Friday Oct 30, 2009
Friday Oct 30, 2009
This entry is mostly for newcomers to Solaris/OpenSolaris from UNIX-like systems. When I had been taught about signal() and sigaction() my understanding was that sigaction() is just a superset of signal() and also POSIX conformant but otherwise they accomplish the same thing. This is indeed the case for some of UNIX-like operating systems. In Solaris, as I only recently discovered (to my dismay :)), it's different.
Consider the following code (please ignore the fact it's not strictly checking return values and that the signal handler is not safe):
#include <stdio.h>
#include <signal.h>
#include <unistd.h>
#include <sys/types.h>
void sig_handler(int s) {
printf("Got signal! Sleeping.\n");
sleep(10);
printf("returning from signal handler\n");
}
int main(void) {
struct sigaction s_action;
printf("Setting signal handler: ");
#ifdef POSIX_SIGNALS
printf("sigaction\n");
(void) sigemptyset(&s_action.sa_mask);
s_action.sa_handler = sig_handler;
s_action.sa_flags = 0;
(void) sigaction(SIGHUP, &s_action, (struct sigaction *) NULL);
#else
printf("signal\n");
signal(SIGHUP, sig_handler);
#endif
printf("Waiting for signal\n");
while(1)
pause();
return (0);
}
Now try to compile and run with and without the -DPOSIX_SIGNALS and send 2 SIGHUP signals to the process within the 10 seconds window (so the second signal is received while the signal handler is still running). With sigaction(), the signal will be caught by the handler in both of the cases. With signal() however, the second signal will cause the process to exit. This is because kernel will reset the signal handler to default upon receiving the signal for the first time. This is described in the signal(3C) man page in a somewhat hidden sentence inside the second paragraph (it really pays out to read man pages slowly and with attention to detail):
If signal() is
used, disp is the address of a signal handler, and sig is
not SIGILL, SIGTRAP, or SIGPWR, the system first sets the
signal's disposition to SIG_DFL before executing the signal
handler.
The sigaction(2) man page has this section:
SA_RESETHAND If set and the signal is caught, the dispo-
sition of the signal is reset to SIG_DFL and
the signal will not be blocked on entry to
the signal handler (SIGILL, SIGTRAP, and
SIGPWR cannot be automatically reset when
delivered; the system silently enforces this
restriction).
sigaction() does not set the flag by default which results in the different behavior. I found out that this behavior has been present since Solaris 2.0 or so.
In fact, signal() routine from libc is implemented via sigaction(). From $SRC/lib/libc/port/sys/signal.c:
58 /*
59 * SVr3.x signal compatibility routines. They are now
60 * implemented as library routines instead of system
61 * calls.
62 */
63
64 void(*
65 signal(int sig, void(*func)(int)))(int)
66 {
67 struct sigaction nact;
68 struct sigaction oact;
69
70 CHECK_SIG(sig, SIG_ERR);
71
72 nact.sa_handler = func;
73 nact.sa_flags = SA_RESETHAND|SA_NODEFER;
74 (void) sigemptyset(&nact.sa_mask);
75
76 /*
77 * Pay special attention if sig is SIGCHLD and
78 * the disposition is SIG_IGN, per sysV signal man page.
79 */
80 if (sig == SIGCHLD) {
81 nact.sa_flags |= SA_NOCLDSTOP;
82 if (func == SIG_IGN)
83 nact.sa_flags |= SA_NOCLDWAIT;
84 }
85
86 if (STOPDEFAULT(sig))
87 nact.sa_flags |= SA_RESTART;
88
89 if (sigaction(sig, &nact, &oact) < 0)
90 return (SIG_ERR);
91
92 return (oact.sa_handler);
93 }
I am pretty sure that the SA_RESETHAND flag is set in signal() in order to preserve backwards compatibility.
This means that to solve this problem with signal(), one should set the signal handler again in the signal handler itself. However, this is not a complete solution since there is still a window where the signal can be delivered and the handler is set to SIG_DFL - the default handler which is exit in case of SIGHUP as the signal.h(3HEAD) man page explains in really useful table:
Name Value Default Event
SIGHUP 1 Exit Hangup (see termio(7I))
...
Now let's look at FreeBSD. Its SIGNAL(3) man page contains this separate paragraph:
The handled signal is unblocked when the function returns and the process
continues from where it left off when the signal occurred. Unlike previ-
ous signal facilities, the handler func() remains installed after a sig-
nal has been delivered.
The second sentence is actually printed in bold letters. I also tried on Linux and NetBSD and the behavior is the same as in FreeBSD.
So, to conclude all of the above: using signal() is really not portable.
tags:
opensolaris
sigaction
signal
signals
solaris
Linkage:
Technorati cosmos
Tuesday Sep 01, 2009
Recently I needed to test a bug fix in in.iked(1M) (should say libike.so with which in.iked is linked) after which the daemon should respond to IKEv2 requests with Notification message telling the peer to fall back to IKEv1 (previously it did not respond to IKEv2 packets at all). This can be tested by:
- installing a OS instance which supports IKEv2 and initiating from there
- writing a simple program (C/Perl/etc.) which will construct the UDP payload
Surely, there should be easier way how to send a UDP paket with arbitrary (in my case ISAKMP) payload. It turns out this is very easy to do just from command line with nc(1) which is available in OpenSolaris (install it via 'pkg install SUNWnetcat'). Let's try to send some garbage first to see if it works:
perl -e 'print "\x41\x41";' | nc -u rpe-foo.czech 500
Yep, tshark(1) (in OpenSolaris shipped by default with Wireshark) reports an IKE packet, malformed one (which is not surprising):
Capturing on eri0 0.000000 10.18.144.12 -> 10.18.144.11 ISAKMP [Malformed Packet] 0000 00 0a e4 2f 61 eb 00 03 ba 4e 3d 38 08 00 45 00 .../a....N=8..E. 0010 00 1e 26 98 40 00 ff 11 20 fb 0a 12 90 0c 0a 12 ..&.@... ....... 0020 90 0b e2 66 01 f4 00 0a 34 57 41 41 ...f....4WAA
Our two A's are there just after the UDP header (Ethernet header 14 bytes, IP header 20 bytes, UDP 8 bytes, in sum 42 bytes and our 2 bytes are just after first 8 bytes on 3rd line).
With that we can go and construct IKEv1 packet first to see if the daemon will react upon it. We will need to construct the payload which is a IKEv1 header. IKEv1 is defined in RFC 2409 (The Internet Key Exchange (IKE)). IKEv1 uses ISAKMP header definition so we need to look into RFC 2408 (Internet Security Association and Key Management Protocol (ISAKMP)) for the actual header definition. It's there in section 3.1:
1 2 3
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
! Initiator !
! Cookie !
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
! Responder !
! Cookie !
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
! Next Payload ! MjVer ! MnVer ! Exchange Type ! Flags !
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
! Message ID !
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
! Length !
