| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
In case of 4way handshake offload, transition disable policy
updated by the AP during EAPOL 3/4 is not updated to the upper layer.
This results in mismatch between transition disable policy
between the upper layer and the driver. This patch addresses this
issue by updating transition disable policy as part of port
authorization indication.
Signed-off-by: Vinayak Yadawad <vinayak.yadawad@broadcom.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
minstrel_priv
Remove unused has_mrr (has multi-rate retry capabilities) member
from struct minstrel_priv (only set once in minstrel_ht_alloc, never
used again).
Signed-off-by: Peter Seiderer <ps.report@gmx.net>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
This should of course be CONFIG_, not CPTCFG_, which is an
artifact from working with backports.
Fixes: 9dd1953846c7 ("wifi: nl80211/mac80211: clarify link ID in control port TX")
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
HE added support for dynamic fragmentation inside aggregation
sessions, but no existing driver ever advertises it. Thus,
remove the code for now, it cannot work as-is in MLO. For it
to properly work in MLO, we'd need to validate that the frag
level is identical across all the link bands/iftypes, which
is a good amount of complex code that's just not worth it as
long as no driver has support for it.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
If the deflink of the station is on 6 GHz, then it won't have HT.
If at the same time we're using MLO, then vif.bss_conf isn't used,
and thus vif.bss_conf.chandef.chan is NULL, causing the code to
crash.
Fix this by just checking for both HT and HE, and refusing the
aggregation session if both are not present. This might be a bit
wrong since it would accept an aggregation session from a peer
that has HE but no HT on 2.4 or 5 GHz, but such a peer shouldn't
exist in the first place, and it probably supports aggregation if
it has HE support.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
This won't work right at least with the code as it is, so
at least for now just assume it's never set for MLO. It may
very well never change, almost no drivers support it.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
Fix the code that sets the DTIM period to always propagate it
into link->conf->dtim_period and not overwrite it, while still
preferring to set it from the beacon data if available.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
If the station has no HT, we deny the aggregation session
but the error message talks about QoS; change it to say HT
instead.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
It's useful to know which link was used for the association,
mark it when printing the links.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
Ensure that the link ID matches in auth/assoc continuation,
otherwise we need to reset all the data.
Fixes: 81151ce462e5 ("wifi: mac80211: support MLO authentication/association with one link")
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
If association to an AP without a link 0 fails, then we crash in
tracing because it assumes that either ap_mld_addr or link 0 BSS
is valid, since we clear sdata->vif.valid_links and then don't
add the ap_mld_addr to the struct.
Since we clear also sdata->vif.cfg.ap_addr, keep a local copy of
it and assign it earlier, before clearing valid_links, to fix
this.
Fixes: 81151ce462e5 ("wifi: mac80211: support MLO authentication/association with one link")
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
Since this frame is addressed from/to an MLD, it should be
built with the correct AP MLD address (in station mode) to
be encrypted properly.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
If any driver relies entirely on the scan request BSSID,
then that would be wrong for internal scans. Initialize
it to the broadcast address since we don't otherwise use
the field.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
Currently, we rely only on the AP capability. If the AP supports
TWT responder we will advertise TWT requester even if the driver
or HW doesn't support it. Fix this by checking the HW capability.
Signed-off-by: Haim Dreyfuss <haim.dreyfuss@intel.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
We should set the parameters here per link, except
unfortunately ap_isolate, but we can't really change
that anymore so it'll remain a quirk in the API in
that you need to change it on one of the valid links
and it'll apply to all.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
We clearly need the link ID here, to know the right BSS
to configure. Use/require it.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
This management frame is intended for the MLD so we
treat it in mac80211 as MLD addressed as well, and
should therefore use the MLD address of the AP for
the BSSID field in the frame, address translation
applies.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
We use this to look up the destination station, so it
needs to be the MLD address of the AP for an MLO; use
ap_addr instead of the BSSID.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
In case the AP returned a non success status for one of the links,
do not activate the link.
