task(struct work_struct *) to
bnxt_fw_reset_timeout_task(void *ctx, int pending) to match FreeBSD
timeout_task callback signature.

- Add inline helpers in bnxt.h to wrap all Linux test_bit/set_bit/
clear_bit/test_and_clear_bit calls on softc->state and
softc->sp_event: bnxt_drv_state_test/set/clear/test_and_clear,
bnxt_sp_event_set/test_and_clear (implemented via bit_test() from
<sys/bitstring.h> and atomic_testandset_long()/
atomic_testandclear_long(), all bit-index based like the Linux
bitops they replace).

- Replace struct mutex lock in bnxt_fw_health with struct mtx;
mutex_init to mtx_init.

- Replace kzalloc/kfree with malloc/free (M_DEVBUF).

- Replace DIV_ROUND_UP with howmany; hweight32 with bitcount32.

- Replace jiffies/HZ with ticks/hz; msleep with pause_sbt(). Since
`ticks` (unlike LinuxKPI's `unsigned long jiffies`) is a 32-bit
`volatile int` that wraps every ~24 days at hz=1000, BNXT_TIME_AFTER
and bnxt_sleep_until_min_elapsed() compute the ticks delta as `int`
(matching the in-tree wraparound-safe idiom used elsewhere, e.g.
sys/dev/axgbe/xgbe-mdio.c) rather than widening it to `long`/
`unsigned long`, so the comparison stays correct across a wrap.

- Replace IS_ERR_OR_NULL/ERR_PTR with plain NULL checks.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D59089

show more ...

if_bnxt: sriov: replace Linux kPIs with FreeBSD native APIs

Convert bnxt_sriov.c and related SRIOV files from Linux to FreeBSD
native equivalents:

- Replace cpu_to_le16/32/64/le16_to_cpu with htole

if_bnxt: sriov: replace Linux kPIs with FreeBSD native APIs

Convert bnxt_sriov.c and related SRIOV files from Linux to FreeBSD
native equivalents:

- Replace cpu_to_le16/32/64/le16_to_cpu with htole16/32/64/le16toh.
- Replace Linux is_valid_ether_addr, ether_addr_equal, ether_addr_copy
with local bnxt_eth_addr_valid/bnxt_eth_addr_equal/bnxt_eth_addr_copy
helpers built on ETHER_IS_MULTICAST/ETHER_IS_ZERO.
- Replace kcalloc/kzalloc/kfree with malloc/free (M_DEVBUF,
M_WAITOK|M_ZERO).
- Replace dma_alloc_coherent/dma_free_coherent with iflib_dma_alloc/
iflib_dma_free; add struct iflib_dma_info hwrm_cmd_req_mem[4] to
bnxt_pf_info to hold the DMA handles.
- Replace the Linux unsigned long *vf_event_bmap with FreeBSD
bitstr_t *, allocated via bit_alloc(); bit_ffs_at()/bit_clear()/
bit_set() replace find_next_bit()/clear_bit()/set_bit() on it
(if_bnxt.c's vf-event producer side included).
- Replace DIV_ROUND_UP with howmany; rcu_assign_pointer with a plain
store.
- Harden bnxt_alloc_vf_resources() error paths to free
partially-allocated DMA pages and the VF array before returning
ENOMEM.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D59088

show more ...

if_bnxt: sysctl: remove unused linux/delay.h include and dead mutex

<linux/delay.h> and the DEFINE_MUTEX(tmp_mutex) it enabled are both
unused in bnxt_sysctl.c: no msleep/udelay/mdelay call remains,

if_bnxt: sysctl: remove unused linux/delay.h include and dead mutex

<linux/delay.h> and the DEFINE_MUTEX(tmp_mutex) it enabled are both
unused in bnxt_sysctl.c: no msleep/udelay/mdelay call remains, and
tmp_mutex itself was never referenced anywhere.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D59087

show more ...

if_bnxt: log: fix missing lock coverage in bnxt_log_live()

bnxt_log_live() walked the loggers TAILQ without holding log_lock,
unlike every other bnxt_log.c function, racing against concurrent
bnxt_r

if_bnxt: log: fix missing lock coverage in bnxt_log_live()

bnxt_log_live() walked the loggers TAILQ without holding log_lock,
unlike every other bnxt_log.c function, racing against concurrent
bnxt_register_logger()/bnxt_unregister_logger() calls. Take log_lock
around the traversal, and bail out early (with the lock dropped) if
live_msgs_len has already reached max_live_buff_size, which would
otherwise underflow the length passed to bnxt_log_info().