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
I'd like to construct a packet which resembles first packet sent by IKEv1 Initiator. So, our packet code (similar to shell code) will look like this (without thinking too much of what should the values look like):
- Initiator's cookie, must not be zero
\x11\x22\x33\x44\x55\x66\x77\x88 - Responder's cookie, must be zero in the initial packet from Initiator
\x00\x00\x00\x00\x00\x00\x00\x00 - next payload, let's try 0 first
\x00 - Major and Minor Version (4 bits each)
\x10 - Exchange Type
Exchange Type Value NONE 0 Base 1 Identity Protection 2 Authentication Only 3 Aggressive 4 Informational 5 ISAKMP Future Use 6 - 31 DOI Specific Use 32 - 239 Private Use 240 - 255So let's try Base first:\x01 - Flags (Initiator)
\x00 - Message ID
\x66\x66\x66\x66 - Length
\x28
We need to massage our packet code into command line. The code:
\x11\x22\x33\x44\x55\x66\x77\x88 \x00\x00\x00\x00\x00\x00\x00\x00 \x00 \x10 \x01 \x00 \x66\x66\x66\x66 \x28
We want source port to be 500 as well because of section 2.5.1 in RFC 2408 so use the -p option (this requires the net_privaddr privilege so either become root or use pfexec(1)). Also, we do not need to wait for the response so use -w option:
perl -e 'print "\x11\x22\x33\x44\x55\x66\x77\x88\x00\x00\x00\x00\x00\x00\x00\x00\x00\x10\x01\x00\x66\x66\x66\x66\x28";' \
| nc -w 1 -p 500 -u rpe-foo.czech 500
The packet was received but there was no reply and tshark still considers this as Malformed Packet. Let's check the header again - oh yeah, the Length field has 4 bytes, not just one. Let's try again:
perl -e 'print "\x11\x22\x33\x44\x55\x66\x77\x88\x00\x00\x00\x00\x00\x00\x00\x00\x00\x10\x01\x00\x66\x66\x66\x66\x28\x00\x00\x00";' \
| nc -w 1 -p 500 -u rpe-foo.czech 500
Okay, this is our Base exchange but still not response:
294.029154 10.18.144.12 -> 10.18.144.11 ISAKMP Base 0000 00 0a e4 2f 61 eb 00 03 ba 4e 3d 38 08 00 45 00 .../a....N=8..E. 0010 00 38 26 a7 40 00 ff 11 20 d2 0a 12 90 0c 0a 12 .8&.@... ....... 0020 90 0b 01 f4 01 f4 00 24 34 71 11 22 33 44 55 66 .......$4q."3DUf 0030 77 88 00 00 00 00 00 00 00 00 00 10 01 00 66 66 w.............ff 0040 66 66 28 00 00 00
Let's try something more provocative and set the Exchange type to Identity protection:
perl -e 'print "\x11\x22\x33\x44\x55\x66\x77\x88\x00\x00\x00\x00\x00\x00\x00\x00\x00\x10\x02\x00\x66\x66\x66\x66\x28\x00\x00\x00";' \
| nc -w 1 -p 500 -u rpe-foo.czech 500
Oh yeah, this finally deserved a response:
383.050874 10.18.144.12 -> 10.18.144.11 ISAKMP Identity Protection (Main Mode) 0000 00 0a e4 2f 61 eb 00 03 ba 4e 3d 38 08 00 45 00 .../a....N=8..E. 0010 00 38 26 a8 40 00 ff 11 20 d1 0a 12 90 0c 0a 12 .8&.@... ....... 0020 90 0b 01 f4 01 f4 00 24 34 71 11 22 33 44 55 66 .......$4q."3DUf 0030 77 88 00 00 00 00 00 00 00 00 00 10 02 00 66 66 w.............ff 0040 66 66 28 00 00 00 ff(... 383.051672 10.18.144.11 -> 10.18.144.12 ISAKMP Informational 0000 00 03 ba 4e 3d 38 00 0a e4 2f 61 eb 08 00 45 00 ...N=8.../a...E. 0010 00 99 d3 8b 40 00 ff 11 73 8c 0a 12 90 0b 0a 12 ....@...s....... 0020 90 0c 01 f4 01 f4 00 85 ed 05 11 22 33 44 55 66 ..........."3DUf 0030 77 88 85 75 8e 0f fa a5 5d de 0b 10 05 00 69 a5 w..u....].....i. 0040 63 e4 00 00 00 7d 00 00 00 61 00 00 00 01 01 10 c....}...a...... 0050 00 1e 11 22 33 44 55 66 77 88 85 75 8e 0f fa a5 ..."3DUfw..u.... 0060 5d de 80 0c 00 01 00 06 00 39 55 44 50 20 50 61 ]........9UDP Pa 0070 63 6b 65 74 20 64 6f 65 73 20 6e 6f 74 20 63 6f cket does not co 0080 6e 74 61 69 6e 20 65 6e 6f 75 67 68 20 64 61 74 ntain enough dat 0090 61 20 66 6f 72 20 49 53 41 4b 4d 50 20 70 61 63 a for ISAKMP pac 00a0 6b 65 74 80 08 00 00 ket....
Now that we proved to ourselves that we can construct semi-valid packet it's time to try IKEv2. IKEv2 header is defined in RFC 4306, section 3.1:
1 2 3
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
! IKE_SA Initiator's SPI !
! !
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
! IKE_SA Responder's SPI !
! !
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
! Next Payload ! MjVer ! MnVer ! Exchange Type ! Flags !
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
! Message ID !
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
! Length !
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
On a first sight, it looks the same (for backward compatibility). However, some of the values are different. For IKE header, the main differences are in the Exchange Type and Flags:
Exchange Type Value
RESERVED 0-33
IKE_SA_INIT 34
IKE_AUTH 35
CREATE_CHILD_SA 36
INFORMATIONAL 37
RESERVED TO IANA 38-239
Reserved for private use 240-255
IKE_SA_INIT is our guy ('echo 0t34=x | mdb' produces 0x22).
The flags are now used to indicate the exchange. Set 3rd bit to say we are the Initiator. We will retain the source port even though IKEv2 supports ports other than 500 and 4500 because we're dealing with IKEv1 implementation. Now slightly change our packet code (don't forget to change the Version field to 2.0):
perl -e 'print "\x11\x22\x33\x44\x55\x66\x77\x88\x00\x00\x00\x00\x00\x00\x00\x00\x00\x20\x22\x08\x66\x66\x66\x66\x28\x00\x00\x00";' \
| nc -w 1 -p 500 -u rpe-foo.czech 500
And we got a nice response (since the responder runs recent version libike.so):
1013.190867 10.18.144.12 -> 10.18.144.11 ISAKMP IKE_SA_INIT
0000 00 0a e4 2f 61 eb 00 03 ba 4e 3d 38 08 00 45 00 .../a....N=8..E.
0010 00 38 26 aa 40 00 ff 11 20 cf 0a 12 90 0c 0a 12 .8&.@... .......
0020 90 0b 01 f4 01 f4 00 24 34 71 11 22 33 44 55 66 .......$4q."3DUf
0030 77 88 00 00 00 00 00 00 00 00 00 20 22 08 66 66 w.......... ".ff
0040 66 66 28 00 00 00 ff(...
1013.192005 10.18.144.11 -> 10.18.144.12 ISAKMP Informational
0000 00 03 ba 4e 3d 38 00 0a e4 2f 61 eb 08 00 45 00 ...N=8.../a...E.
0010 00 83 d3 8d 40 00 ff 11 73 a0 0a 12 90 0b 0a 12 ....@...s.......
0020 90 0c 01 f4 01 f4 00 6f 66 da 11 22 33 44 55 66 .......of.."3DUf
0030 77 88 5c 36 e3 75 a2 7b 8e fe 0b 10 05 00 87 03 w.\6.u.{........
0040 0c f5 00 00 00 67 00 00 00 4b 00 00 00 01 01 10 .....g...K......
0050 00 05 11 22 33 44 55 66 77 88 5c 36 e3 75 a2 7b ..."3DUfw.\6.u.{
0060 8e fe 80 0c 00 01 00 06 00 23 49 6e 76 61 6c 69 .........#Invali
0070 64 20 49 53 41 4b 4d 50 20 6d 61 6a 6f 72 20 76 d ISAKMP major v
0080 65 72 73 69 6f 6e 20 6e 75 6d 62 65 72 80 08 00 ersion number...
0090 00
The only thing which is not nice is our terminal since nc(1) dumped the binary packet to it. Let's try again with some post-processing:
# perl -e 'print "\x11\x22\x33\x44\x55\x66\x77\x88\x00\x00\x00\x00\x00\x00\x00\x00\x00\x20\x22\x08\x66\x66\x66\x66\x28\x00\x00\x00";' | nc -w 1 -p 500 -u rpe-foo.czech 500 | od -c 0000000 021 " 3 D U f w 210 237 y 254 264 351 333 007 344 0000020 013 020 005 \0 [ 251 244 j \0 \0 \0 g \0 \0 \0 K 0000040 \0 \0 \0 001 001 020 \0 005 021 " 3 D U f w 210 0000060 237 y 254 264 351 333 007 344 200 \f \0 001 \0 006 \0 # 0000100 I n v a l i d I S A K M P m 0000120 a j o r v e r s i o n n u m 0000140 b e r 200 \b \0 \0 0000147
The fix is obviously in place.
tags:
construction
factory
ike
nc
netcat
packet
solaris
Linkage:
Technorati cosmos
Monday May 04, 2009
I have just integrated couple of changes which I believe are the first contributed externally to the Testing community as an open-source contribution. The changes add couple of new tests to the nc test suite to cover the enhancement described in PSARC/2008/680 (which is present in Nevada since build 106). This is the stuff which allows you to run nc(1) in client mode with complex portlist specifications. Previously it was possible only to use simple port ranges like 22-80, with this change one can connect to e.g. 22,24,50-80,66,1024-2048. Little example how it might be useful:
$ nc -v -z grok.czech 22,25,80-88,8080 Connection to grok.czech 22 port [tcp/ssh] succeeded! nc: connect to 129.157.71.49 port 25 [host grok.czech] (tcp) failed: Connection refused Connection to grok.czech 80 port [tcp/*] succeeded! nc: connect to 129.157.71.49 port 81 [host grok.czech] (tcp) failed: Connection refused nc: connect to 129.157.71.49 port 82 [host grok.czech] (tcp) failed: Connection refused nc: connect to 129.157.71.49 port 83 [host grok.czech] (tcp) failed: Connection refused nc: connect to 129.157.71.49 port 84 [host grok.czech] (tcp) failed: Connection refused nc: connect to 129.157.71.49 port 85 [host grok.czech] (tcp) failed: Connection refused nc: connect to 129.157.71.49 port 86 [host grok.czech] (tcp) failed: Connection refused nc: connect to 129.157.71.49 port 87 [host grok.czech] (tcp) failed: Connection refused nc: connect to 129.157.71.49 port 88 [host grok.czech] (tcp) failed: Connection refused Connection to grok.czech 8080 port [tcp/*] succeeded!
