On 2020-08-14 10:37:00 [-0700], Davidlohr Bueso wrote: > ipcomp_compress() will serialize access to the ipcomp_scratches percpu buffer by > disabling BH and preventing a softirq from coming in and running ipcom_decompress(), > which is never called from process context. This of course won't work on RT and > the buffer can get corrupted; there have been similar issues with in the past with > such assumptions, ie: ebf255ed6c44 (net: add back the missing serialization in > ip_send_unicast_reply()). > > Similarly, this patch addresses the issue with locallocks allowing RT to have a > percpu spinlock and do the correct serialization. > > Addressing such races on an individual basis seemed like a game of whack a mole, > until afaict local_bh_enable() was reworked to use locallocks in 96fac673174 > (softirq: Add preemptible softirq) which is why the BUG is not seen in newer kernels. So what you are saying that this patch isn't needed v5.0.19-rt10+ because local_bh_disable() acts there as a lock. Repost is then with Tom in Cc because he will be the first one to apply to the v4.19 tree. And please shrink the patch description to a sane amount of information. For upstream, could you please get rid of get_cpu() in ipcomp_decompress()? If that function is always called with BH disabled then something like this_cpu_ptr() would be enough. Also I highly discourage code where half of the stuff is done in the part of the function where you define local variables. Removing get_cpu() would be nice because some implementation of crypto_comp_decompress() acquire sleeping locks. That would be: - nx842_crypto_decompress() - zip_comp_decompress() - nx842_crypto_decompress() > Signed-off-by: Davidlohr Bueso <dbueso@xxxxxxx> > --- > > This applies on top of all v4.x based kernels -rt. Sebastian
