Fix broken fragment/mssfix with NCP
PATCH 1/1 Fix broken fragment/mssfix with NCP
From: Lev Stipakov <[email protected]...>
NCP negotiation replaces worst cast crypto overhead with actual one in data channel frame. That frame params are used by mssfix.
Fragment frame still contains worst case overhead. Because of that TCP packets are fragmented, since MSS value exceeds max fragment size.
Fix by replacing worst case crypto overhead with actual one for fragment frame, as it is done for data channel frame.
Trac #1140
--- src/openvpn/forward.c | 1 + src/openvpn/init.c | 12 +++++++++++- src/openvpn/openvpn.h | 1 + src/openvpn/push.c | 9 ++++++++- src/openvpn/ssl.c | 19 ++++++++++++++++++- src/openvpn/ssl.h | 13 ++++++++----- 6 files changed, 47 insertions(+), 8 deletions(-)
src/openvpn/forward.c +1 -0
@@ -1090,6 +1090,7 @@ process_incoming_link_part1(struct context *c, struct link_socket_info *lsi, booif (is_hard_reset (opcode, c->options.key_method)) { c->c2.frame = c->c2.frame_initial; + c->c2.frame_fragment = c->c2.frame_fragment_initial; }interval_action (&c->c2.tmp_int);
src/openvpn/init.c +11 -1
@@ -2300,9 +2300,18 @@ do_deferred_options(struct context *c, const unsigned int found) {tls_poor_mans_ncp (&c->options, c->c2.tls_multi->remote_ciphername); } +struct frame *frame_fragment =NULL ; +#ifdef ENABLE_FRAGMENT +if (c->options.ce.fragment) + { + frame_fragment = &c->c2.frame_fragment; + } +#endif +/* Do not regenerate keys if server sends an extra push reply */ if (!session->key[KS_PRIMARY ].crypto_options.key_ctx_bi.initialized - && !tls_session_update_crypto_params (session, &c->options, &c->c2.frame)) + && !tls_session_update_crypto_params (session, &c->options, &c->c2.frame, + frame_fragment)) {msg (D_TLS_ERRORS ,"OPTIONS ERROR: failed to import crypto options" );return false;
@@ -3082,6 +3091,7 @@ do_init_frame(struct context *c)*/ c->c2.frame_fragment = c->c2.frame;frame_subtract_extra (&c->c2.frame_fragment, &c->c2.frame_fragment_omit); + c->c2.frame_fragment_initial = c->c2.frame_fragment;#endif #if defined(ENABLE_FRAGMENT ) && defined(ENABLE_OCC )
src/openvpn/openvpn.h +1 -0
@@ -265,6 +265,7 @@ struct context_2/* Object to handle advanced MTU negotiation and datagram fragmentation */ struct fragment_master *fragment;struct frame frame_fragment; +struct frame frame_fragment_initial;struct frame frame_fragment_omit;#endif
src/openvpn/push.c +8 -1
@@ -274,11 +274,18 @@ incoming_push_message(struct context *c, const struct buffer *buffer) {if (c->options.mode ==MODE_SERVER ) { +struct frame *frame_fragment =NULL ; +#ifdef ENABLE_FRAGMENT +if (c->options.ce.fragment) + { + frame_fragment = &c->c2.frame_fragment; + } +#endif struct tls_session *session = &c->c2.tls_multi->session[TM_ACTIVE ];/* Do not regenerate keys if client send a second push request */ if (!session->key[KS_PRIMARY ].crypto_options.key_ctx_bi.initialized && !tls_session_update_crypto_params (session, &c->options, - &c->c2.frame)) + &c->c2.frame, frame_fragment)) {msg (D_TLS_ERRORS ,"TLS Error: initializing data channel failed" ); goto error;
src/openvpn/ssl.c +18 -1
@@ -1986,7 +1986,8 @@ cleanup:bool tls_session_update_crypto_params (struct tls_session *session, -struct options *options,struct frame *frame) +struct options *options,struct frame *frame, +struct frame *frame_fragment) {if (!session->opt->server && 0 != strcmp(options->ciphername, session->opt->config_ciphername)