Back to the testing part. The putback (yes, stcnv-gate is still using Teamware) log for this change looks like this (I have modified Erik's e-mail a bit):
6786859 portranges_complex_spec is missing the listener 6754842 extended port list specification needs to be tested Code contributed by Erik Trauschke <erik.trauschke AT freenet.de>
I think this is really nice example of the ideal state - the contributor not only did the feature part but also the testing part. It shows a great degree of responsibility - not just throwing some code "over the fence" but fully participating in the process to ensure the quality even in the long term.
The tests are both positive and negative. Each purpose in portranges directory is numbered and the following numbers match the test purpose numbers:
- 5-12 ensure nc treats ports 0 and 65536 as invalid
Previously, it was possible to listen on ports 0 and 65536, with Erik's changes this is no longer true so we need to add regression tests for both cases (client/server) and both ports. - 13-19 see if various malformed port list specifications are considered as invalid
Each purpose needs not only positive tests which make sure the functionality actually works but also negative tests which ensure it does not misbehave. In this case, invalid port list specifications are thrown at nc(1) to see it reacts accordingly (with error, that is). - 20-25 test the functionality of complex port lists
This is the bunch of test which see if the functionality actually works. - 26 tests reallocation
Since the internal representation of the port list is now dynamically allocated and there is a default pre-allocated value which is reallocated if needed we need to test the case of reallocation.
To be able to do such integration there is now a Test development process. It's similar to the process used in ON community but it's more lightweight. The main difference is that the request-sponsor part is done informally via the testing-discuss mailing list and there is no list of bugs to pick up from. But don't be shy, whether you're adding new functionality or completely new program, the Testing community is here to help you.
tags:
contribute
contributions
nc
netcat
opensolaris
solaris
testing
Linkage:
Technorati cosmos
Tuesday Mar 31, 2009
Since the days when John Beck added command line editing to zonecfg Mark Phalan did similar thing to Kerberos utilities and Huie-Ying Lee to sftp. IPsec utilities (ipseckey(1M) and ikeadm(1M)) offered the ability to enter commands in interactive mode for a long time but only since Nevada build 112, the commands support command line editing and history too. Again, thanks to libtecla (shipped with Solaris/OpenSolaris).
Lessons learned:
- adding full-blown command line editing support is hard.
Adding the initial support is quite easy. However, more advanced features could require substantial work. This is especially true for tab completion. For sftp Huie-Ying decided to add tab completion in the future phase because of the ambiguities when completing names of files (when to complete local files versus remote files).
I did the same with tab completion for IPsec utilities - the integration only delivers basic command line editing support, without tab completion. The problem with ipseckey(1M) and ikeadm(1M) is that their grammar is quite bifurcated and has contexts. For example, you cannot use encr_alg with AH SAs in ipseckey. Or, it would be erroneous to tab complete a valid command in the middle of entering a key if the key hex sequence contained sub-string of a valid command. The hardest part is I think offering the right tables of valid commands in given context. E.g. in our case a command line in our case must start with top-level command. Each top-level command offers several valid sub-commands and we do not offer invalid sub-commands for given top-level command so there is a necessity to track the state of the finite state machine describing the grammar contexts.
Also, after the user entered src we do not want to allow him to enter it again on the same command line. Also, if the user already entered say add esp spi we are expecting SPI number, not a command name.
Ideally, to solve this problem in nice way there should be a meta library (or additional API in libtecla) which would offer the ability to link command tables and set the contexts. - interruptible cycles in command line mode
ipseckey's monitor command reads from a PF_KEY socket in a loop. The loop is normally interruptible by SIGINT. To do so in libtecla environment (we do not want to exit the command line upon SIGINT and yet still need to interrupt the cycle), something like this is needed:static void monitor_catch(int signal) { if (!interactive) errx(signal, gettext("Bailing on signal %d."), signal); } void doreadcycle(void) { ... /* Catch ^C. */ newsig.sa_handler = monitor_catch; newsig.sa_flags = 0; (void) sigemptyset(&newsig.sa_mask); (void) sigaddset(&newsig.sa_mask, SIGINT); (void) sigaction(SIGINT, &newsig, &oldsig); for (; ; ) { rc = read(keysock, samsg, sizeof (get_buffer)); /* handle the data */ } /* restore old behavior */ if (interactive) (void) sigaction(SIGINT, &oldsig, NULL); } - interaction with SMF
While it's fine to bail out in interactive mode with error, due to the nature of IPsec commands (they can read the config files using the same routines as for interactive mode and they are used as SMF services to bring up IPsec policy and keys after boot) we need to distinguish the interactive and non-interactive mode. - maximum command line history value
It seems that the second parameter to new_GetLine() - histlen is commonly misunderstood. This variable does not express the number of maximum lines in the history but instead maximum size of the history buffer in bytes. If the buffer becomes full, libtecla does not trim the last line but shifts instead.
Given the first parameter to new_GetLine() expresses maximum command line size (in bytes) one needs to do some calculations and estimates on what will be needed too avoid too big buffer - ipseckey is used to enter key material so the line could become quite long. Say we wanted to keep 1024 lines. If the maximum length of the line is 1024 this will give us 1 megabyte buffer which seems too much for a simple application. Thus I did some guessing and set the buffer size accordingly:
For "common" ipseckey configuration commands (think moderately bifurcated 'add') it's cca 300 characters. Mostly however, the users enter query commands like 'flush esp', 'dump ah' and the like so this is somewhere around say 30 characters. Say 30% of the commands are configuration and the rest is queries. To hold 100 such commands only cca 10K memory is required. In the end I chose 64K to be able to hold 15 of the biggies (4K) commands.
tags:
command
history
ipsec
line
opensolaris
solaris
Linkage:
Technorati cosmos
Friday Nov 07, 2008
I will start this one a little bit generically..
Part of standard OpenSolaris/Solaris development process is code review. To facilitate a review, a so-called webrev is needed. A webrev is set of HTML/text/PDF pages and documents which display all the changes between local repository containing the changes and its parent repository. To produce a webrev, simply switch to a repository and run the webrev script (it is part of SUNWonbld package, which can be downloaded from OpenSolaris download center.):
$ cd /local/Source/bugfix-314159.onnv $ webrev
Assuming /opt/onbld/bin is present in your PATH a webrev will be generated under /local/Source/bugfix-314159.onnv/webrev/ directory.
For OpenSolaris changes, the webrev is usually uploaded to cr.opensolaris.org
(every OpenSolaris member has an account automatically created for him) which serves it under http://cr.opensolaris.org/~OSol_username/
(where OSol_username is your OpenSolaris username) and a request for review with a link to the webrev is sent
to one of the mailing lists relevant to the change.
Dan Price has written a script which produces
RSS feed out of recently uploaded webrevs which is pretty handy substitute for feeds from news/headlines/magazines :)
For a long time I was basically doing the following:
$ cd /local/Source/bugfix-foo.onnv && webrev $ scp -r webrev cr.opensolaris.org:bugfix-foo.onnv
This had two flaws: first it was slow (because of rcp protocol over SSH channel) and second I had to delete it via separate command (use sftp(1) and rename the old webrev to .trash directory) before uploading new version of the webrev (otherwise couple of permissions errors would follow).
To solve the first problem, rsync (with SSH transport) can be used which makes the upload nearly instantaneous. Second problem can be worked around by using incremental webrevs. Still, this does not seem good enough for code reviews with many iterations.
