On Fri 16 Aug 2019 at 20:56, Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx> wrote: > On Fri, 16 Aug 2019 15:04:44 +0000, Vlad Buslov wrote: >> >> [ 401.511871] RSP: 002b:00007ffca2a9fad8 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 >> >> [ 401.511875] RAX: ffffffffffffffda RBX: 0000000000000002 RCX: 00007fad892d30f8 >> >> [ 401.511878] RDX: 0000000000000002 RSI: 000055afeb072a90 RDI: 0000000000000001 >> >> [ 401.511881] RBP: 000055afeb072a90 R08: 00000000ffffffff R09: 000000000000000a >> >> [ 401.511884] R10: 000055afeb058710 R11: 0000000000000246 R12: 0000000000000002 >> >> [ 401.511887] R13: 00007fad893a8780 R14: 0000000000000002 R15: 00007fad893a3740 >> >> >> >> I don't think it is correct approach to try to call these callbacks with >> >> rcu protection because: >> >> >> >> - Cls API uses sleeping locks that cannot be used in rcu read section >> >> (hence the included trace). >> >> >> >> - It assumes that all implementation of classifier ops reoffload() don't >> >> sleep. >> >> >> >> - And that all driver offload callbacks (both block and classifier >> >> setup) don't sleep, which is not the case. >> >> >> >> I don't see any straightforward way to fix this, besides using some >> >> other locking mechanism to protect block_ing_cb_list. >> >> >> >> Regards, >> >> Vlad >> > >> > Maybe get the mutex flow_indr_block_ing_cb_lock for both lookup, add, delete? >> > >> > the callbacks_lists. the add and delete is work only on modules init case. So the >> > >> > lookup is also not frequently(ony [un]register) and can protect with the locks. >> >> That should do the job. I'll send the patch. > > Hi Vlad! > > While looking into this, would you mind also add the missing > flow_block_cb_is_busy() calls in the indirect handlers in the drivers? > > LMK if you're too busy, I don't want this to get forgotten :) Hi Jakub, I've checked the code and it looks like only nfp driver is affected: - I added check in nfp to lookup cb_priv with nfp_flower_indr_block_cb_priv_lookup() and call flow_block_cb_is_busy() if cb_priv exists. - In mlx5 en_rep.c there is already a check that indr_priv exists, so trying to lookup block_cb->cb_indent==indr_priv is redundant. - Switch drivers (mlxsw and ocelot) take reference to block_cb on FLOW_BLOCK_BIND, so they should not require any modifications. Tell me if I missed anything. Sending the patch for nfp. Regards, Vlad