bnxt_start_logging_driver_coredump() now drops log_lock before
invoking logger->log_live_op() (which calls back into
bnxt_log_live()) and re-acquires it afterward, resetting live_msgs
to NULL so a later bnxt_log_live() call (e.g. from a VF async event
handler) can't write into a coredump buffer the caller has already
freed.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D59086

show more ...

if_bnxt: rework interrupt coalescing onto the AGGINT_QCAPS scheme

Query firmware-advertised coalescing capabilities and program
per-direction rx/tx coalescing settings against them, replacing the
ol

if_bnxt: rework interrupt coalescing onto the AGGINT_QCAPS scheme

Query firmware-advertised coalescing capabilities and program
per-direction rx/tx coalescing settings against them, replacing the
old hardcoded scheme. Add sysctls for the coalescing mode, budget,
and stats-timer knobs the new scheme exposes.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58621

show more ...

if_bnxt: drop epoch_arr[] to detect EPOCH bit toggle

The EPOCH array size is 4096 but the driver does not enforce any limit
on ring size, so setting ring size >4096 corrupts adjacent memory
and lead

if_bnxt: drop epoch_arr[] to detect EPOCH bit toggle

The EPOCH array size is 4096 but the driver does not enforce any limit
on ring size, so setting ring size >4096 corrupts adjacent memory
and leads to undefined behavior.

Toggle epoch_bit directly at every ring-producer wrap point (TX/RX
encap and refill, MPC crypto commands, kTLS presync/replay) instead of
looking it up per-index in the doorbell path, dropping the now-unused
epoch_arr[] snapshot array.

Reported by: gallatin
Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58620
Change-Id: I4673d64670ee0e8ea256b8a8932724bbd73f710d

show more ...

if_bnxt: drop redundant PCI disable in bnxt_fw_reset_close()

bnxt_fw_reset_close() unconditionally called pci_disable_device()
after freeing the Rx IRQs, on every firmware reset, not just the
BNXT_S

if_bnxt: drop redundant PCI disable in bnxt_fw_reset_close()

bnxt_fw_reset_close() unconditionally called pci_disable_device()
after freeing the Rx IRQs, on every firmware reset, not just the
BNXT_STATE_FW_FATAL_COND path which already disables the device via
bnxt_fw_fatal_close(). BNXT_FW_RESET_STATE_ENABLE_DEV in
bnxt_fw_reset_task() already re-enables the device unconditionally,
so this wasn't leaving the device disabled, just adding an
unnecessary PCI disable/re-enable cycle around every non-fatal reset.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58619

show more ...

if_bnxt: support more Tx rings than Rx rings on P5+

Add an unsupported-config guard rejecting nrxqsets > ntxqsets, and
teach the P5+ NQ datapath to handle the opposite asymmetric case:
extra Tx ring

if_bnxt: support more Tx rings than Rx rings on P5+

Add an unsupported-config guard rejecting nrxqsets > ntxqsets, and
teach the P5+ NQ datapath to handle the opposite asymmetric case:
extra Tx rings beyond the Rx ring count. softc->nq_rings[i].type now
tags each NQ as SHARED_NQ (paired with an Rx CQ) or BNXT_TX_ONLY_NQ
(dedicated to an extra Tx CQ with no backing Rx ring); bnxt_init()
and bnxt_hwrm_resource_free() gate all Rx ring group/ring/stat-ctx
alloc and free on IS_SHARED_NQ() so the extra Tx-only NQs don't touch
nonexistent Rx ring state. Gate the new type-tagging and IS_SHARED_NQ
checks on BNXT_CHIP_P5_PLUS(), since non-P5+ chips never allocate
softc->nq_rings and would otherwise access a nonexistent array.

BNXT_TX_ONLY_NQ is a distinct value from the existing MPC-private
TX_CP_NQ; the two are unrelated and kept apart to avoid confusion.