So, the change made in CR 6752000 introduces the following command line options for automatic webrev upload:
- -U uploads the webrev
- -n suppresses webrev generation
- -t allows to specify custom upload target
webrev.1 man page has been updated to explain the usage. For common OpenSolaris code reviews the usage will probably mostly look like this:
$ webrev -O -U
This will upload the webrev to cr.opensolaris.org under directory named according to local repository name. Further invocations
will replace the remote webrev with fresh version.
But it is possible to get more advanced. After the initial webrev is posted, an incremental webrev can be both generated and posted.
Assuming you're switched to the repository (via bldenv) and we're dealing with 4th round of code review the following
command will perform the task:
webrev_name=`basename $CODEMGR_WS`
webrev -O -U -o $CODEMGR_WS/${webrev_name}.rd4 \
-p $CODEMGR_WS/${webrev_name}.rd3
The above commands hide maybe not-so-obvious behavior so I'll try to explain it in the table:
+---------------------------+------------------------+-----------------------------------------------------+ | command | local webrev directory | remote webrev directory | +---------------------------+------------------------+-----------------------------------------------------+ | webrev -O -U | $CODEMGR_WS/webrev/ | cr.opensolaris.org/~OSOLuser/`basename $CODEMGR_WS` | +---------------------------+------------------------+-----------------------------------------------------+ | webrev -O -o \ | $CODEMGR_WS/my_webrev/ | cr.opensolaris.org/~OSOLuser/my_webrev | | $CODEMGR_WS/my_webrev | | | +---------------------------+------------------------+-----------------------------------------------------+ | webrev -O \ | $CODEMGR_WS/fix.rd2/ | cr.opensolaris.org/~OSOLuser/fix.rd2 | | -p $CODEMGR_WS/fix.rd1 \ | | | | -o $CODEMGR_WS/fix.rd2 | | | +---------------------------+------------------------+-----------------------------------------------------+
Basically, without the -o flag webrev will generate the webrev to local directory named 'webrev' but it will upload it to the directory named after basename of local repository. With the -o flag webrev will use the name of root directory of the repository it is called from for both local and remote storage. This is done to keep the default behavior of generating local webrev to directory named 'webrev'. At the same time, uploading different webrevs to the same remote directory named 'webrev' does not make sense.
NOTE: This behavior is also valid in the case when not enclosed in a repository via ws or bldenv, I have just used $CODEMGR_WS to express root directory of a workspace/repository.
Also, now it is possible to call webrev from within Cadmium Mercurial plugin, so all webrev commands can be prefixed with hg.
All in all, it was fun dealing with webrevs of webrev. I am looking forward to more entries in the RSS feed :)NOTE: It will take some time before the changes appear in SUNWonbld packages offered by the download center so it's better to update the sources from the ssh://anon@hg.opensolaris.org/hg/onnv/onnv-gate repository and build and upgrade the SUNWonbld package from there.
tags:
automatic
opensolaris
solaris
upload
webrev
Linkage:
Technorati cosmos
Sunday Apr 13, 2008
During nc(1) preintegration testing, short time before it went back
I had found that 'cat /etc/passwd | nc localhost 4444' produced endless loop with
100% CPU utilization, looping in calls doing poll(2) (I still remember my laptop suddenly getting much warmer
than it should be and CPU fan cranking up).
'nc localhost 4444 < /etc/password' was not exhibiting that behavior.
The cause was a difference between poll(2) implementation on BSD and Solaris. Since I am working on Netcat
in Solaris again (adding more features, stay tuned), it's time to take a look back
and maybe even help people porting similar software from BSD to Solaris.
The issue appears because POLLHUP is set in read events bitfield for stdin after pipe is closed (or to be more precise - after the producer/write end is done) on Solaris. poll.c (which resembles readwrite() function from nc) illustrates the issue:
01 #include <stdio.h>
02 #include <poll.h>
03
04 #define LEN 1024
05
06 int main(void) {
07 int timeout = -1;
08 int n;
09 char buf[LEN];
10 int plen = LEN;
11
12 struct pollfd pfd;
13
14 pfd.fd = fileno(stdin);
15 pfd.events = POLLIN;
16
17 while (pfd.fd != -1) {
18 if ((n = poll(&pfd, 1, timeout)) < 0) {
19 err(1, "Polling Error");
20 }
21 fprintf(stderr, "revents = 0x%x [ %s %s ]\n",
22 pfd.revents,
23 pfd.revents & POLLIN ? "POLLIN" : "",
24 pfd.revents & POLLHUP ? "POLLHUP" : "");
25
26 if (pfd.revents & (POLLIN|POLLHUP)) {
27 if ((n = read(fileno(stdin), buf, plen)) < 0) {
28 fprintf(stderr,
29 "read() returned neg. val (%d)\n", n);
30 return;
31 } else if (n == 0) {
32 fprintf(stderr, "read() returned 0\n", n);
33 pfd.fd = -1;
34 pfd.events = 0;
35 } else {
36 fprintf(stderr, "read: %d bytes\n", n);
37 }
38 }
39 }
40 }
Running it on NetBSD (chosen because my personal non-work mailbox is hosted on a machine running it) produces the following:
otaku[~]% ( od -N 512 -X -v /dev/zero | sed 's/ [ \t]*/ /g'; sleep 3 ) | ./poll revents = 0x1 [ POLLIN ] read: 1024 bytes revents = 0x1 [ POLLIN ] read: 392 bytes revents = 0x11 [ POLLIN POLLHUP ] read() returned 0
I had to post-process the output of od(1) (because of difference between output of od(1) on NetBSD and Solaris) and slow the execution down a bit (via sleep) in order to make things more visible (try to run the command without the sleep and the pipe will be closed too quickly). On OpenSolaris the same program produces different pattern:
moose:~$ ( od -N 512 -X -v /dev/zero | sed 's/ [ \t]*/ /g' ; sleep 3 ) | ./poll revents = 0x1 [ POLLIN ] read: 1024 bytes revents = 0x1 [ POLLIN ] read: 392 bytes revents = 0x10 [ POLLHUP ] read() returned 0
So, the program is now obviously correct. Had the statement on line 26 checked only POLLIN, the command above (with or without the sleep) would go into endless loop on Solaris:
revents = 0x11 [ POLLIN POLLHUP ] read: 1024 bytes revents = 0x11 [ POLLIN POLLHUP ] read: 392 bytes revents = 0x10 [ POLLHUP ] revents = 0x10 [ POLLHUP ] revents = 0x10 [ POLLHUP ] ...
Both OSes set POLLHUP after the pipe is closed. The difference is that while BSD always
indicates POLLIN (even if there is nothing to read), Solaris strips it after data stream ended.
So, which one is correct ?
poll() function as
described by OpenGroup says that
"POLLHUP and POLLIN are not mutually exclusive
".
This means both implementations seem to conform to the IEEE Std 1003.1, 2004 Edition standard
(part of POSIX) in this respect.
However, the POSIX standard also says:
-
In each pollfd structure, poll ( ) shall clear the revents member, except that where the application
requested a report on a condition by setting one of the bits of events listed above, poll ( ) shall set
the corresponding bit in revents if the requested condition is true. In addition, poll ( ) shall set the
POLLHUP, POLLERR, and POLLNVAL flag in revents if the condition is true, even if the
application did not set the corresponding bit in events.
This might be still ok even though POLLIN flag remains to be set in NetBSD's poll() even after no data are available for reading (try to comment out lines 33,34 and run as above) because the standard says about POLLIN flag: For STREAMS, this flag is set in revents even if the message is of zero length.
Without further reading it is hard to tell how exactly should POSIX compliant poll() look like. On the Austin group mailing list there was a thread about poll() behavior w.r.t. POLLHUP suggesting this is fuzzy area.
Anyway, to see where exactly is POLLHUP set for pipes in OpenSolaris go to fifo_poll(). The function _sets_ the revents bit field to POLLHUP so the POLLIN flag is wiped off after that. fifo_poll() is part of fifofs kernel module which has been around in Solaris since late eighties (I was still in elementary school the year fifovnops.c appeared in SunOS code base :)). NetBSD has fifofs too but the POLLHUP flag gets set via bit logic operation in pipe_poll() which is part of syscall processing code. The difference between OpenSolaris and NetBSD (whoa, NetBSD project uses OpenGrok !) POLLHUP attitude (respectively) is now clear:
tags:
opengrok
opensolaris
poll
pollhup
posix
solaris
syscall
Linkage:
Technorati cosmos
Friday Apr 04, 2008
The flashback is still alive even weeks after: the day before my presentation at FIRST Technical
Colloquium in Prague I brought my 2 years old laptop with the work-in-progress slides to the office.