Signed-off-by: Ilan Peer <ilan.peer@intel.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
When processing an association response frame for a Multi-Link
connection, extract the per station profile for each additional
link, and use it for parsing the link elements.
As the Multi-Link element might be fragmented, add support for
reassembling a fragmented element. To simplify memory management
logic, extend 'struct ieee802_11_elems' to hold a scratch buffer,
which is used for the defragmentation. Once an element is
reconstructed in the scratch area, point the corresponding element
pointer to it. Currently only defragmentation of Multi-Link element
and the contained per-STA profile subelement is supported.
Signed-off-by: Ilan Peer <ilan.peer@intel.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
For assoc and connect result APIs, support reporting
failed links; they should still come with the BSS
pointer in the case of assoc, so they're released
correctly. In the case of connect result, this is
optional.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
During link switching, the active links change, so we need to
recalculate the aggregate data in the stations.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
Add the link address to the per-link information, but only if we are
using MLO.
Signed-off-by: Benjamin Berg <benjamin.berg@intel.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
Create debugfs data per-link. For drivers, there is a new operation
link_sta_add_debugfs which will always be called.
For non-MLO, the station directory will be used directly rather than
creating a corresponding subdirectory. As such, non-MLO drivers can
simply continue to create the data from sta_debugfs_add.
Signed-off-by: Benjamin Berg <benjamin.berg@intel.com>
[add missing inlines if !CONFIG_MAC80211_DEBUGFS]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
While often not needed, this considerably simplifies going from a link
to the STA. This helps in cases such as debugfs where a single pointer
should allow accessing a specific link and the STA.
Signed-off-by: Benjamin Berg <benjamin.berg@intel.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
flexible array
To work around a misbehavior of the compiler's ability to see into
composite flexible array structs (as detailed in the coming memcpy()
hardening series[1]), split the memcpy() of the header and the payload
so no false positive run-time overflow warning will be generated.
[1] https://lore.kernel.org/linux-hardening/20220901065914.1417829-2-keescook@chromium.org/
Signed-off-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Gustavo A. R. Silva <gustavoars@kernel.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
Syzkaller reports buffer overflow false positive as follows:
------------[ cut here ]------------
memcpy: detected field-spanning write (size 8) of single field
"&compat_event->pointer" at net/wireless/wext-core.c:623 (size 4)
WARNING: CPU: 0 PID: 3607 at net/wireless/wext-core.c:623
wireless_send_event+0xab5/0xca0 net/wireless/wext-core.c:623
Modules linked in:
CPU: 1 PID: 3607 Comm: syz-executor659 Not tainted
6.0.0-rc6-next-20220921-syzkaller #0
[...]
Call Trace:
<TASK>
ioctl_standard_call+0x155/0x1f0 net/wireless/wext-core.c:1022
wireless_process_ioctl+0xc8/0x4c0 net/wireless/wext-core.c:955
wext_ioctl_dispatch net/wireless/wext-core.c:988 [inline]
wext_ioctl_dispatch net/wireless/wext-core.c:976 [inline]
wext_handle_ioctl+0x26b/0x280 net/wireless/wext-core.c:1049
sock_ioctl+0x285/0x640 net/socket.c:1220
vfs_ioctl fs/ioctl.c:51 [inline]
__do_sys_ioctl fs/ioctl.c:870 [inline]
__se_sys_ioctl fs/ioctl.c:856 [inline]
__x64_sys_ioctl+0x193/0x200 fs/ioctl.c:856
do_syscall_x64 arch/x86/entry/common.c:50 [inline]
do_syscall_64+0x35/0xb0 arch/x86/entry/common.c:80
entry_SYSCALL_64_after_hwframe+0x63/0xcd
[...]
</TASK>
Wireless events will be sent on the appropriate channels in
wireless_send_event(). Different wireless events may have different
payload structure and size, so kernel uses **len** and **cmd** field
in struct __compat_iw_event as wireless event common LCP part, uses
**pointer** as a label to mark the position of remaining different part.
Yet the problem is that, **pointer** is a compat_caddr_t type, which may
be smaller than the relative structure at the same position. So during
wireless_send_event() tries to parse the wireless events payload, it may
trigger the memcpy() run-time destination buffer bounds checking when the
relative structure's data is copied to the position marked by **pointer**.
