• in vt_fb_setpixel(): re-order offset calculation in the 90 degree case
• vt_fb_bitblt_bitmap(): add break after the last case

• reduce number of fb_flags to two
• use switch/case in vt_fb_setpixel() and vt_fb_bitblt_bitmap() rotation code

flags base 16 in fbio.h, style(9) in vt_fb.c

Author: Stefan Grundmann <sg2342@googlemail.com>

i was glad that there was no dimension validation in vt_fb_drawrect(), otherwise https://reviews.freebsd.org/D34221 would be more complicated to do.

subtract one from vt_logo_sprite_height as well

This commit changed the behavior of dwc_attach() in a way that broke ethernet on my Allwinner A20 based cubieboard2:

In D26137#611017, @mmacy wrote:

It's a real bug insofar as people will run this virtualized. It's probably not possible to reproduce this on physical hardware because the OS is delivering new packets to a closed socket almost half a second after the socket has been closed. This suggests that the bhyve thread was blocked from running after the packet arrived but before the socket was closed.

FreeBSD ifnet is full of life cycle issues, it would take fairly thorough revamping, eliminating ifnet pointers floating around the stack that no one is going to pay for. So I'm loath to put too much additional time in to this. This is further compounded in a virtual environment where vcpu scheduling can create a situation where one vcpu destroys an object and another vcpu runs hundreds of milliseconds later accessing it. I've already added considerable overhead to the data path as it is for things that no one would actually see in production that I would at some point like to see taken out.

Since we're just dealing with hypervisor scheduling at this point the only thing we can do here is add a delay:

diff --git a/sys/dev/if_wg/module/module.c b/sys/dev/if_wg/module/module.c
index 0d5aca904ec..cf2f10fb697 100644
--- a/sys/dev/if_wg/module/module.c
+++ b/sys/dev/if_wg/module/module.c
@@ -340,6 +340,7 @@ wg_detach(if_ctx_t ctx)
        taskqgroup_drain_all(qgroup_if_io_tqg);
        pause("link_down", hz/4);
        wg_peer_remove_all(sc);
+       pause("link_down", hz);
        mtx_destroy(&sc->sc_mtx);
        rw_destroy(&sc->sc_index_lock);
        taskqgroup_detach(qgroup_if_io_tqg, &sc->sc_handshake);

moved the test setup to a different machine and after 1 hour and 19 minutes of running the test setup i got a panic here:

In D26137#610764, @mmacy wrote:

I don't have time to test right now, but this is an analogous double free fix in the wg_encap path that I did earlier in the wg_decap path.

i did not test Diff 21 79917 ( Mon, Nov 23, 10:04 PM ) since it does not seem to involve the code responsible for the device destruction panics.

with 79581 applied, device destruction can still panic the kernel (same setup as before: loop create, ping, sleep, destroy on FreeBSD and ping flood the wg address of the FreeBSD machine)

with the same setup (on FreeBSD: while true; do ifconfig wg0 create .....; ping -c 1 PEERIP; sleep 1; ifconfig wg0 destroy; done and on the Linux peer: ping -f FreeBSDwgIP) i can also get a different panic: here the gtaskqueue_drain thread got to wg_deliver_in(...) but peer->p_sc->sc_socket->so_so4 is 0x0

with 78675 applied (and clone_setdefcallback changed to clone_setdefcallback_prefix in ifwg.c), i got another
kernel panic (on wg device destruction):

Another kernel panic triggered by interface destruction: incoming upd traffic from the wg peer arrives in wg_input() where sc is already gone.

In order to have working wg in VIMAGE jails:

This might be out of scope of this review:
the WGC_SET ioctl is not priv(9) checked (and there is no PRIV_NET_WG entry in sys/priv.h)

IMHO wg_get() (sys/dev/if_wg/module/module.c:526) should not expose private-key
and wireguard_status() (sbin/ifconfig/ifwg.c:546) should not print it.

When a peer has more than one AllowedIPs, dump_peer() (sys/ifconfig/ifwg.c:270)
will print the address (but not the mask) of the first entry multiple times
because of sys/ifconfig/ifwg.c:306 which should be

In D26137#587248, @mmacy wrote:

In D26137#586721, @sg2342_googlemail.com wrote:

According to ifconfig(8): Cloned interfaces are members of their interface family group by default.

The use of iflib_clone_register() in module.c prevents that this is done for if_wg.

How so? iflib_clone_register calls if_clone_simple which is what vxlan does as well.

According to ifconfig(8): Cloned interfaces are members of their interface family group by default.

if the wg interface has an ipv6 address, SIOCIFDESTROY can panic the kernel.

after reading D24586 :

update path in diff, add PORTREVISION

fix RESCUE: include lib/libjail/jail.c in librescue if necessary

cap_jail.c: improve allocation and error handling in service command

cap_sysctl.c: resolve names to mibs when limits are set.

• rename cap_jail_get -> cap_jail and system.cap_jail_get -> system.cap_jail
• cap_jail:
  • fix copyright
  • add man page
• cap_jail.c:
  • style(9) changes
  • use dnvlist_* in service command
  • split nvlist -> iov function in two: nvl_to_iov_s is used by the service and makes sure there is space before memcpy
• jls.c: use caph_enter_casper
• cap_sysctl.c: style
• cap_sysclt.3: reference sysctl(3)

I removed to kernel changes and used libcasper to obtain sysctl and jail_get functionality needed for jls(1).

caph_cache_catpages(3) before cap_enter(2)