Since I wanted to finish the slides in the evening a live-upgrade
process was fired off on the laptop to get fresh Nevada
version. (of course, to show off during the presentation ;))
LU
is very I/O intensive process and the red Ferrari notebooks tend to get _very_ hot. In the afternoon I
noticed that the process failed. To my astonishment, the I/O operations started to fail.
After couple of reboots (and zpool status / fmadm faulty commands)
it was obvious that the disk cannot be trusted anymore. I was able to rescue some data
from the ZFS pool which was spanning the biggest slice of the internal disk but not all data.
(ZFS is not willing to get corrupted data out.) My slides were lost as well as other data.
After some time I stumbled upon James Gosling's
blog entry about ZFS mirroring on laptop.
This get me started (or more precisely I was astonished and wondered how is it possible
that this idea escaped me because at that time ZFS had been in Nevada for a long time)
and I have discovered several similar and more
in-depth
blog entries about the topic.
After some experiments with borrowed USB disk it was time to make it reality
on a new laptop.
The process was a multi-step one:
-
First I had to extend the free slice #7 on the internal disk so it spans the remaining space
on the disk because it was trimmed after the experiments. In the end the slices look like this
in format(1) output:
Part Tag Flag Cylinders Size Blocks 0 root wm 3 - 1277 9.77GB (1275/0/0) 20482875 1 unassigned wm 1278 - 2552 9.77GB (1275/0/0) 20482875 2 backup wm 0 - 19442 148.94GB (19443/0/0) 312351795 3 swap wu 2553 - 3124 4.38GB (572/0/0) 9189180 4 unassigned wu 0 0 (0/0/0) 0 5 unassigned wu 0 0 (0/0/0) 0 6 unassigned wu 0 0 (0/0/0) 0 7 home wm 3125 - 19442 125.00GB (16318/0/0) 262148670 8 boot wu 0 - 0 7.84MB (1/0/0) 16065 9 alternates wu 1 - 2 15.69MB (2/0/0) 32130
-
Then the USB drive was connected to the system and recognized via format(1):
AVAILABLE DISK SELECTIONS: 0. c0d0/pci@0,0/pci-ide@12/ide@0/cmdk@0,0 1. c5t0d0 /pci@0,0/pci1025,10a@13,2/storage@4/disk@0,0 - Live upgrade boot environment which was not active was deleted via ludelete(1M) and the slice was commented out in /etc/vfstab. This was needed to make zpool(1M) happy.
-
ZFS pool was created out of the slice on the internal disk (c0d0s7) and external
USB disk (c5t0d0). I had to force it cause zpool(1M) complained about the overlap
of c0d0s2 (slice spanning the whole disk) and c0d0s7:
# zpool create -f data mirror c0d0s7 c5t0d0
For a while I have struggled with finding a name for the pool (everybody seems either to stick to the 'tank' name or come up with some double-cool-stylish name which I wanted to avoid because of the likely degradation of the excitement from that name) but then chosen the ordinary data (it's what it is, after all). - I have verified that it is possible to disconnect the USB disk
and safely connect it while an I/O operation is in progress:
root:moose:/data# mkfile 10g /data/test & [1] 10933 root:moose:/data# zpool status pool: data state: ONLINE scrub: none requested config: NAME STATE READ WRITE CKSUM data ONLINE 0 0 0 mirror ONLINE 0 0 0 c0d0s7 ONLINE 0 0 0 c5t0d0 ONLINE 0 0 0 errors: No known data errors
It survived it without a hitch (okay, I had to wait for the zpool command to complete a little bit longer due to the still ongoing I/O but that was it) and resynced the contents automatically after the USB disk was reconnected:root:moose:/data# zpool status pool: data state: DEGRADED status: One or more devices are faulted in response to persistent errors. Sufficient replicas exist for the pool to continue functioning in a degraded state. action: Replace the faulted device, or use 'zpool clear' to mark the device repaired. scrub: resilver in progress for 0h0m, 3.22% done, 0h5m to go config: NAME STATE READ WRITE CKSUM data DEGRADED 0 0 0 mirror DEGRADED 0 0 0 c0d0s7 ONLINE 0 0 0 c5t0d0 FAULTED 0 0 0 too many errors errors: No known data errors
Also, with heavy I/O it is needed to mark the zpool as clear after the resilver completes via zpool clear data because the USB drive is marked as faulty. Normally this will not happen (unless the drive really failed) because I will be connecting and disconnecting the drive only when powering on or shutting down the laptop, respectively. - After that I have used
Mark Shellenbaum's
blog entry
about ZFS delegated administration (it was Mark who did
the integration)
and ZFS Delegated Administration
chapter from OpenSolaris
ZFS Administration Guide and created permissions set for my local user and assigned
those permissions to the ZFS pool 'data' and the user:
# chmod A+user:vk:add_subdirectory:fd:allow /data # zfs allow -s @major_perms clone,create,destroy,mount,snapshot data # zfs allow -s @major_perms send,receive,share,rename,rollback,promote data # zfs allow -s @major_props copies,compression,quota,reservation data # zfs allow -s @major_props snapdir,sharenfs data # zfs allow vk @major_perms,@major_props data
All of the commands had to be done under root. - Now the user is able to create a home directory for himself:
$ zfs create data/vk
- Time to setup the environment of data sets and prepare it for data.
I have separated the data sets according to a 'service level'. Some data are very
important (e.g. presentations ;)) to me so I want them multiplied via the
ditto blocks
mechanism so they are actually present 4 times in case of copies dataset
property set to 2. Also, documents are not usually accompanied by executable
code so the exec property was set to off which will prevent
running scripts or programs from that dataset.
Some data are volatile and in high quantity so they do not need any additional protection and it is good idea to compress them with better compression algorithm to save some space. The following table summarizes the setup:dataset properties comment +-------------------+--------------------------------+------------------------+ | data/vk/Documents | copies=2 | presentations | | data/vk/Saved | compression=on exec=off | stuff from web | | data/vk/DVDs | compression=gzip-8 exec=off | Nevada ISOs for LU | | data/vk/CRs | compression=on copies=2 | precious source code ! | +-------------------+--------------------------------+------------------------+
So the commands will be:$ zfs create -o copies=2 data/vk/Documents $ zfs create -o compression=gzip-3 -o exec=off data/vk/Saved ...
- Now it is possible to migrate all the data, change home directory of the user to /data/vk (e.g. via /usr/ucb/vipw) and relogin.
However, this is not the end of it but just beginning. There are many things to make the setup even better, to name a few:
- setup some sort of automatic snapshots for selected datasets
The set of scripts and SMF service for doing ZFS snapshots and backup (see ZFS Automatic For The People and related blog entries) made by Tim Foster could be used for this task. - make zpool scrub run periodically
- detect failures of the disks
This would be ideal to see in Gnome panel or Gnome system monitor. - setup off-site backup via
SunSSH + zfs send
This could be done using the hooks provided by Tim's scripts (see above). - Set quotas and reservations for some of the datasets.
- Install ZFS scripts for Gnome nautilus so I will be able to browse, perform and destroy snapshots in nautilus. Now which set of scripts to use ? Chris Gerhard's or Tim Foster's ? Or should I just wait for the official ZFS support for nautilus to be integrated ?
- Find how exactly will the recovery scenario (in case of laptop drive failure) will look like.
To import the ZFS pool from the USB disk should suffice but during my experiments I was not able to complete it successfully.
With all the above the data should be safe from disk failure (after all disks are often called "spinning rust" so they are going to fail sooner or later) and also the event of loss of both laptop and USB disk.
Lastly, a philosophical thought: One of my colleagues considers hardware as a necessary (and very faulty) layer which is only needed to make it possible to express the ideas in software. This might seem extreme but come to think of it. ZFS is special in this sense - being a software which provides that bridge, it's core idea to isolate the hardware faults.
tags:
laptop
mirror
notebook
opensolaris
raid
solaris
usb
zfs
Linkage:
Technorati cosmos
Friday Feb 15, 2008
Two weeks ago (yeah, I am a slacker) FIRST technical colloquium was held in Prague and we (me and Sasha) were given the opportunity to attend (the fact the Derrick serves as FIRST chair in the steering comittee has of course something to do with it).
I only attended one day of the technical colloquium (Tuesday 29th).
The day was filled with various talks and presentations. Most of them were
performed by various CERT teams members from around the world. This was because
this event was a joint meeting of FIRST and TF-CSIRT.
It was definitely
interesting to see very different approaches to the shared problem set (dealing with
incidents, setting up honey pots, building forensic analysis labs, etc.).
Not only these differences stemmed from sizes of the networks and organizations
but also (and that was kind of funny) from nationalities.