@@ -2030,6 +2031,22 @@ tls_session_update_crypto_params(struct tls_session *session,frame_init_mssfix (frame, options);frame_print (frame,D_MTU_INFO ,"Data Channel MTU parms" ); +/* +* mssfix uses data channel framing, which at this point contains +* actual overhead. Fragmentation logic uses frame_fragment, which +* still contains worst case overhead. Replace it with actual overhead +* to prevent unneeded fragmentation. +*/ + +if (frame_fragment) + { +frame_remove_from_extra_frame (frame_fragment,crypto_max_overhead ()); +crypto_adjust_frame_parameters (frame_fragment, &session->opt->key_type, + options->replay, packet_id_long_form); +frame_set_mtu_dynamic (frame_fragment, options->ce.fragment,SET_MTU_UPPER_BOUND ); +frame_print (frame_fragment,D_MTU_INFO ,"Fragmentation MTU parms" ); + } +return tls_session_generate_data_channel_keys (session); }
src/openvpn/ssl.h +8 -5
@@ -474,15 +474,18 @@ void tls_update_remote_addr(struct tls_multi *multi,* Update TLS session crypto parameters (cipher and auth) and derive data * channel keys based on the supplied options. * -* @param session The TLS session to update. -* @param options The options to use when updating session. -* @param frame The frame options for this session (frame overhead is -* adjusted based on the selected cipher/auth). +* @param session The TLS session to update. +* @param options The options to use when updating session. +* @param frame The frame options for this session (frame overhead is +* adjusted based on the selected cipher/auth). +* @param frame_fragment The fragment frame options. * * @return true if updating succeeded, false otherwise. */ bool tls_session_update_crypto_params (struct tls_session *session, -struct options *options,struct frame *frame); +struct options *options, +struct frame *frame, +struct frame *frame_fragment);/** * "Poor man's NCP": Use peer cipher if it is an allowed (NCP) cipher.
Signature · 1 line
2.7.4
Hi,
Quoted 10 lines · Lev Stipakov
From: Lev Stipakov <[email protected]...>
NCP negotiation replaces worst cast crypto overhead with actual one in data channel frame. That frame params are used by mssfix.
Fragment frame still contains worst case overhead. Because of that TCP packets are fragmented, since MSS value exceeds max fragment size.
Quoted 5 lines
Fix by replacing worst case crypto overhead with actual one for fragment frame, as it is done for data channel frame.
Trac #1140
Thanks for digging into this!
Quoted 15 lines
Signed-off-by Lev Stipakov <[email protected]...>--- src/openvpn/forward.c | 1 + src/openvpn/init.c | 12 +++++++++++- src/openvpn/openvpn.h | 1 + src/openvpn/push.c | 9 ++++++++- src/openvpn/ssl.c | 19 ++++++++++++++++++- src/openvpn/ssl.h | 13 ++++++++----- 6 files changed, 47 insertions(+), 8 deletions(-)src/openvpn/forward.c +0
-0
src/openvpn/forward.c process_incoming_link_part1(struct context *c, struct link_socket_info *lsi, boo · line 1093
Quoted 4 lines
src/openvpn/forward.c +1
-0@@ -1090,6 +1090,7 @@ process_incoming_link_part1(struct context *c, struct link_socket_info *lsi, booif (is_hard_reset (opcode, c->options.key_method)) { c->c2.frame = c->c2.frame_initial; + c->c2.frame_fragment = c->c2.frame_fragment_initial;
This breaks --disable-fragment builds:
src/openvpn/forward.c:1096:27: error: no member named 'frame_fragment' in 'struct context_2'
c->c2.frame_fragment = c->c2.frame_fragment_initial;
~~~~~ ^
src/openvpn/forward.c:1096:50: error: no member named
'frame_fragment_initial' in 'struct context_2'
c->c2.frame_fragment = c->c2.frame_fragment_initial;
~~~~~ ^2 errors generated.