This patch solves it by introducing flexible-array field **ptr_bytes**,
to mark the position of the wireless events remaining part next to
LCP part. What's more, this patch also adds **ptr_len** variable in
wireless_send_event() to improve its maintainability.
Reported-and-tested-by: syzbot+473754e5af963cf014cf@syzkaller.appspotmail.com
Link: https://lore.kernel.org/all/00000000000070db2005e95a5984@google.com/
Suggested-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Hawkins Jiawei <yin31149@gmail.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
STP topology change notification packets only have a payload of 7 bytes,
so they get dropped due to the skb->len < hdrlen + 8 check.
Fix this by removing the extra 8 from the skb->len check and checking the
return code on the skb_copy_bits calls.
Fixes: 2d1c304cb2d5 ("cfg80211: add function for 802.3 conversion with separate output buffer")
Reported-by: Chad Monroe <chad.monroe@smartrg.com>
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
Properly handle TX stop for internal queues (iTXQs) within mac80211.
mac80211 must not stop netdev queues when using mac80211 iTXQs.
For these drivers the netdev interface is created with IFF_NO_QUEUE.
While netdev still drops frames for IFF_NO_QUEUE interfaces when we stop
the netdev queues, it also prints a warning when this happens:
Assuming the mac80211 interface is called wlan0 we would get
"Virtual device wlan0 asks to queue packet!" when netdev has to drop a
frame.
This patch is keeping the harmless netdev queue starts for iTXQ drivers.
Signed-off-by: Alexander Wetzel <alexander@wetzel-home.de>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
Since AP_VLAN interfaces are not passed to the driver, check offload_flags
on the bss vif instead.
Reported-by: Howard Hsu <howard-yh.hsu@mediatek.com>
Fixes: 80a915ec4427 ("mac80211: add rx decapsulation offload support")
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
Unlock before returning -EOPNOTSUPP.
Fixes: 3c06e91b40db ("wifi: mac80211: Support POWERED_ADDR_CHANGE feature")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
At some point a few kernel debug prints started appearing which
indicated something was sending invalid IEs:
"bad VHT capabilities, disabling VHT"
"Invalid HE elem, Disable HE"
Turns out these were being printed because the local hardware
supported HE/VHT but the peer/AP did not. Bad/invalid indicates,
to me at least, that the IE is in some way malformed, not missing.
For the HE print (ieee80211_verify_peer_he_mcs_support) it will
now silently fail if the HE capability element is missing (still
prints if the element size is wrong).
For the VHT print, it has been removed completely and will silently
set the DISABLE_VHT flag which is consistent with how DISABLE_HT
is set.
Signed-off-by: James Prestwood <prestwoj@gmail.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
When building the probe request IEs HE support is checked for
the 6GHz band (wiphy->bands[NL80211_BAND_6GHZ]). If supported
the HE capability IE should be included according to the spec.
The problem is the 16-bit capability is obtained from the
band object (sband) that was passed in, not the 6GHz band
object (sband6). If the sband object doesn't support HE it will
result in a warning.
Fixes: 7d29bc50b30e ("mac80211: always include HE 6GHz capability in probe request")
Signed-off-by: James Prestwood <prestwoj@gmail.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
Since STP TCN frames are only 7 bytes, the pskb_may_pull call returns an error.
Instead of dropping those packets, bump them back to the slow path for proper
processing.
Fixes: 49ddf8e6e234 ("mac80211: add fast-rx path")
Reported-by: Chad Monroe <chad.monroe@smartrg.com>
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
|
|
This parse_opts will set invalid opts.rfd/wfd in case of failure which
we already check, but it is not clear for readers that parse_opts error
are handled in p9_fd_create: clarify this by explicitely checking the
return value.