In the morning I talked about the integration of
Netcat into Solaris,
describing the process, current features and planned enhancements and extensions.
The most anticipated talk was by Adam Laurie who is entertaining guy involved in many hacker-like activities (see e.g. A hacker games the hotel
article by Wired) directed at proving insecurities in
many publicly used systems.
Adam (brother of Ben Laurie, author of Apache-SSL
and OpenSSL contributor) first started with intro about satellite scanning, insecure hotel safes
(with backdoors installed by the manufacturers which can be overcome by a screwdriver).
Then he proceeded to talk about RFID chips, mainly about cloning.
Also, at the "social event" in the evening I had the pleasure to share a table with Ken van Wyk who is overall cool fellow and the author of Secure coding and Incident response books from O'Reilly.
In overall, it was interesting to see so many security types in a room and get to know some of them.
tags:
cert
conference
first
netcat
prague
security
solaris
tf-csirt
Linkage:
Technorati cosmos
Monday Feb 11, 2008
I have been working on a tough bug for some non-trivial time. The bug is a combination of race condition and data consistency issues. To debug this I am using multi-threaded apache process and dtrace heavily. The logs produced by dtrace are huge and contain mostly dumps of internal data structures.
The excerpt from such log looks e.g. like this:
1 50244 pk11_return_session:return 128 8155.698 1 50223 pk11_RSA_verify:entry 128 8155.7069 1 50224 pk11_get_session:entry 128 8155.7199 1 50234 pk11_get_session:return 128 8155.7266 PK11_SESSION: pid = 1802 session handle = 0x00273998 rsa_pub_key handle -> 0x00184e70 rsa_priv_key handle -> 0x00274428 rsa_pub = 0x00186248 rsa_priv = 0x001865f8 1 50224 pk11_get_session:entry 128 8155.7199 1 50244 pk11_return_session:return 128 8155.698
Side note: This is post-processed log (probe together with their bodies are timestamp sorted, time stamps are converted to miliseconds - see Coloring dtrace output entry for details and scripts).
Increasingly, I was asking myself questions which resembled this one: "when function foo() was called with data which contained value bar ?"
This quickly lead to a script which does the tedious job for me. dtrace-grep.pl accepts 2 or 3 parameters. First 2 are probe pattern and data pattern, respectively. Third, which is optional, is input file (if not supplied, stdin will be used). Example of use on the above pasted file looks like this:
~/bin$ ./dtrace-grep.pl pk11_get_session 0x00186248 /tmp/test 1 50234 pk11_get_session:return 128 8155.7266 PK11_SESSION: pid = 1802 session handle = 0x00273998 rsa_pub_key handle -> 0x00184e70 rsa_priv_key handle -> 0x00274428 rsa_pub = 0x00186248 rsa_priv = 0x001865f8 ~/bin$
Now I have to get back to work, to do some more pattern matching. Oh, and the script is here.
tags:
dtrace
grep
logs
solaris
Linkage:
Technorati cosmos
Thursday Jan 17, 2008
It seems that many developers and dtrace users found themselves in a position where they wanted to add some SDT probes to a module to get more insight into what's going on but the had to pause and were thinking "okay, more probes. But where to put them ? Do I really need the additional probes when I already have the fbt ones ?". To do this, systematic approach is needed in order not to over-do or under-do. I will use KSSL (Solaris kernel SSL proxy [1]) for illustration.
With CR 6556447, tens of SDT probes were introduced into KSSL module and other modules which interface with it. Also, in addition to the new SDT probes, in KSSL we got rid of the KSSL_DEBUG macros compiled only in DEBUG kernels and substituted them with SDT probes. As a result, much better observability and error detection was achieved with both debug and non-debug kernels. The other option would be to create KSSL dtrace provider but that would be too big gun for what is needed to achieve.
Generically, the following interesting data points for data gathering/observation can be identified in code:
- data paths
When there is a more than one path how data could flow into a subsystem. E.g. for TCP we have couple of cases how SSL data could reach KSSL input queue. To identify where from exactly was tcp_kssl_input() called we use SDT probes:if (tcp->tcp_listener || tcp->tcp_hard_binding) { ... if (tcp->tcp_kssl_pending) { DTRACE_PROBE1(kssl_mblk__ksslinput_pending, mblk_t *, mp); tcp_kssl_input(tcp, mp); } else { tcp_rcv_enqueue(tcp, mp, seg_len); } } else { ... /* Does this need SSL processing first? */ if ((tcp->tcp_kssl_ctx != NULL) && (DB_TYPE(mp) == M_DATA)) { DTRACE_PROBE1(kssl_mblk__ksslinput_data1, mblk_t *, mp); tcp_kssl_input(tcp, mp); } else { putnext(tcp->tcp_rq, mp); if (!canputnext(tcp->tcp_rq)) tcp->tcp_rwnd -= seg_len; } ... - data processed in while/for cycles
To observe what happens in each iteration of the cycle. Can be used in code like this:while (mp != NULL) { DTRACE_PROBE1(kssl_mblk__handle_record_cycle, mblk_t *, mp); /* process the data */ ... mp = mp->b_cont; } - switch statements
If significant/non-trivial processing happens inside switch it may be useful to add SDT probes there too. E.g.:content_type = (SSL3ContentType)mp->b_rptr[0]; switch(content_type) { /* select processing according to type */ case content_alert: DTRACE_PROBE1(kssl_mblk__content_alert, mblk_t *, mp); ... break; case content_change_cipher_spec: DTRACE_PROBE1(kssl_mblk__change_cipher_spec, mblk_t *, mp); ... break; default: DTRACE_PROBE1(kssl_mblk__unexpected_msg, mblk_t *, mp); break; } - labels which cannot be (easily) identified in other way
Useful if code which follows the label is generic (assignments, no function calls), e.g.:/* * Give this session a chance to fall back to * userland SSL */ if (ctxmp == NULL) goto no_can_do; ... no_can_do: DTRACE_PROBE1(kssl_no_can_do, tcp_t *, tcp); listener = tcp->tcp_listener; ind_mp = tcp->tcp_conn.tcp_eager_conn_ind; ASSERT(ind_mp != NULL);
You've surely noticed that same of the probe definitions above have common prefix (kssl_mblk-). This is one of the things which make SDT probes so attractive. With prefixes it is possible to do the e.g. following:
sdt:::kssl_err-*
{
trace(timestamp);
printf("hit error in %s\n", probefunc);
stack(); ustack();
}
The important part is that we do not specify module of function name. The implicit wildcard (funcname/probename left out) combined with explicit wildcard (prefix + asterisk) will lead to all KSSL error probes to be activated regardless of in which module or function there are defined. This is obviously very useful for technologies which span multiple Solaris subsystems or modules (such as KSSL).
The nice thing about the error probes is that they could be leveraged in test suites. For each test case we can first run dtrace script with the above probeset covering all KSSL errors in the background and after the test completes just check if it produced some data. If it did, then the test case can be considered as failed. No need to check kstat(1M) (and other counters), log files, etc.
Also, thanks to the way how dtrace probes are activated we can have both generic probeset (using this for lack of better term) as above with addition of probe specific action, e.g.:
/* probeset of all KSSL error probes */
sdt:::kssl_err-*
{
trace(timestamp);
printf("hit error in %s\n", probefunc);
}
/*
the probe definition is:
DTRACE_PROBE2(kssl_err__bad_record_size,
uint16_t, rec_sz, int, spec->cipher_bsize);
*/
sdt:kssl:kssl_handle_record:kssl_err-bad_record_size
{
trace(timestamp);
tracemem(arg0, 32);
printf("rec_sz = %d , cipher_bsize = %d\n", arg1, arg2);
}
If probe kssl_err-bad_record_size gets activated the generic probe will be activated (and fires) too because the probeset contains the probe.
Similarly to the error prefix, we can have data specific prefix. For KSSL it is kssl_mblk- prefix which could be used for tracing all mblks (msgb(9S)) as they flow through TCP/IP, STREAMS and KSSL modules. With such probes it is possible to do e.g. the following:
/* how many bytes from a mblk to dump */ #define DUMP_SIZE 48 /* use macros from*/ #define MBLKL(mp) ((mp)->b_wptr - (mp)->b_rptr) #define DB_FLAGS(mp) ((mp)->b_datap->db_flags) #define PRINT_MBLK_INFO(mp) \ printf("mblk = 0x%p\n", mp); \ printf("mblk size = %d\n", MBLKL((mblk_t *)mp)); \ PRINT_MBLK_PTRS(mp); #define PRINT_MBLK(mp) \ trace(timestamp); \ printf("\n"); \ PRINT_MBLK_INFO(mp); \ printf("DB_FLAGS = 0x%x", DB_FLAGS((mblk_t *)mp)); \ tracemem(((mblk_t *)mp)->b_rptr, DUMP_SIZE); \ tracemem(((mblk_t *)mp)->b_wptr - DUMP_SIZE, \ DUMP_SIZE); sdt:::kssl_mblk-* { trace(timestamp); printf("\n"); PRINT_MBLK(arg0) }
This is actually an excerpt from my (currently internal) KSSL debugging suite.
