Yes. Every time it turns out that you really need irqsave locks for rxe it is because some maniac is calling into the verbs API at interrupt level. I have run into this a couple of times. Bob -----Original Message----- From: Jason Gunthorpe <jgg@xxxxxxxxxx> Sent: Friday, October 28, 2022 12:33 PM To: Bob Pearson <rpearsonhpe@xxxxxxxxx> Cc: zyjzyj2000@xxxxxxxxx; linux-rdma@xxxxxxxxxxxxxxx Subject: Re: [PATCH for-next] RDMA/rxe: Convert spin_lock_irqsave to spin_lock On Thu, Oct 06, 2022 at 12:04:55PM -0500, Bob Pearson wrote: > Currently the rxe driver uses a spin_lock_irqsave call to protect the > producer end of a completion queue from multiple QPs which may be > sharing the cq. Since all accesses to the cq are from tasklets this is > overkill and a simple spin_lock will be sufficient. A tasklet is a softirq, which means it needs to use the BH suffix for all acquires that are in a process context if it the spinlock is obtained from a tasklet There is a nice table in Documentation/kernel-hacking/locking.rst that explains this. Jason