Link: https://lkml.kernel.org/r/20220921210921.1654735-1-floridsleeves@gmail.com
Signed-off-by: Li Zhong <floridsleeves@gmail.com>
[Dominique: reworded commit message to clarify this is NOOP]
Signed-off-by: Dominique Martinet <asmadeus@codewreck.org>
|
|
xen transport was missing annotations
Link: https://lkml.kernel.org/r/20220909103546.73015-1-xiujianfeng@huawei.com
Signed-off-by: Xiu Jianfeng <xiujianfeng@huawei.com>
Signed-off-by: Dominique Martinet <asmadeus@codewreck.org>
|
|
Shamelessly copying the explanation from Tetsuo Handa's suggested
patch[1] (slightly reworded):
syzbot is reporting inconsistent lock state in p9_req_put()[2],
for p9_tag_remove() from p9_req_put() from IRQ context is using
spin_lock_irqsave() on "struct p9_client"->lock but trans_fd
(not from IRQ context) is using spin_lock().
Since the locks actually protect different things in client.c and in
trans_fd.c, just replace trans_fd.c's lock by a new one specific to the
transport (client.c's protect the idr for fid/tag allocations,
while trans_fd.c's protects its own req list and request status field
that acts as the transport's state machine)
Link: https://lore.kernel.org/r/20220904112928.1308799-1-asmadeus@codewreck.org
Link: https://lkml.kernel.org/r/2470e028-9b05-2013-7198-1fdad071d999@I-love.SAKURA.ne.jp [1]
Link: https://syzkaller.appspot.com/bug?extid=2f20b523930c32c160cc [2]
Reported-by: syzbot <syzbot+2f20b523930c32c160cc@syzkaller.appspotmail.com>
Reported-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Reviewed-by: Christian Schoenebeck <linux_oss@crudebyte.com>
Signed-off-by: Dominique Martinet <asmadeus@codewreck.org>
|
|
reference
Gwangun Jung reported a slab-out-of-bounds access in fib_nh_match:
fib_nh_match+0xf98/0x1130 linux-6.0-rc7/net/ipv4/fib_semantics.c:961
fib_table_delete+0x5f3/0xa40 linux-6.0-rc7/net/ipv4/fib_trie.c:1753
inet_rtm_delroute+0x2b3/0x380 linux-6.0-rc7/net/ipv4/fib_frontend.c:874
Separate nexthop objects are mutually exclusive with the legacy
multipath spec. Fix fib_nh_match to return if the config for the
to be deleted route contains a multipath spec while the fib_info
is using a nexthop object.
Fixes: 493ced1ac47c ("ipv4: Allow routes to use nexthop objects")
Fixes: 6bf92d70e690 ("net: ipv4: fix route with nexthop object delete warning")
Reported-by: Gwangun Jung <exsociety@gmail.com>
Signed-off-by: David Ahern <dsahern@kernel.org>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Tested-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Fix to return error code -EINVAL from the error handling
case instead of 0, as done elsewhere in this function.
Fixes: 94160108a70c ("net/ieee802154: fix uninit value bug in dgram_sendmsg")
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Link: https://lore.kernel.org/r/20220919160830.1436109-1-weiyongjun@huaweicloud.com
Signed-off-by: Stefan Schmidt <stefan@datenfreihafen.org>
|
|
syzbot is reporting hung task at p9_fd_close() [1], for p9_mux_poll_stop()
from p9_conn_destroy() from p9_fd_close() is failing to interrupt already
started kernel_read() from p9_fd_read() from p9_read_work() and/or
kernel_write() from p9_fd_write() from p9_write_work() requests.
Since p9_socket_open() sets O_NONBLOCK flag, p9_mux_poll_stop() does not
need to interrupt kernel_read()/kernel_write(). However, since p9_fd_open()
does not set O_NONBLOCK flag, but pipe blocks unless signal is pending,
p9_mux_poll_stop() needs to interrupt kernel_read()/kernel_write() when
the file descriptor refers to a pipe. In other words, pipe file descriptor
needs to be handled as if socket file descriptor.