An example of output from such probe can be seen in my
Coloring dtrace output
post.
For more complex projects it would be waste to stop here. Prefixes could be further
structured. However, this has some drawbacks. In particular, I was thinking about having
kssl_mblk- and kssl_err- prefixes. Now what to do for places
where an error condition occurred _and_ we would like to see the associated mblk ?
Using something like kssl_mblk_err-* comes to ones mind. However, there is a problem
with that - what about the singleton cases (only mblk, only err). Sure, using multiple
wildcards in dtrace is possible (e.g. syscall::*read*:) but this will
make it ugly and complicated given the number of mblk+err cases (it's probably safe to
assume that the number of such cases will be low). Simply, it's not worth the hassle.
Rather, I went with 2 probes.
To conclude, using structured prefixes is highly beneficial only for set of probes where
categories/sub-prefixes create non-intersecting sets (e.g. data type and debug level).
Of course, all of the above is not valid only for kernel but also for custom userland probes !
[1] High-level description of KSSL can be found in blueprint 819-5782.
tags:
debugging
dtrace
observability
opensolaris
probes
sdt
solaris
Linkage:
Technorati cosmos
Saturday Dec 01, 2007
CR 4664622 has been integrated into Nevada and will be part of build 80 (which means it will not be part of next SXDE release but I can live with that :)).
During the course of getting the process done I have stumbled upon several interesting obstacles. For example, during ingress Open Source Review I was asked by our open-source caretaker what will be the "support model" for Netcat once it is integrated. I was puzzled. Because, for Netcat, support is not really needed since it has been around for ages (ok, since 1996 according to wikipedia) and is pretty stable piece of code which is basically no longer developed. Nonetheless, this brings some interesting challenges with move to a community model where more and more projects are integrated by people outside Sun (e.g. ksh93 project).
The nc(1) man page will be delivered in build 81. In the meantime you can read
Netcat review blog entry which contains the link to updated man page.
The older version of the man page is contained in the mail communication for PSARC 2007/389.
Note: I have realized that the ARC case directory does not have to include most up-to-date man page at the time of integration.
Only when something _architectural_ changes, then the man page has to be updated (which was not the case with Netcat since we only added new section describing how to setup nc(1) with RBAC). Thanks to Jim for the explanation.
I have some ideas how to make Netcat in Solaris even better and will work to get them done over time. In particular, there are following RFEs: 6515928, 6573202. However, this does not mean that there is only single person who can work on nc(1). Since it is now part of ONNV, anyone is free to hack it.
So, I'd like to invite everyone to participate - if you have an idea how to extend Netcat, what features to add, it is sitting in ONNV waiting for new RFEs (or bugs) and sponsor requests (be sure to read Jayakara Kini's explanation of how to contribute if you're not OpenSolaris contributor yet).
Also, if you're Netcat user and use Netcat in a cool way, I want to hear that !
tags:
contribute
nc
netcat
opensolaris
opensource
participate
solaris
Linkage:
Technorati cosmos
Saturday Nov 03, 2007
As you might know, Netcat implementation is going to be part of OpenSolaris. The initial Netcat integration is based on a reimplementation from OpenBSD (here's why).
As Jeff Bonwick said, open sourced code is nothing compared to the fact that all design discussions and decisions suddenly happen in the public (loosely paraphrased). This is a great wave to ride and I have jumped on it when it was not really small so I have at least posted the webrev pointer for initial Netcat integration (CR 4664622) to the opensolaris-code mailing list (which is roughly the equivalent of freebsd-hackers, openbsd-tech or similar mailing lists) to get some code review comments.
Since then couple of things changed. Thanks to Dan Price and others it's now possible to upload webrevs to cr.opensolaris.org. I have started using the service so the new and official place for the Netcat webrev is
The webrev has moved location but what I said in the opensolaris-code post still holds true:
Any constructive notes regarding the code are welcome. (I am mainly looking for functional/logic errors, packaging flaws or parts of code which could mismatch the PSARC case)
The following things could help any potential code reviewer:
- Summary of the changes done to the original implementation in order to adapt it to Solaris environment..
- PSARC 2007/389 case covering interfaces delivered by this project. For more information about ARCs see Architecture Process and Tools community pages.
- SUNWnetcat package (x86) which contains /usr/bin/nc binary
- webrev
of the differences between my version of Netcat and the one which is
currently in OpenBSD.
Only the *.[ch] files matter, of course. (This is very easy thing to do with distributed SCM since it only requires one to reparent and regenerate webrev against new parent workspace) - Updated manual page This is slightly different from the man page in the PSARC materials because it contains new section about using nc with privileges and associated set of examples in the EXAMPLES section. The man page in the PSARC materials will not be updated because after a case is approved, the man page is updated only in case some architectural changes were needed. In the case of privileges, it is only addition describing specific usage, no architectural changes.
The conclusion for non code reviewers ? I hope it is clear the in (Open)Solaris land we value quality and transparency. Peer reviews and architectural reviews are just (albeit crucial) pieces which help to achieve that goal.
tags:
code
netcat
review
solaris
Linkage:
Technorati cosmos
Tuesday Aug 28, 2007
In my previous entry about BSD compatibility gap closure process I have promised to provide a guide on how to get new code into libc. I will use changes done via CR 6495220 to illustrate the process with examples.
Process related and technical changes which are usually needed:
- get PSARC case done
- File a CR to create a manual page according to the man page draft supplied with the PSARC case. You will probably need to go through the functions being added and assign them MT-Level according to attributes(5) man page (if this was not done prior to filing the PSARC case).
- actually add the code into libc
This includes moving/introducing files from the SCM point of view and doing necessary changes to the Makefiles.
In terms of symbols, the functions need to be actually delivered twice. Once as underscored (strong) symbol and second as WEAK alias to the strong symbol. This allows libraries use their own private implementation of the functions. (This is because the weak symbol is silently overridden by the private symbol in runtime linker) - add entries to c_synonyms.h and synonyms.h
synonyms.h is used in libc for symbol alias contruction (see above). c_synonyms.h provides access to underscored symbols for other (non-libc) libraries. This provides a way how to call the underscored symbols directly without risking namespace clashes/pollution.
This step is actually needed to be used in conjunction with the previous step. nm(1) can be used to check this worked as expected:$ nm -x /usr/lib/libc.so.1 | grep '|_\?err$' [5783] |0x00049c40|0x00000030|FUNC |GLOB |0 |13 |_err [6552] |0x00049c40|0x00000030|FUNC |WEAK |0 |13 |err
- Do the necessary packaging changes
If you're adding new header file change SUNWhea's prototype_* files (most probably just prototype_com)
If the file was previously installed into proto area during build it needs to be removed from the exception files (for i386 and sparc). - modify lib's mapfile
This is needed for the symbols to become visible and versioned. Public symbols belong to the latest SUNW section. After you have compiled the changes you can check this via command similar to the following:pvs -dsv -N SUNW_1.23 /usr/lib/libc.so.1 \ | sed -n '1,/SUNW.*:/p' | egrep '((v?errx?)|(v?warnx?));' vwarnx; ...If you're adding private (underscored) symbols do not forget to add them to the SUNWprivate section. This is usually the case because the strong symbols are accompanied by weak symbols. Weak symbols go to the global part of the most recent SUNW section and strong symbols go to global part of SUNWprivate section. - update libc's lint library
If you are adding private symbols then add them as well. See the entries _vwarnfp et al. for example.
