Messages in this thread | | | Subject | Re: [PATCH net-next 4/6] net: ipa: ensure hardware has power in ipa_start_xmit() | From | Alex Elder <> | Date | Mon, 16 Aug 2021 12:56:40 -0500 |
| |
On 8/16/21 9:20 AM, Alex Elder wrote: > On 8/16/21 9:15 AM, Jakub Kicinski wrote: >> On Fri, 13 Aug 2021 21:25:23 -0500 Alex Elder wrote: >>>> This is racy, what if the pm work gets scheduled on another CPU and >>>> calls wake right here (i.e. before you call netif_stop_queue())? >>>> The queue may never get woken up? >>> >>> I haven't been seeing this happen but I think you may be right. >>> >>> I did think about this race, but I think I was relying on the >>> PM work queue to somehow avoid the problem. I need to think >>> about this again after a good night's sleep. I might need >>> to add an atomic flag or something. >> >> Maybe add a spin lock? Seems like the whole wake up path will be >> expensive enough for a spin lock to be in the noise. You can always >> add complexity later. > > Exactly what I just decided after trying to work out a > clever way without using a spinlock... I'll be sending > out a fix today. Thanks.
I'm finding this isn't an easy problem to solve (or even think about). While I ponder the best course of action I'm going to send out another series (i.e., *before* I send a fix for this issue) because I'd like to get everything I have out for review this week. I *will* address this potential race one way or another, possibly later today.
-Alex
> > -Alex >
| |