We somehow need to interrupt kernel_read()/kernel_write() on pipes.
A minimal change, which this patch is doing, is to set O_NONBLOCK flag
from p9_fd_open(), for O_NONBLOCK flag does not affect reading/writing
of regular files. But this approach changes O_NONBLOCK flag on userspace-
supplied file descriptors (which might break userspace programs), and
O_NONBLOCK flag could be changed by userspace. It would be possible to set
O_NONBLOCK flag every time p9_fd_read()/p9_fd_write() is invoked, but still
remains small race window for clearing O_NONBLOCK flag.
If we don't want to manipulate O_NONBLOCK flag, we might be able to
surround kernel_read()/kernel_write() with set_thread_flag(TIF_SIGPENDING)
and recalc_sigpending(). Since p9_read_work()/p9_write_work() works are
processed by kernel threads which process global system_wq workqueue,
signals could not be delivered from remote threads when p9_mux_poll_stop()
from p9_conn_destroy() from p9_fd_close() is called. Therefore, calling
set_thread_flag(TIF_SIGPENDING)/recalc_sigpending() every time would be
needed if we count on signals for making kernel_read()/kernel_write()
non-blocking.
Link: https://lkml.kernel.org/r/345de429-a88b-7097-d177-adecf9fed342@I-love.SAKURA.ne.jp
Link: https://syzkaller.appspot.com/bug?extid=8b41a1365f1106fd0f33 [1]
Reported-by: syzbot <syzbot+8b41a1365f1106fd0f33@syzkaller.appspotmail.com>
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Tested-by: syzbot <syzbot+8b41a1365f1106fd0f33@syzkaller.appspotmail.com>
Reviewed-by: Christian Schoenebeck <linux_oss@crudebyte.com>
[Dominique: add comment at Christian's suggestion]
Signed-off-by: Dominique Martinet <asmadeus@codewreck.org>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
Pull vfs d_path updates from Al Viro.
* tag 'pull-d_path' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
d_path.c: typo fix...
dynamic_dname(): drop unused dentry argument
|
|
Allow the caller to force a disconnection of the RPC client so that we
can clear any pending requests that are buffered in the socket.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Add the helper rpc_cancel_tasks(), which uses a caller-defined selection
function to define a set of in-flight RPC calls to cancel. This is
mainly intended for pNFS drivers which are subject to a layout recall,
and which may therefore want to cancel all pending I/O using that layout
in order to redrive it after the layout recall has been satisfied.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Ensure that we immediately call rpc_exit_task() after waking up, and
that the tk_rpc_status cannot get clobbered by some other function.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/sschmidt/wpan
Stefan Schmidt says:
====================
pull-request: ieee802154 for net 2022-10-05
Only two patches this time around. A revert from Alexander Aring to a patch
that hit net and the updated patch to fix the problem from Tetsuo Handa.
* tag 'ieee802154-for-net-2022-10-05' of git://git.kernel.org/pub/scm/linux/kernel/git/sschmidt/wpan:
net/ieee802154: don't warn zero-sized raw_sendmsg()
Revert "net/ieee802154: reject zero-sized raw_sendmsg()"
====================
Link: https://lore.kernel.org/r/20221005144508.787376-1-stefan@datenfreihafen.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
child qdiscs"
taprio_attach() has this logic at the end, which should have been
removed with the blamed patch (which is now being reverted):
/* access to the child qdiscs is not needed in offload mode */
if (FULL_OFFLOAD_IS_ENABLED(q->flags)) {
kfree(q->qdiscs);
q->qdiscs = NULL;
}
because otherwise, we make use of q->qdiscs[] even after this array was
deallocated, namely in taprio_leaf(). Therefore, whenever one would try
to attach a valid child qdisc to a fully offloaded taprio root, one
would immediately dereference a NULL pointer.