After you're done it's time to run nightly with lint checks and fix the noise. (see below) - Add per-symbol filters
If you are moving stuff from a library to libc you will probably want to preserve the existing interfaces. To accomplish this per-symbol filters can be added to the library you're moving from. So, if symbol foo is moved from libbar to libc then change the line in the global section of libbar's mapfile to look like this:foo = FUNCTION FILTER libc.so.1;
This was done with the *fp functions in libipsecutils' mapfile. The specialty in that case was that the *fp functions were renamed to underscored variants while moving them via redefines in errfp.h. - Fix build/lint noise introduced by the changes
There could be the following noises:- build noise
Can be caused by symbol type clash (there is symbol of the same name defined in libc as FUNC and in $SRC/cmd as OBJT) which is not harmful because ld(1) will do due diligence and prefer the non-libc symbol. This can be fixed by renaming the local symbol. There could also be #define clash caused by inclusion of c_synonyms.h. Fixed via renaming as well. - lint noise
In the 3rd pass of the lint checks an inconsistency in function declarations can be found such as this:/builds/.../usr/include/err.h", line 43: warning: function argument declared inconsistently: warn(arg 1) in utils.c(62) char * and llib-lc:err.h(43) const char * (E_INCONS_ARG_DECL2)
The problem with this output is that there are cca 23 files named utils.c in ONNV. CR 6585621 is waiting someone to provide remedy for that via adding -F flag to LINTFLAGS in $SRC/lib and $SRC/cmd.
After the right file(s) are found the fix is usually renaming again. Where the renaming is not possible -erroff=E_FUNC_DECL_VAR_ARG2 can be passed to lint(1).
- build noise
- Make sure there are not duplicate symbols in libc after the changes
This is necessary because it might confuse debugging tools (mdb, dtrace). For err/warn stuff there was one such occurence:[6583] | 301316| 37|FUNC |GLOB |0 |13 |_warnx [1925] | 320000| 96|FUNC |LOCL |0 |13 |_warnx
This can be usually solved by renaming the local variant. - Test thoroughly
- test with different compilers
SunStudio does different things than gcc so it is good idea to test the code with both. - Try to compile different consolidations (e.g. Companion CD, SFW) on top of the changes. For err/warn project a bug was filed to get RPM build fixed.
- Test if the WEAK symbols actually work
- Test the programs in ONNV affected by the changes
e.g. the programs which needed to be modified because of the build/lint noise.
- test with different compilers
- Produce annotated putback list explaining the changes
This is handy for a CRT advocate and saves time. - If the change requires some sort of action from fellow gatelings, send a heads-up, e.g. like heads-up for err/warn.
- If you are actually adding code to libc (this includes moving code from other libraries to libc) send an e-mail similar to the heads-up e-mail to opensolaris-code mailing list, e.g. like this message about err/warn.
tags:
bsd
libc
opensolaris
solaris
Linkage:
Technorati cosmos
Thursday Aug 23, 2007
I do not get to meet customers very often but I clearly remember the last time where I participated in a pseudo-technical session with a new customer. The engineers were keen on learning details about all features which make Solaris outstanding but they were also bewildered by the lack of common functions such as daemon() (see CR 4471189). Yes, there is a number of private implementations in various places in ONNV, however this is not very useful. Until recently, this was also the case of err/warn function family.
With the putback of CR 6495220 there are now the following functions living in libc:
void err(int, const char *, ...); void verr(int, const char *, va_list); void errx(int, const char *, ...); void verrx(int, const char *, va_list); void warn(const char *, ...); void vwarn(const char *, va_list); void warnx(const char *, ...); void vwarnx(const char *, va_list);
These functions were present in BSD systems for a long time (they've been in FreeBSD since 1994). The configure scripts of various pieces of software contain checks for presence and functionality of the err/warn functions in libc (and setting the HAVE_ERR_H define). For Solaris, those checks have now become enabled too.
The err(3C) man page covering these functions will be delivered in the same build as the changes, that is build 72.
The change is covered by PSARC 2006/662 architectural review and the stability level for all the functions is Committed
(see Architecture Process and Tools for more details on how this works).
Unfortunately, the case is still closed. Hopefully it will be opened soon.
Update 09-28-2007: the PSARC/2006/662 case is now open, including onepager document and e-mail discussion. Thanks to John Plocher, Darren Reed and Bill Sommerfeld.
As I prompted in the err/warn Heads-up there is now time to stop creating new private implementations and to look at purging duplicates (there are many of them, however not all can be gotten rid of in favour of err/warn from libc).
I will write about how to get code into libc in general from a more detailed perspective next time.
However, this does not mean there is nothing left to do in this area. Specifically, FreeBSD's err(3) contains functions err_set_exit(), err_set_file() and errc(), warnc(), verrc() and vwarnc(). These functions could be ported over too. Also, there is __progname or getprogname(3). Moreover, careful (code) reader has realized that err.c contains private implementations of functions with fp suffix. This function (sub)family could be made Committed too. So, there is still lot of work which could be done. Anyone is free to work on any of these. (see Jayakara Kini's blog entry on how to become OpenSolaris contributor if you're not already)
tags:
bsd
err
opensolaris
solaris
warn
Linkage:
Technorati cosmos
Tuesday Mar 27, 2007
During stressful days of libmd backport to Solaris 10 update 4 I managed to break my iPod mini by sliding on my chair backwards and running to get next coffee cup. (The catch was that I still had my headphones connected to the iPod on so the iPod fell down on the floor. It now only reports the unhappy face of Macintosh.)
Only after that I have realized how music is important for my day-to-day tasks. I cannot simply continue in the same pace as before without steady flow of rhythm into my ears.
Before buying another iPod I wanted to get some interim relief. This is when I entered the not-so-happy world of Solaris audio drivers. My workstation is Acer Aspire E300 which came with integrated TV card and whatnot but also with integrated sound card. The sound card is supported in Solaris but the level of noise coming from it was unbearable. (and I am no HiFi snob, listening mostly to 32kbps radio Akropolis streams) After some googling I have realized that there is a driver for Sound Blaster Live! 5.1 PCI card which I had in my Solaris test workstation at home. The driver was backported by Jürgen Keil (frequent OpenSolaris contributor) from NetBSD among other drivers.
The relevant part of prtconf -v output looks like this:
pci1102,8027 (driver not attached)
Hardware properties:
name='assigned-addresses' type=int items=5
value=81024810.00000000.00009400.00000000.00000020
name='reg' type=int items=10
value=00024800.00000000.00000000.00000000.00000000.01024
810.00000000.00000000.00000000.00000020
name='compatible' type=string items=7
value='pci1102,2.1102.8027.7' + 'pci1102,2.1102.8027' +
'pci1102,8027' + 'pci1102,2.7' + 'pci1102,2' + 'pciclass,040100' + 'pciclass,0401'
name='model' type=string items=1
value='Audio device'
It's quite easy to get it working:
- disable the on-board sound card in BIOS
- get the sources and extract them
/usr/sfw/bin/wget http://www.tools.de/files/opensource/solaris/audio/beta/audio-1.9beta.tar.bz2 bzcat audio-1.9beta.tar.bz2 | tar -xf -
- compile (set the PATH if needed to get access to gcc/ar)
export PATH=/usr/bin:/usr/sbin:/usr/sfw/bin:/usr/sfw/sbin:/usr/ccs/bin cd audio-1.9beta && make
- install the driver (audioemu) and its dependency (audiohlp) on x86:
cp drv/emu/audioemu /platform/i86pc/kernel/drv cp drv/emu/amd64/audioemu /platform/i86pc/kernel/drv/amd64 cp drv/emu/audioemu.conf /platform/i86pc/kernel/drv cp misc/audiohlp /kernel/misc cp misc/amd64/audiohlp /kernel/misc/amd64 cp misc/audiohlp.conf /kernel/misc
- attach the driver (see instructions in drv/emu/Makefile)
add_drv -i '"pci1102,2" "pci1102,4"' audioemu
- reboot
Yes, I could have used the infrastructure provided in Makefiles to create the package and install it but I wanted to have minimalistic install and have just the things which are really needed.
After the reboot you should be able to play sound via e.g. xmms. (installed e.g. from Blastwave) Check for audioemu in modinfo(1M) output and dmesg(1M) output for error messages if something goes wrong. So far it has been working for me rather flawlessly. (no kernel panics ;))
During the search for the driver I have discovered number of complaints from the users trying OpenSolaris for the first time that their Sound Blaster Live! was not recognized.
Looking into Device drivers community not much is going on about soundcard drivers.
I wonder how hard would it be to get the audioemu NetBSD-based driver into ON.. the CR number is 6539690.
2007-08-08 update: After asking on opensolaris-discuss mailing list I have realized that there is a project underway which will deliver OSS into (Open)Solaris. Some info can be found in PSARC/2007/238, however there is no project page at opensolaris.org (yet). Hopefully, RFE 6539690 will be closed after OSS integrates into ONNV.
tags:
blaster
opensolaris
solaris
sound
Linkage:
Technorati cosmos