Makefile:676: recipe for target 'forward.o' failed
Quoted 127 lines
src/openvpn/forward.c +0
-0}interval_action (&c->c2.tmp_int);src/openvpn/init.c +11
-1@@ -2300,9 +2300,18 @@ do_deferred_options(struct context *c, const unsigned int found) {tls_poor_mans_ncp (&c->options, c->c2.tls_multi->remote_ciphername); } +struct frame *frame_fragment =NULL ; +#ifdef ENABLE_FRAGMENT +if (c->options.ce.fragment) + { + frame_fragment = &c->c2.frame_fragment; + } +#endif +/* Do not regenerate keys if server sends an extra push reply */ if (!session->key[KS_PRIMARY ].crypto_options.key_ctx_bi.initialized - && !tls_session_update_crypto_params (session, &c->options, &c->c2.frame)) + && !tls_session_update_crypto_params (session, &c->options, &c->c2.frame, + frame_fragment)) {msg (D_TLS_ERRORS ,"OPTIONS ERROR: failed to import crypto options" );return false;@@ -3082,6 +3091,7 @@ do_init_frame(struct context *c)*/ c->c2.frame_fragment = c->c2.frame;frame_subtract_extra (&c->c2.frame_fragment, &c->c2.frame_fragment_omit); + c->c2.frame_fragment_initial = c->c2.frame_fragment;#endif #if defined(ENABLE_FRAGMENT ) && defined(ENABLE_OCC )src/openvpn/openvpn.h +1
-0@@ -265,6 +265,7 @@ struct context_2/* Object to handle advanced MTU negotiation and datagram fragmentation */ struct fragment_master *fragment;struct frame frame_fragment; +struct frame frame_fragment_initial;struct frame frame_fragment_omit;#endif src/openvpn/push.c +8
-1@@ -274,11 +274,18 @@ incoming_push_message(struct context *c, const struct buffer *buffer) {if (c->options.mode ==MODE_SERVER ) { +struct frame *frame_fragment =NULL ; +#ifdef ENABLE_FRAGMENT +if (c->options.ce.fragment) + { + frame_fragment = &c->c2.frame_fragment; + } +#endif struct tls_session *session = &c->c2.tls_multi->session[TM_ACTIVE ];/* Do not regenerate keys if client send a second push request */ if (!session->key[KS_PRIMARY ].crypto_options.key_ctx_bi.initialized && !tls_session_update_crypto_params (session, &c->options, - &c->c2.frame)) + &c->c2.frame, frame_fragment)) {msg (D_TLS_ERRORS ,"TLS Error: initializing data channel failed" ); goto error;src/openvpn/ssl.c +18
-1@@ -1986,7 +1986,8 @@ cleanup:bool tls_session_update_crypto_params (struct tls_session *session, -struct options *options,struct frame *frame) +struct options *options,struct frame *frame, +struct frame *frame_fragment) {if (!session->opt->server && 0 != strcmp(options->ciphername, session->opt->config_ciphername)@@ -2030,6 +2031,22 @@ tls_session_update_crypto_params(struct tls_session *session,frame_init_mssfix (frame, options);frame_print (frame,D_MTU_INFO ,"Data Channel MTU parms" ); +/* +* mssfix uses data channel framing, which at this point contains +* actual overhead. Fragmentation logic uses frame_fragment, which +* still contains worst case overhead. Replace it with actual overhead +* to prevent unneeded fragmentation. +*/ + +if (frame_fragment) + { +frame_remove_from_extra_frame (frame_fragment,crypto_max_overhead ()); +crypto_adjust_frame_parameters (frame_fragment, &session->opt->key_type, + options->replay, packet_id_long_form); +frame_set_mtu_dynamic (frame_fragment, options->ce.fragment,SET_MTU_UPPER_BOUND ); +frame_print (frame_fragment,D_MTU_INFO ,"Fragmentation MTU parms" ); + } +return tls_session_generate_data_channel_keys (session); }src/openvpn/ssl.h +5
-4@@ -474,15 +474,18 @@ void tls_update_remote_addr(struct tls_multi *multi,* Update TLS session crypto parameters (cipher and auth) and derive data * channel keys based on the supplied options. * -* @param session The TLS session to update. -* @param options The options to use when updating session. -* @param frame The frame options for this session (frame overhead is -* adjusted based on the selected cipher/auth). +* @param session The TLS session to update. +* @param options The options to use when updating session. +* @param frame The frame options for this session (frame overhead is +* adjusted based on the selected cipher/auth). +* @param frame_fragment The fragment frame options. *
src/openvpn/ssl.h void tls_update_remote_addr(struct tls_multi *multi, · line 491
Quoted 10 lines
src/openvpn/ssl.h +3
-1* @return true if updating succeeded, false otherwise. */ bool tls_session_update_crypto_params (struct tls_session *session, -struct options *options,struct frame *frame); +struct options *options, +struct frame *frame, +struct frame *frame_fragment);/** * "Poor man's NCP": Use peer cipher if it is an allowed (NCP) cipher.