Suggested by: gallatin
Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58618

show more ...

if_bnxt: fix HWRM failures/timeouts after repeated FW resets

Consecutive firmware-initiated reset cycles produced HWRM
failures/timeouts and traffic didn't come back. Fix the FW-reset
recovery path:

if_bnxt: fix HWRM failures/timeouts after repeated FW resets

Consecutive firmware-initiated reset cycles produced HWRM
failures/timeouts and traffic didn't come back. Fix the FW-reset
recovery path:

- bnxt_fw_reset_close() called bnxt_stop() and
bnxt_hwrm_func_drv_unrgtr(), sending HWRM ring/VNIC/filter free
commands to a firmware that's already mid-reset and unresponsive,
producing the observed timeouts. Both are unnecessary since
firmware comes back with fresh state anyway; drop them along with
the now-redundant iflib_request_reset().
- bnxt_func_reset() now skips bnxt_hwrm_resource_free() entirely
while BNXT_STATE_IN_FW_RESET is set, for the same reason.
- bnxt_open() (used to reopen after a firmware reset) now issues
bnxt_hwrm_func_reset() up front and drives reinit through iflib's
own reset machinery (iflib_request_reset() plus a new
bnxt_iflib_reset_sync() that waits for
IFF_DRV_RUNNING/OACTIVE to flip 1->0->1) instead of calling
bnxt_init()/bnxt_intr_enable() directly.
- Route repeated PORT_QSTATS/PORT_QSTATS_EXT HWRM timeouts through
bnxt_log_live() instead of device_printf(), so a stats poll timing
out during a reset doesn't spam dmesg on every tick.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58617

show more ...

if_bnxt: Never ARM Tx CQ as they are not interrupt-driven

Tx completion rings are not interrupt-driven, so there is no need to
ring the doorbell with CQ_ARMALL and toggle bits set for them. Stop
arm

if_bnxt: Never ARM Tx CQ as they are not interrupt-driven

Tx completion rings are not interrupt-driven, so there is no need to
ring the doorbell with CQ_ARMALL and toggle bits set for them. Stop
arming the Tx CQ doorbell on P5+, which makes tx_cp_rings[]->toggle
dead, so drop the now-unnecessary toggle bookkeeping for it in
process_nq() too.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58616

show more ...

if_bnxt: avoid FTQM/STQM pg_info alias on reset

Copying the STQM backing-store context to seed FTQM's also copied
STQM's pg_info pointer verbatim. On cold load STQM's pg_info is still
NULL so this i

if_bnxt: avoid FTQM/STQM pg_info alias on reset

Copying the STQM backing-store context to seed FTQM's also copied
STQM's pg_info pointer verbatim. On cold load STQM's pg_info is still
NULL so this is harmless, but on a firmware reset STQM's pg_info is
already non-NULL, making FTQM alias STQM's backing-store pages. The
allocator then skips FTQM since it looks already allocated, and FTQM
is later indexed as its own array, reading out of bounds into STQM's
buffer and dereferencing a bogus DMA address.

Clear ctxm->pg_info after the memcpy so FTQM always gets its own
backing store.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58615

show more ...

if_bnxt: add support for HW generic stats

HW generic stats provide information about PCIe transactions and
out-of-order doorbell writes, useful for debug. Allocate a DMA buffer
for them (disabling B

if_bnxt: add support for HW generic stats

HW generic stats provide information about PCIe transactions and
out-of-order doorbell writes, useful for debug. Allocate a DMA buffer
for them (disabling BNXT_FW_CAP_GENERIC_STATS if the allocation
fails), fetch them via HWRM_STAT_GENERIC_QSTATS once per admin-status
tick on P5+ chips, and expose them to userspace under
dev.bnxt.<unit>.hwstats.generic_stats.* via sysctl.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58614

show more ...

if_bnxt: Restore Rx CQ doorbell skip-when-idle optimization on P5

The unconditional-doorbell fix for Whp/older NICs (always ring the Rx
CQ doorbell after processing Rx interrupts) had also been appl

if_bnxt: Restore Rx CQ doorbell skip-when-idle optimization on P5

The unconditional-doorbell fix for Whp/older NICs (always ring the Rx
CQ doorbell after processing Rx interrupts) had also been applied to
P5+ chips, discarding process_nq()'s work-done signal entirely and
losing the doorbell-skip optimization on P5+ when process_nq() finds
no CQ notification to act on.

