Re: [PATCH v2 1/2] mwifiex: dispatch/rotate from reorder table atomically
From: Kalle Valo
Date: Thu Jun 27 2019 - 12:51:36 EST
Brian Norris <briannorris@xxxxxxxxxxxx> wrote:
> mwifiex_11n_scan_and_dispatch() and
> mwifiex_11n_dispatch_pkt_until_start_win() share similar patterns, where
> they perform a few different actions on the same table, using the same
> lock, but non-atomically. There have been other attempts to clean up
> this sort of behavior, but they have had problems (incomplete;
> introducing new deadlocks).
> We can improve these functions' atomicity by queueing up our RX packets
> in a list, to dispatch at the end of the function. This avoids problems
> of another operation modifying the table in between our dispatch and
> rotation operations.
> This was inspired by investigations around this:
> Subject: [4.20 PATCH] Revert "mwifiex: restructure rx_reorder_tbl_lock usage"
> While the original (now-reverted) patch had good intentions in
> restructuring some of the locking patterns in this driver, it missed an
> important detail: we cannot defer to softirq contexts while already in
> an atomic context. We can help avoid this sort of problem by separating
> the two steps of:
> (1) iterating / clearing the mwifiex reordering table
> (2) dispatching received packets to upper layers
> This makes it much harder to make lock recursion mistakes, as these
> two steps no longer need to hold the same locks.
> Testing: I've played with a variety of stress tests, including download
> stress tests on the same APs which caught regressions with commit
> 5188d5453bc9 ("mwifiex: restructure rx_reorder_tbl_lock usage"). I've
> primarily tested on Marvell 8997 / PCIe, although I've given 8897 / SDIO
> a quick spin as well.
> Signed-off-by: Brian Norris <briannorris@xxxxxxxxxxxx>
> Acked-by: Ganapathi Bhat <gbhat@xxxxxxxxxxx>
2 patches applied to wireless-drivers-next.git, thanks.
ce2e942e32e8 mwifiex: dispatch/rotate from reorder table atomically
8a7f9fd8a3e0 mwifiex: don't disable hardirqs; just softirqs