Fixes:
commit 52a515d187
"udhcp: use poll() instead of select()"
Feb 16 2017
udhcp_sp_read() is meant to check whether signal pipe indeed has some data to read.
In the above commit, it was changed as follows:
- if (!FD_ISSET(signal_pipe.rd, rfds))
+ if (!pfds[0].revents)
return 0;
The problem is, the check was working for select() purely by accident.
Caught signal interrupts select()/poll() syscalls, they return with EINTR
(regardless of SA_RESTART flag in sigaction). _Then_ signal handler is invoked.
IOW: they can't see any changes to fd state caused by signal haldler
(in our case, signal handler makes signal pipe ready to be read).
For select(), it means that rfds[] bit array is unmodified, bit of signal
pipe's read fd is still set, and the above check "works": it thinks select()
says there is data to read.
This accident does not work for poll(): .revents stays clear, and we do not
try reading signal pipe as we should. In udhcpd, we fall through and block
in socket read. Further SIGTERM signals simply cause socket read to be
interrupted and then restarted (since SIGTERM handler has SA_RESTART=1).
Fixing this as follows: remove the check altogether. Set signal pipe read fd
to nonblocking mode. Always read it in udhcp_sp_read().
If read fails, assume it's EAGAIN and return 0 ("no signal seen").
udhcpd avoids reading signal pipe on every recvd packet by looping if EINTR
(using safe_poll()) - thus ensuring we have correct .revents for all fds -
and calling udhcp_sp_read() only if pfds[0].revents!=0.
udhcpc performs much fewer reads (typically it sleeps >99.999% of the time),
there is no need to optimize it: can call udhcp_sp_read() after each poll
unconditionally.
To robustify socket reads, unconditionally set pfds[1].revents=0
in udhcp_sp_fd_set() (which is before poll), and check it before reading
network socket in udhcpd.
TODO:
This might still fail: if pfds[1].revents=POLLIN, socket read may still block.
There are rare cases when select/poll indicates that data can be read,
but then actual read still blocks (one such case is UDP packets with
wrong checksum). General advise is, if you use a poll/select loop,
keep all your fds nonblocking.
Maybe we should also do that to our network sockets?
function old new delta
udhcp_sp_setup 55 65 +10
udhcp_sp_fd_set 54 60 +6
udhcp_sp_read 46 36 -10
udhcpd_main 1451 1437 -14
udhcpc_main 2723 2708 -15
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 2/3 up/down: 16/-39) Total: -23 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
The kernel broke the name years ago, but didn't notice until it was much
too late. Rename the node to match expectations of userland software,
and what the kernel itself documents in its Kconfig help:
This provides a device that's usually called /dev/hwrng, ...
URL: https://marc.info/?l=linux-crypto-vger&m=144249767024990&w=2
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
As usual, by multiplying directories - "dhcpd_eth0", "dhcpd_wlan1"
you can run many servers on different interfaces.
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
The realpath utility requires all paths exist when canonicalizing
symlinks. If /etc/resolv.conf points to a tmpfs, then it might
not exist initially. Use `readlink -f` so that we follow all
symlinks that are available.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Sequential run of concurrent mdev's was too simplistic:
they waited for /dev/mdev.seq to match. This could sometimes
cause cumulative loss of time on the order of a second.
Added SIGCHLD signaling from exiting mdev to all other mdev's.
Added debugging required to see that code actually works as intended.
Example of /dev/mdev.log (with "woken up" elevated from dbg lvl 3 to 2):
mdev[1023]: first seq written
^^^^ seq, not pid
mdev[1023]: 35.022395 ACTION:add SUBSYSTEM:module DEVNAME:(null) DEVPATH:/module/lib80211
mdev[1023]: rule matched, line -1
^^^^^^^ means "default rule"
mdev[1023]: 35.022676 exiting
^^^^^^^^^ second,usec timestamp
mdev[1024]: 35.069691 ACTION:add SUBSYSTEM:vc DEVNAME:vcs9 DEVPATH:/devices/virtual/vc/vcs9
mdev[1024]: dev 7,9
mdev[1025]: 35.069889 waiting for '1024'
mdev[1026]: 35.069946 waiting for '1024'
mdev[1027]: 35.070151 waiting for '1024'
mdev[1024]: rule matched, line -1
mdev[1024]: mknod vcs9 (7,9) 20660 0:0
mdev[1024]: 35.070346 exiting
mdev[1025]: woken up
mdev[1026]: woken up
mdev[1025]: 35.071213 ACTION:add SUBSYSTEM:vc DEVNAME:vcsa9 DEVPATH:/devices/virtual/vc/vcsa9
^^^^^^^^^ took only a millisecond to start running after prev mdev exited
mdev[1025]: dev 7,137
mdev[1027]: woken up
mdev[1025]: rule matched, line -1
mdev[1025]: mknod vcsa9 (7,137) 20660 0:0
mdev[1025]: 35.072109 exiting
function old new delta
mdev_main 849 1372 +523
curtime - 59 +59
dirAction 87 134 +47
static.ts - 8 +8
keywords 19 12 -7
make_device 2189 2119 -70
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Some shells (like dash) are lame and omit the POSIX increment/decrement
feature (because it is listed as optional). Tweak the shell script to
work in all POSIX variants.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Often it is desirable to have /etc/ be on read-only storage (well, the
whole rootfs) but have things like /etc/resolv.conf be symlinks to a
writable location. Tweak the simple script to support that.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
It was colliding with matching of devnames with slashes.
We need a more generic way to examine env.vars in rules anyway.
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
