net: macb: fix sleep inside spinlock
authorSascha Hauer <s.hauer@pengutronix.de>
Fri, 8 Sep 2023 11:29:13 +0000 (13:29 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Tue, 19 Sep 2023 10:28:10 +0000 (12:28 +0200)
commitd5d315cf7687dd9b8dcf9da09d9145f86b30fe2a
tree1762926fb0f27668a9281440988af0b9a44f3e47
parent7aa720c350c7108179a37c3cec0b80087ab7479c
net: macb: fix sleep inside spinlock

[ Upstream commit 403f0e771457e2b8811dc280719d11b9bacf10f4 ]

macb_set_tx_clk() is called under a spinlock but itself calls clk_set_rate()
which can sleep. This results in:

| BUG: sleeping function called from invalid context at kernel/locking/mutex.c:580
| pps pps1: new PPS source ptp1
| in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 40, name: kworker/u4:3
| preempt_count: 1, expected: 0
| RCU nest depth: 0, expected: 0
| 4 locks held by kworker/u4:3/40:
|  #0: ffff000003409148
| macb ff0c0000.ethernet: gem-ptp-timer ptp clock registered.
|  ((wq_completion)events_power_efficient){+.+.}-{0:0}, at: process_one_work+0x14c/0x51c
|  #1: ffff8000833cbdd8 ((work_completion)(&pl->resolve)){+.+.}-{0:0}, at: process_one_work+0x14c/0x51c
|  #2: ffff000004f01578 (&pl->state_mutex){+.+.}-{4:4}, at: phylink_resolve+0x44/0x4e8
|  #3: ffff000004f06f50 (&bp->lock){....}-{3:3}, at: macb_mac_link_up+0x40/0x2ac
| irq event stamp: 113998
| hardirqs last  enabled at (113997): [<ffff800080e8503c>] _raw_spin_unlock_irq+0x30/0x64
| hardirqs last disabled at (113998): [<ffff800080e84478>] _raw_spin_lock_irqsave+0xac/0xc8
| softirqs last  enabled at (113608): [<ffff800080010630>] __do_softirq+0x430/0x4e4
| softirqs last disabled at (113597): [<ffff80008001614c>] ____do_softirq+0x10/0x1c
| CPU: 0 PID: 40 Comm: kworker/u4:3 Not tainted 6.5.0-11717-g9355ce8b2f50-dirty #368
| Hardware name: ... ZynqMP ... (DT)
| Workqueue: events_power_efficient phylink_resolve
| Call trace:
|  dump_backtrace+0x98/0xf0
|  show_stack+0x18/0x24
|  dump_stack_lvl+0x60/0xac
|  dump_stack+0x18/0x24
|  __might_resched+0x144/0x24c
|  __might_sleep+0x48/0x98
|  __mutex_lock+0x58/0x7b0
|  mutex_lock_nested+0x24/0x30
|  clk_prepare_lock+0x4c/0xa8
|  clk_set_rate+0x24/0x8c
|  macb_mac_link_up+0x25c/0x2ac
|  phylink_resolve+0x178/0x4e8
|  process_one_work+0x1ec/0x51c
|  worker_thread+0x1ec/0x3e4
|  kthread+0x120/0x124
|  ret_from_fork+0x10/0x20

The obvious fix is to move the call to macb_set_tx_clk() out of the
protected area. This seems safe as rx and tx are both disabled anyway at
this point.
It is however not entirely clear what the spinlock shall protect. It
could be the read-modify-write access to the NCFGR register, but this
is accessed in macb_set_rx_mode() and macb_set_rxcsum_feature() as well
without holding the spinlock. It could also be the register accesses
done in mog_init_rings() or macb_init_buffers(), but again these
functions are called without holding the spinlock in macb_hresp_error_task().
The locking seems fishy in this driver and it might deserve another look
before this patch is applied.

Fixes: 633e98a711ac0 ("net: macb: use resolved link config in mac_link_up()")
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
Link: https://lore.kernel.org/r/20230908112913.1701766-1-s.hauer@pengutronix.de
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
drivers/net/ethernet/cadence/macb_main.c