Otherwise this patch looks good to me. I recall finding weird behaviour when I tested this patch earlier, but can't reproduce it anymore based on my notes from then. I tried (large) pings and TCP transfers, using --fragment 500 and 1000 and inspected the packet sizes using tcpdump. Both for client-to-server and server-to-client traffic. Although --fragment for non-TCP traffic still seems to behave suboptimal, this patch does improve behaviour for TCP streams. Since that is probably the majority or traffic where people need performance, this is very useful.
Once the --disable-fragment builds have been fixed, I have no objections to merging that patch. But I would prefer to have one of the networking guys to take a good look at this too before merging.
-Steffan
PATCH v2 1/1 Fix broken fragment/mssfix with NCP
From: Lev Stipakov <[email protected]...>
NCP negotiation replaces worst cast crypto overhead with actual one in data channel frame. That frame params are used by mssfix.
Fragment frame still contains worst case overhead. Because of that TCP packets are fragmented, since MSS value exceeds max fragment size.
Fix by replacing worst case crypto overhead with actual one for fragment frame, as it is done for data channel frame.
Trac #1140
---
v2: fix --disable-fragment build
src/openvpn/forward.c | 3 +++ src/openvpn/init.c | 12 +++++++++++- src/openvpn/openvpn.h | 1 + src/openvpn/push.c | 9 ++++++++- src/openvpn/ssl.c | 19 ++++++++++++++++++- src/openvpn/ssl.h | 13 ++++++++----- 6 files changed, 49 insertions(+), 8 deletions(-)
src/openvpn/forward.c +3 -0
@@ -1093,6 +1093,9 @@ process_incoming_link_part1(struct context *c, struct link_socket_info *lsi, booif (is_hard_reset (opcode, c->options.key_method)) { c->c2.frame = c->c2.frame_initial; +#ifdef ENABLE_FRAGMENT + c->c2.frame_fragment = c->c2.frame_fragment_initial; +#endif }interval_action (&c->c2.tmp_int);
src/openvpn/init.c +11 -1
@@ -2298,9 +2298,18 @@ do_deferred_options(struct context *c, const unsigned int found) {tls_poor_mans_ncp (&c->options, c->c2.tls_multi->remote_ciphername); } +struct frame *frame_fragment =NULL ; +#ifdef ENABLE_FRAGMENT +if (c->options.ce.fragment) + { + frame_fragment = &c->c2.frame_fragment; + } +#endif +/* Do not regenerate keys if server sends an extra push reply */ if (!session->key[KS_PRIMARY ].crypto_options.key_ctx_bi.initialized - && !tls_session_update_crypto_params (session, &c->options, &c->c2.frame)) + && !tls_session_update_crypto_params (session, &c->options, &c->c2.frame, + frame_fragment)) {msg (D_TLS_ERRORS ,"OPTIONS ERROR: failed to import crypto options" );return false;
@@ -3082,6 +3091,7 @@ do_init_frame(struct context *c)*/ c->c2.frame_fragment = c->c2.frame;frame_subtract_extra (&c->c2.frame_fragment, &c->c2.frame_fragment_omit); + c->c2.frame_fragment_initial = c->c2.frame_fragment;#endif #if defined(ENABLE_FRAGMENT ) && defined(ENABLE_OCC )
src/openvpn/openvpn.h +1 -0
@@ -265,6 +265,7 @@ struct context_2/* Object to handle advanced MTU negotiation and datagram fragmentation */ struct fragment_master *fragment;struct frame frame_fragment; +struct frame frame_fragment_initial;struct frame frame_fragment_omit;#endif
src/openvpn/push.c +8 -1