Make process_nq() return whether it saw a CQ notification, and only
skip the Rx CQ doorbell on P5+ when it didn't; non-P5+ chips continue
to always ring it, since process_nq() isn't used on that path.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58613

show more ...

if_bnxt: ktls: Reject new kTLS sessions once driver is detaching

bnxt_tls_snd_tag_alloc() only gated on BNXT_STATE_OPEN, which is set
once by bnxt_open()/bnxt_attach_pre() but never cleared again, s

if_bnxt: ktls: Reject new kTLS sessions once driver is detaching

bnxt_tls_snd_tag_alloc() only gated on BNXT_STATE_OPEN, which is set
once by bnxt_open()/bnxt_attach_pre() but never cleared again, so it
could not reject new sessions once bnxt_detach() started tearing
things down. The snd_tag_alloc_ref-based wait in bnxt_detach(), which
waits for in-flight allocators to exit, was already in place, but
without this gate new sessions could keep being created for the
entire duration of that wait, extending it indefinitely instead of
just draining existing ones.

Reject new sessions once softc->detached is set, and reset it back to
false in bnxt_attach_pre() so it doesn't wrongly persist across a
detach/reattach cycle.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58612

show more ...

bnxt_re: Fix witness reported calltrace while unloading driver

bnxt_unregister_dev() called synchronize_rcu() while holding
bp->en_ops_lock, which witness flags as a sleeping-while-locked
violation

bnxt_re: Fix witness reported calltrace while unloading driver

bnxt_unregister_dev() called synchronize_rcu() while holding
bp->en_ops_lock, which witness flags as a sleeping-while-locked
violation on driver unload. Drop the lock around the RCU grace period
and reacquire it afterward.

Also gate the RoCE extended-stats hex dump in
bnxt_qplib_qext_stat() behind bootverbose instead of printing it
unconditionally.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58611

show more ...

if_bnxt: Collect hw stats every 250 milliseconds

bnxt_if_timer() only scheduled bnxt_update_admin_status(), which
polls HW stats, once per second. Drop the interval to 250ms so stats
collection is t

if_bnxt: Collect hw stats every 250 milliseconds

bnxt_if_timer() only scheduled bnxt_update_admin_status(), which
polls HW stats, once per second. Drop the interval to 250ms so stats
collection is timely.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58610

show more ...

if_bnxt: Pass HWRM cmd response to apps even if FW fails it

The driver was not passing the HWRM command response to management
apps when the firmware failed the command, which is not what apps
expec

if_bnxt: Pass HWRM cmd response to apps even if FW fails it

The driver was not passing the HWRM command response to management
apps when the firmware failed the command, which is not what apps
expect. Copy the response out whenever firmware populated one
(resp_len != 0), regardless of the command's overall return code, and
stop bailing out of the mgmt ioctl path early on a failed passthrough
so DMA'd indirect data and the response still get copied to
userspace.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58609

show more ...

if_bnxt: honor firmware min wait before VER_GET on reset

The firmware advertises a minimum wait time it needs before it is
ready. The driver was not honoring it and sent VER_GET too early
during fir

if_bnxt: honor firmware min wait before VER_GET on reset

The firmware advertises a minimum wait time it needs before it is
ready. The driver was not honoring it and sent VER_GET too early
during firmware error-recovery, which could return an incomplete
VER_GET buffer. Enforce that minimum wait, measured from when the
RESET_NOTIFY async event arrived, before sending VER_GET.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58608

show more ...

if_bnxt: Fix HWRM mailbox/DMA teardown race on detach

iflib's generic device-deregister path never drains the admin task
before calling IFDI_DETACH, so bnxt_update_admin_status() (scheduled
once/sec

if_bnxt: Fix HWRM mailbox/DMA teardown race on detach

iflib's generic device-deregister path never drains the admin task
before calling IFDI_DETACH, so bnxt_update_admin_status() (scheduled
once/sec) could still run concurrently with bnxt_detach(), racing on
softc->hwrm_lock and the shared HWRM request/response DMA buffer that
bnxt_detach() destroys. That race could leave the firmware-side HWRM
mailbox inconsistent, surfacing as "Timeout sending HWRM_VER_GET" on
the next module load.

Add a softc->detached guard: set it as the first statement in
bnxt_detach(), and bail out of bnxt_update_admin_status() immediately
when it's set.

Reviewed by: gallatin
MFC after: 2 weeks
Sponsored by: Broadcom Inc.
Differential Revision: https://reviews.freebsd.org/D58607

show more ...