- User Since
- May 28 2014, 2:27 PM (189 w, 6 d)
Fri, Jan 5
Dec 7 2017
Dec 6 2017
Dec 5 2017
Nov 28 2017
Nov 22 2017
Nov 21 2017
My fault, thanks for fixing.
Nov 13 2017
Nov 8 2017
Oct 19 2017
D12405 looks good to me. Matches set of changed I had with 4.9.0. Is reduced actually, since tcpdump has incorporated some of our changes since.
Hi! I think phabrickator is useless for such kinds of things. Did you follow instructions in vendor/tcpdump/FreeBSD-upgrade?
Oct 12 2017
Oct 11 2017
Oct 10 2017
SUS specifies ECONNRESET for write(2), but BSD systems never did that. SUS also specifies EPIPE, and doesn't provide any clue when ECONNRESET is preferred over EPIPE. They got very similar descriptions.
I'm sorry for accepting the previous patch too fast, without investing time into proper review.
My suggestion is to commit in two patches. First, revert r324405. Second, move of SS_ISCONNECTED check.
Oct 9 2017
The change doesn't affect sfbufs at all.
Oct 7 2017
Yes, it will be. It is split in commits in my git repo.
Oct 6 2017
Can you please grab the PR? :)
Oct 2 2017
How is that possible that tcp_usr_send leads to tcp_connect and in_pcbbind? Shouldn't we get immediate ENOTCONN if we try to sendfile on a not connected socket?
That's an interesting finding! However, I don't like that we are blindly following write() or send() logic. Copy-pasting the ERESTART is definitely superfluous. I'm pretty sure that protocols shall never return that. More complex question is that why are we sure that EAGAIN confirms that all bytes sent were actually sent? Could be bug not in sendfile, but in write :) Let me dig more on that.
Does this fix this bug? https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=217401
Sep 26 2017
Sep 16 2017
Sep 14 2017
It is possible to separate white space changes from the rest of changes? Do you have this in git or are you using just modified subversion workdir?
Sep 13 2017
Aug 31 2017
Aug 24 2017
Aug 21 2017
Aug 18 2017
Aug 9 2017
Sorry for clicking wrong button. I'm seconding ae@: IPSEC_SUPPORT should remain, but IPSEC should be removed. The ipsec.ko and tcpmd5.ko should be loadable, not statically compiled in.
Jun 25 2017
Jun 15 2017
Jun 14 2017
Jun 13 2017
Did a style review, now working on understanding the gist of changes.
Jun 12 2017
The change consist also some changes to the protocol. Some new bits are parsed. I'm pretty sure that my particular device doesn't have any pressure. All I can do is to confirm that my device is still working.
Ack. Will test before 1st of July.
Jun 10 2017
I had this dialogue with jhb@ before. To not repeat myself, please read these two emails.
Jun 9 2017
I'm going to write python scripts for gdb.
Oh, they use struct pfsync_state. :( So we are limited to 512 byte subheaders for the time being of this 5.x protocol.
Jun 8 2017
I apologize, Marcel. Didn't notice the update. So, what do you plan to add for compatibility with older FreeBSD versions? My suggestion is that if we ever receive a bulk update request with version 5.0, we must downgrade to version 5.0. If this is implemented, I am all for this change.
Actually my comment was an objection. And you read me wrong on compatibility history. All FreeBSD versions that support pfsync, starting with FreeBSD 5.4 and ending in FreeBSD 12.0 are compatible. They are also compatible with OpenBSD versions that used protocol version 5.
I plan to commit this tomorrow.
- Fix check for the listen queue overflow. We shall not do this
- Revert unnecessary changes to syncache KPI.
- Put back SO_ACCEPTFILTER on the listening socket. It is inherited by
- Copy so_options holding the lock. This closes race with accept
- Two unrelated changes: accept filter teardown and close race.
- Fix a degenerate case when soisdisconnected() would call soisconnected().
- Improve fe263daf185e1b88691f9d1df5c92b56a496e050: check sol_accept_filter
Jun 7 2017
Jun 2 2017
Merge remote-tracking branch 'FreeBSD/master' into solisten
Jun 1 2017
- Merge commit '153f9fe91de8a2d0aa5910076e983a7c9d76100a' into solisten
- Check if socket is not listening or is listening with queue overflow.
- SOCK_LOCK is no longer synonym for locking receive socket buffer.
- Access to so_error and to receive socket buffer are not protected
- Allow listening on 1-to-many sockets, but not actually convert them
- Merge commit 'f92f6763f17a7dff5cd5e4b96b5c433c33931811' into solisten
- Use the SCTP_PCB_FLAGS_ACCEPTING flags to check for listeners.