$ tc qdisc replace dev eno0 handle 8001: parent root taprio \
num_tc 8 \
map 0 1 2 3 4 5 6 7 \
queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \
max-sdu 0 0 0 0 0 200 0 0 \
base-time 200 \
sched-entry S 80 20000 \
sched-entry S a0 20000 \
sched-entry S 5f 60000 \
flags 2
$ max_frame_size=1500
$ data_rate_kbps=20000
$ port_transmit_rate_kbps=1000000
$ idleslope=$data_rate_kbps
$ sendslope=$(($idleslope - $port_transmit_rate_kbps))
$ locredit=$(($max_frame_size * $sendslope / $port_transmit_rate_kbps))
$ hicredit=$(($max_frame_size * $idleslope / $port_transmit_rate_kbps))
$ tc qdisc replace dev eno0 parent 8001:7 cbs \
idleslope $idleslope \
sendslope $sendslope \
hicredit $hicredit \
locredit $locredit \
offload 0
Unable to handle kernel NULL pointer dereference at virtual address 0000000000000030
pc : taprio_leaf+0x28/0x40
lr : qdisc_leaf+0x3c/0x60
Call trace:
taprio_leaf+0x28/0x40
tc_modify_qdisc+0xf0/0x72c
rtnetlink_rcv_msg+0x12c/0x390
netlink_rcv_skb+0x5c/0x130
rtnetlink_rcv+0x1c/0x2c
The solution is not as obvious as the problem. The code which deallocates
q->qdiscs[] is in fact copied and pasted from mqprio, which also
deallocates the array in mqprio_attach() and never uses it afterwards.
Therefore, the identical cleanup logic of priv->qdiscs[] that
mqprio_destroy() has is deceptive because it will never take place at
qdisc_destroy() time, but just at raw ops->destroy() time (otherwise
said, priv->qdiscs[] do not last for the entire lifetime of the mqprio
root), but rather, this is just the twisted way in which the Qdisc API
understands error path cleanup should be done (Qdisc_ops :: destroy() is
called even when Qdisc_ops :: init() never succeeded).
Side note, in fact this is also what the comment in mqprio_init() says:
/* pre-allocate qdisc, attachment can't fail */
Or reworded, mqprio's priv->qdiscs[] scheme is only meant to serve as
data passing between Qdisc_ops :: init() and Qdisc_ops :: attach().
[ this comment was also copied and pasted into the initial taprio
commit, even though taprio_attach() came way later ]
The problem is that taprio also makes extensive use of the q->qdiscs[]
array in the software fast path (taprio_enqueue() and taprio_dequeue()),
but it does not keep a reference of its own on q->qdiscs[i] (you'd think
that since it creates these Qdiscs, it holds the reference, but nope,
this is not completely true).
To understand the difference between taprio_destroy() and mqprio_destroy()
one must look before commit 13511704f8d7 ("net: taprio offload: enforce
qdisc to netdev queue mapping"), because that just muddied the waters.
In the "original" taprio design, taprio always attached itself (the root
Qdisc) to all netdev TX queues, so that dev_qdisc_enqueue() would go
through taprio_enqueue().
It also called qdisc_refcount_inc() on itself for as many times as there
were netdev TX queues, in order to counter-balance what tc_get_qdisc()
does when destroying a Qdisc (simplified for brevity below):
if (n->nlmsg_type == RTM_DELQDISC)
err = qdisc_graft(dev, parent=NULL, new=NULL, q, extack);
qdisc_graft(where "new" is NULL so this deletes the Qdisc):
for (i = 0; i < num_q; i++) {
struct netdev_queue *dev_queue;
dev_queue = netdev_get_tx_queue(dev, i);
old = dev_graft_qdisc(dev_queue, new);
if (new && i > 0)
qdisc_refcount_inc(new);
qdisc_put(old);
~~~~~~~~~~~~~~
this decrements taprio's refcount once for each TX queue
}
notify_and_destroy(net, skb, n, classid,
rtnl_dereference(dev->qdisc), new);
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
and this finally decrements it to zero,
making qdisc_put() call qdisc_destroy()
The q->qdiscs[] created using qdisc_create_dflt() (or their
replacements, if taprio_graft() was ever to get called) were then
privately freed by taprio_destroy().
