On 7/30/20 5:33 AM, Johannes Berg wrote:
On Wed, 2020-06-10 at 13:40 -0700, greearb@xxxxxxxxxxxxxxx wrote:
From: Ben Greear <greearb@xxxxxxxxxxxxxxx>
I backported out-of-tree ax200 driver from backport-iwlwifi to my
5.4 kernel so that I could run ax200 beside other radios (backports
mac80211 otherwise is incompatible and other drivers will crash).
Always possible that upstream kernel doesn't suffer from exactly this
case, but upstream ax200 is too unstable to even get this far, so...
The ax200 firmware crash often causes the kernel to deadlock due to the
while (sta->sta_state == IEEE80211_STA_AUTHORIZED)
loop in __sta_info_Destroy_part. If sta_info_move_state does not
make progress, then it will loop forever. In my case, sta_info_move_state
fails due to the sdata-in-driver check.
Interesting. I don't think I've seen this in our testing before.
So, put a few ax200 NICs (two is plenty) in a system and run any significant
upload. It crashes FW in < 1 minute on stock 5.7, like completely and utterly
unstable. The out-of-tree backports ax200 (which is what I was actually using here)
is a bit better, closer to 5.2 kernel ax200 stability,
but still crashes reliably on upload traffic with 2+ radios
in a chassis. It will crash on download too, but less often.
If you actually test this and have any trouble reproducing problems, please
let me know.
The firmware hard crash that causes ax200 phy to go away and cause the
sdata-in-driver / EIO busy spin is a bit harder to reproduce, but certainly
it happens often enough.
iwlwifi 0000:12:00.0: dma_pool_destroy iwlwifi:bc, 00000000d859bd4c busy
Ugh, yeah, as an aside - we still leak stuff there... need to dig into
that.
Signed-off-by: Ben Greear <greearb@xxxxxxxxxxxxxxx>
---
net/mac80211/sta_info.c | 23 +++++++++++++++++++++--
1 file changed, 21 insertions(+), 2 deletions(-)
diff --git a/net/mac80211/sta_info.c b/net/mac80211/sta_info.c
index e2a04fc..31a3856 100644
--- a/net/mac80211/sta_info.c
+++ b/net/mac80211/sta_info.c
@@ -1092,6 +1092,7 @@ static void __sta_info_destroy_part2(struct sta_info *sta)
struct ieee80211_sub_if_data *sdata = sta->sdata;
struct station_info *sinfo;
int ret;
+ int count = 0;
/*
* NOTE: This assumes at least synchronize_net() was done
@@ -1104,6 +1105,13 @@ static void __sta_info_destroy_part2(struct sta_info *sta)
while (sta->sta_state == IEEE80211_STA_AUTHORIZED) {
ret = sta_info_move_state(sta, IEEE80211_STA_ASSOC);
WARN_ON_ONCE(ret);
+ if (++count > 1000) {
+ /* WTF, bail out so that at least we don't hang the system. */
+ sdata_err(sdata, "Could not move state after 1000 tries, ret: %d state: %d\n",
+ ret, sta->sta_state);
+ WARN_ON_ONCE(1);
+ break;
+ }
I guess that should be
if (WARN_ON_ONCE()) ...
If we spin 1000 times, it is worth a second warning. Or do you mean the WARN_ON_ONCE(ret) should have if in front of it?
etc.
int err = drv_sta_state(sta->local, sta->sdata, sta,
sta->sta_state, new_state);
- if (err)
- return err;
+ if (err == -EIO) {
+ /* Sdata-not-in-driver, we are out of sync, but probably
+ * best to carry on instead of bailing here, at least maybe
+ * we can clean this up.
+ */
It _could_ be the driver itself returning -EIO, so why not check the
sdata-in-driver flag?
Right, but if driver is complaining here, we need to bail out regardless of
sdata-in-driver or not, unless you think a driver could return EIO and then
a small bit later start working for the same request?
Anyway, that mostly looks good and would make mac80211 more robust, but
like I just said in the other patch I think you need to consider
mac80211 changes more from mac80211's POV, not from an arbitrary
driver's POV.
Really here that mostly applies to the commit log, which should probably
say something like
mac80211: deadlock due to driver misbehaviour
or so, and then go on to explain what it does in *mac80211*, and show
the iwlwifi parts only as an *example*.
Its not really driver mis-behaviour per se. The root cause is that the
firmware crashes too badly for the driver to recover (ok, so driver might
could be better, but I've also seen cases where ath10k NIC falls off the PCI
bus, so nothing the driver can do in that case I think).
Per my other patches, I've seen this sdata-in-driver crap in the past, so
I think I probably hit a similar bug in both ax200 and ath10k, but since
ax200 is so easy to crash, it is much more likely to hit this bug than any
other driver I'm aware of.
I'll try to re-word the commit message though, I don't really care what it
says so long as the code gets in.
Thanks,
Ben
Thanks,
johannes
--
Ben Greear <greearb@xxxxxxxxxxxxxxx>
Candela Technologies Inc http://www.candelatech.com
