Re: [PATCH net v1] mlxbf_gige: open() should call request_irq() after NAPI init

From: Paolo Abeni
Date: Tue Mar 19 2024 - 10:32:35 EST


On Fri, 2024-03-15 at 10:56 -0400, David Thompson wrote:
> This patch fixes an exception that occurs during open()
> when kdump is enabled. The sequence to reproduce the
> exception is as follows:
> a) enable kdump
> b) trigger kdump via "echo c > /proc/sysrq-trigger"
> c) kdump kernel executes
> d) kdump kernel loads mlxbf_gige module
> e) the mlxbf_gige module runs its open() as the
> the "oob_net0" interface is brought up
> f) mlxbf_gige module will experience an exception
> during its open(), something like:
>
> Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
> Mem abort info:
> ESR = 0x0000000086000004
> EC = 0x21: IABT (current EL), IL = 32 bits
> SET = 0, FnV = 0
> EA = 0, S1PTW = 0
> FSC = 0x04: level 0 translation fault
> user pgtable: 4k pages, 48-bit VAs, pgdp=00000000e29a4000
> [0000000000000000] pgd=0000000000000000, p4d=0000000000000000
> Internal error: Oops: 0000000086000004 [#1] SMP
> CPU: 0 PID: 812 Comm: NetworkManager Tainted: G OE 5.15.0-1035-bluefield #37-Ubuntu
> Hardware name: https://www.mellanox.com BlueField-3 SmartNIC Main Card/BlueField-3 SmartNIC Main Card, BIOS 4.6.0.13024 Jan 19 2024
> pstate: 80400009 (Nzcv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
> pc : 0x0
> lr : __napi_poll+0x40/0x230
> sp : ffff800008003e00
> x29: ffff800008003e00 x28: 0000000000000000 x27: 00000000ffffffff
> x26: ffff000066027238 x25: ffff00007cedec00 x24: ffff800008003ec8
> x23: 000000000000012c x22: ffff800008003eb7 x21: 0000000000000000
> x20: 0000000000000001 x19: ffff000066027238 x18: 0000000000000000
> x17: ffff578fcb450000 x16: ffffa870b083c7c0 x15: 0000aaab010441d0
> x14: 0000000000000001 x13: 00726f7272655f65 x12: 6769675f6662786c
> x11: 0000000000000000 x10: 0000000000000000 x9 : ffffa870b0842398
> x8 : 0000000000000004 x7 : fe5a48b9069706ea x6 : 17fdb11fc84ae0d2
> x5 : d94a82549d594f35 x4 : 0000000000000000 x3 : 0000000000400100
> x2 : 0000000000000000 x1 : 0000000000000000 x0 : ffff000066027238
> Call trace:
> 0x0
> net_rx_action+0x178/0x360
> __do_softirq+0x15c/0x428
> __irq_exit_rcu+0xac/0xec
> irq_exit+0x18/0x2c
> handle_domain_irq+0x6c/0xa0
> gic_handle_irq+0xec/0x1b0
> call_on_irq_stack+0x20/0x2c
> do_interrupt_handler+0x5c/0x70
> el1_interrupt+0x30/0x50
> el1h_64_irq_handler+0x18/0x2c
> el1h_64_irq+0x7c/0x80
> __setup_irq+0x4c0/0x950
> request_threaded_irq+0xf4/0x1bc
> mlxbf_gige_request_irqs+0x68/0x110 [mlxbf_gige]
> mlxbf_gige_open+0x5c/0x170 [mlxbf_gige]
> __dev_open+0x100/0x220
> __dev_change_flags+0x16c/0x1f0
> dev_change_flags+0x2c/0x70
> do_setlink+0x220/0xa40
> __rtnl_newlink+0x56c/0x8a0
> rtnl_newlink+0x58/0x84
> rtnetlink_rcv_msg+0x138/0x3c4
> netlink_rcv_skb+0x64/0x130
> rtnetlink_rcv+0x20/0x30
> netlink_unicast+0x2ec/0x360
> netlink_sendmsg+0x278/0x490
> __sock_sendmsg+0x5c/0x6c
> ____sys_sendmsg+0x290/0x2d4
> ___sys_sendmsg+0x84/0xd0
> __sys_sendmsg+0x70/0xd0
> __arm64_sys_sendmsg+0x2c/0x40
> invoke_syscall+0x78/0x100
> el0_svc_common.constprop.0+0x54/0x184
> do_el0_svc+0x30/0xac
> el0_svc+0x48/0x160
> el0t_64_sync_handler+0xa4/0x12c
> el0t_64_sync+0x1a4/0x1a8
> Code: bad PC value
> ---[ end trace 7d1c3f3bf9d81885 ]---
> Kernel panic - not syncing: Oops: Fatal exception in interrupt
> Kernel Offset: 0x2870a7a00000 from 0xffff800008000000
> PHYS_OFFSET: 0x80000000
> CPU features: 0x0,000005c1,a3332a5a
> Memory Limit: none
> ---[ end Kernel panic - not syncing: Oops: Fatal exception in interrupt ]---
>
> The exception happens because there is a pending RX interrupt before the
> call to request_irq(RX IRQ) executes. Then, the RX IRQ handler fires
> immediately after this request_irq() completes. The RX IRQ handler runs
> "napi_schedule()" before NAPI is fully initialized via "netif_napi_add()"
> and "napi_enable()", both which happen later in the open() logic.
>
> This patch fixes the issue by re-ordering the logic in mlxbf_gige open()
> so that the request_irq() calls execute after NAPI is fully initialized.

It also adds a missing phy_stop(phydev) in the error path. It's worth
mentioning in the changelog or possibly place into a separate patch.

Cheers,

Paolo