On 24-1-2017 10:40, Johannes Berg wrote: > >> + * @max_sched_scan_reqs: maximum number of scheduled scan requests >> that >> + * the device can run concurrently. > > Perhaps we should get rid of WIPHY_FLAG_SUPPORTS_SCHED_SCAN and just > set this to 1 for such devices? Otherwise we have two different > requirements, and we need to track that 0 is an invalid value here if > WIPHY_FLAG_SUPPORTS_SCHED_SCAN is set, or something like that? Ok. Doesn't that cause issues in user-space. Or do you only want to get rid of it in cfg80211 api and report the flag to user-space when max_sched_scan_reqs equals 1? >> + * @NL80211_ATTR_SCHED_SCAN_MAX_REQS: indicates maximum number of >> scheduled >> + * scan request that may be active for the device (u8). > > I'd make that a u32 - not that I believe we'll ever want to change this > in the future, but there's simply no value in making it a u8 since it > uses the same amount of space in a netlink message. Ok. >> + list_for_each_entry_safe(pos, tmp, &rdev- >>> sched_scan_req_list, list) { >> + cfg80211_stop_sched_scan_req(rdev, pos, false); >> + } > > nit: don't really need braces here. True. >> + if ((wiphy->flags & WIPHY_FLAG_SUPPORTS_SCHED_SCAN) && >> + !wiphy->max_sched_scan_reqs) >> + wiphy->max_sched_scan_reqs = 1; > > Yeah, this. Why bother? > > (should even be simple to come up with an spatch to change all the > drivers, but there are only five anyway) Done. >> + nla_put_u8(msg, >> NL80211_ATTR_SCHED_SCAN_MAX_REQS, >> + rdev->wiphy.max_sched_scan_reqs) || >> nla_put_u8(msg, >> NL80211_ATTR_MAX_NUM_SCHED_SCAN_SSIDS, >> rdev->wiphy.max_sched_scan_ssids) || > > This might break older userspace - you'll have to put it in a later > portion of the code. > > I'm also a bit surprised the attributes aren't actually optional for > when sched scan isn't supported, I'd make the new one optional and I > guess we can fix the others later too, if desired. Why would it break user-space. Is the order in which attributes are added into the stream something user-space relies on. >> + bool want_multi; > > That's bool > >> + want_multi = !!info->attrs[NL80211_ATTR_SCHED_SCAN_MULTI]; > > so you don't really need the !! as it's implied by the rules for bool > :) I see. >> + /* leave request id zero for legacy request >> + * or if driver does not support multi-scheduled scan >> + */ >> + if (want_multi && rdev->wiphy.max_sched_scan_reqs > 1) { > > Why do the >1 check here? It probably doesn't really make a difference > since only one can be running at a time, but it might be nicer - at > least for debug in userspace - to have a real value for all multi > scans? > >> + while (!sched_scan_req->reqid) > > Pretty sure we won't run over the u64 ... but I guess it doesn't matter > much :) > > > I don't see you sending the reqid/cookie back to userspace here though, > that's missing? Indeed that is in the second patch. Maybe I should put that commit first in the series. >> static int nl80211_stop_sched_scan(struct sk_buff *skb, >> struct genl_info *info) >> { >> + struct cfg80211_sched_scan_request *req; >> struct cfg80211_registered_device *rdev = info->user_ptr[0]; >> + u64 cookie; >> >> if (!(rdev->wiphy.flags & WIPHY_FLAG_SUPPORTS_SCHED_SCAN) || >> !rdev->ops->sched_scan_stop) >> return -EOPNOTSUPP; >> >> - return __cfg80211_stop_sched_scan(rdev, false); >> + if (info->attrs[NL80211_ATTR_COOKIE]) { >> + cookie = nla_get_u64(info- >>> attrs[NL80211_ATTR_COOKIE]); >> + return __cfg80211_stop_sched_scan(rdev, cookie, >> false); >> + } else { >> + req = list_first_or_null_rcu(&rdev- >>> sched_scan_req_list, >> + struct >> cfg80211_sched_scan_request, >> + list); >> + if (!req || req->reqid || >> + (req->owner_nlportid && >> + req->owner_nlportid != info->snd_portid)) >> + return -ENOENT; > > Shouldn't this also check that it's non-multi? non-multi == (req->reqid == 0). non-multi/legacy and multi can not be active at the same time so we can use list_first_or_null_rcu here. If req->reqid of first entry is non-zero there is no non-multi to stop here. >> +void cfg80211_add_sched_scan_req(struct cfg80211_registered_device >> *rdev, >> + struct cfg80211_sched_scan_request >> *req) >> +{ >> + list_add_rcu(&req->list, &rdev->sched_scan_req_list); >> +} >> + >> +static void cfg80211_del_sched_scan_req(struct >> cfg80211_registered_device *rdev, >> + struct >> cfg80211_sched_scan_request *req) >> +{ >> + list_del_rcu(&req->list); >> + kfree_rcu(req, rcu_head); >> +} > > Some locking assertions in these would be good, I think. Will do. >> +static struct cfg80211_sched_scan_request * >> +cfg80211_find_sched_scan_req(struct cfg80211_registered_device >> *rdev, u64 reqid) >> +{ >> + struct cfg80211_sched_scan_request *pos; >> + >> + list_for_each_entry(pos, &rdev->sched_scan_req_list, list) { >> + if (pos->reqid == reqid) >> + return pos; >> + } >> + return ERR_PTR(-ENOENT); >> +} > > Here too, I guess, since you don't actually use RCU. So should I use RCU here? Not sure what is the better choice here. Regards, Arend