This is still what is happening after commit 13511704f8d7 ("net: taprio
offload: enforce qdisc to netdev queue mapping"), but only for software
mode.
In full offload mode, the per-txq "qdisc_put(old)" calls from
qdisc_graft() now deallocate the child Qdiscs rather than decrement
taprio's refcount. So when notify_and_destroy(taprio) finally calls
taprio_destroy(), the difference is that the child Qdiscs were already
deallocated.
And this is exactly why the taprio_attach() comment "access to the child
qdiscs is not needed in offload mode" is deceptive too. Not only the
q->qdiscs[] array is not needed, but it is also necessary to get rid of
it as soon as possible, because otherwise, we will also call qdisc_put()
on the child Qdiscs in qdisc_destroy() -> taprio_destroy(), and this
will cause a nasty use-after-free/refcount-saturate/whatever.
In short, the problem is that since the blamed commit, taprio_leaf()
needs q->qdiscs[] to not be freed by taprio_attach(), while qdisc_destroy()
-> taprio_destroy() does need q->qdiscs[] to be freed by taprio_attach()
for full offload. Fixing one problem triggers the other.
All of this can be solved by making taprio keep its q->qdiscs[i] with a
refcount elevated at 2 (in offloaded mode where they are attached to the
netdev TX queues), both in taprio_attach() and in taprio_graft(). The
generic qdisc_graft() would just decrement the child qdiscs' refcounts
to 1, and taprio_destroy() would give them the final coup de grace.
However the rabbit hole of changes is getting quite deep, and the
complexity increases. The blamed commit was supposed to be a bug fix in
the first place, and the bug it addressed is not so significant so as to
justify further rework in stable trees. So I'd rather just revert it.
I don't know enough about multi-queue Qdisc design to make a proper
judgement right now regarding what is/isn't idiomatic use of Qdisc
concepts in taprio. I will try to study the problem more and come with a
different solution in net-next.
Fixes: 1461d212ab27 ("net/sched: taprio: make qdisc_leaf() see the per-netdev-queue pfifo child qdiscs")
Reported-by: Muhammad Husaini Zulkifli <muhammad.husaini.zulkifli@intel.com>
Reported-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Link: https://lore.kernel.org/r/20221004220100.1650558-1-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
net/sunrpc/xprtrdma/frwr_ops.c:151:32: warning: variable 'rc' is uninitialized when used here [-Wuninitialized]
trace_xprtrdma_frwr_alloc(mr, rc);
^~
net/sunrpc/xprtrdma/frwr_ops.c:127:8: note: initialize the variable 'rc' to silence this warning
int rc;
^
= 0
1 warning generated.
The tracepoint is intended to record the error returned from
ib_alloc_mr(). In the current code there is no other purpose for
@rc, so simply replace it.
Reported-by: kernel test robot <lkp@intel.com>
Fixes: d8cf39a280c3b0 ('xprtrdma: MR-related memory allocation should be allowed to fail')
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
Many memory allocations that xprtrdma does can fail safely. Let's
use this fact to avoid some potential deadlocks: Replace GFP_KERNEL
with GFP flags that do not try hard to acquire memory.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
An attempt to establish a connection can always fail and then be
retried. GFP_KERNEL allocation is not necessary here.
Like MR allocation, establishing a connection is always done in a
worker thread. The new GFP flags align with the flags that would
be returned by rpc_task_gfp_mask() in this case.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
|
xprtrdma always drives a retry of MR allocation if it should fail.
It should be safe to not use GFP_KERNEL for this purpose rather
than sleeping in the memory allocator.
In theory, if these weaker allocations are attempted first, memory
exhaustion is likely to cause xprtrdma to fail fast and not then
invoke the RDMA core APIs, which still might use GFP_KERNEL.
Also note that rpc_task_gfp_mask() always sets __GFP_NORETRY and
__GFP_NOWARN when an RPC-related allocation is being done in a
worker thread. MR allocation is already always done in worker
threads.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
|