@@ -287,11 +287,18 @@ incoming_push_message(struct context *c, const struct buffer *buffer) {if (c->options.mode ==MODE_SERVER ) { +struct frame *frame_fragment =NULL ; +#ifdef ENABLE_FRAGMENT +if (c->options.ce.fragment) + { + frame_fragment = &c->c2.frame_fragment; + } +#endif struct tls_session *session = &c->c2.tls_multi->session[TM_ACTIVE ];/* Do not regenerate keys if client send a second push request */ if (!session->key[KS_PRIMARY ].crypto_options.key_ctx_bi.initialized && !tls_session_update_crypto_params (session, &c->options, - &c->c2.frame)) + &c->c2.frame, frame_fragment)) {msg (D_TLS_ERRORS ,"TLS Error: initializing data channel failed" ); goto error;
src/openvpn/ssl.c +18 -1
@@ -1990,7 +1990,8 @@ cleanup:bool tls_session_update_crypto_params (struct tls_session *session, -struct options *options,struct frame *frame) +struct options *options,struct frame *frame, +struct frame *frame_fragment) {if (!session->opt->server && 0 != strcmp(options->ciphername, session->opt->config_ciphername)
@@ -2034,6 +2035,22 @@ tls_session_update_crypto_params(struct tls_session *session,frame_init_mssfix (frame, options);frame_print (frame,D_MTU_INFO ,"Data Channel MTU parms" ); +/* +* mssfix uses data channel framing, which at this point contains +* actual overhead. Fragmentation logic uses frame_fragment, which +* still contains worst case overhead. Replace it with actual overhead +* to prevent unneeded fragmentation. +*/ + +if (frame_fragment) + { +frame_remove_from_extra_frame (frame_fragment,crypto_max_overhead ()); +crypto_adjust_frame_parameters (frame_fragment, &session->opt->key_type, + options->replay, packet_id_long_form); +frame_set_mtu_dynamic (frame_fragment, options->ce.fragment,SET_MTU_UPPER_BOUND ); +frame_print (frame_fragment,D_MTU_INFO ,"Fragmentation MTU parms" ); + } +return tls_session_generate_data_channel_keys (session); }
src/openvpn/ssl.h +8 -5
@@ -476,15 +476,18 @@ void tls_update_remote_addr(struct tls_multi *multi,* Update TLS session crypto parameters (cipher and auth) and derive data * channel keys based on the supplied options. * -* @param session The TLS session to update. -* @param options The options to use when updating session. -* @param frame The frame options for this session (frame overhead is -* adjusted based on the selected cipher/auth). +* @param session The TLS session to update. +* @param options The options to use when updating session. +* @param frame The frame options for this session (frame overhead is +* adjusted based on the selected cipher/auth). +* @param frame_fragment The fragment frame options. * * @return true if updating succeeded, false otherwise. */ bool tls_session_update_crypto_params (struct tls_session *session, -struct options *options,struct frame *frame); +struct options *options, +struct frame *frame, +struct frame *frame_fragment);/** * "Poor man's NCP": Use peer cipher if it is an allowed (NCP) cipher.
Signature · 1 line
2.7.4
Am 21.01.19 um 21:04 schrieb Lev Stipakov:
Quoted 6 lines
From: Lev Stipakov <[email protected]...>
NCP negotiation replaces worst cast crypto overhead with actual one in data channel frame. That frame params are used by mssfix.
Quoted 7 lines
Fragment frame still contains worst case overhead. Because of that TCP packets are fragmented, since MSS value exceeds max fragment size.
Fix by replacing worst case crypto overhead with actual one for fragment frame, as it is done for data channel frame.
The change looks good and is tested.
Steffan Karger
Acked-byQuoted 11 lines · Arne Schwabe
Am 21.01.19 um 21:04 schrieb Lev Stipakov:
Quoted 10 lines
From: Lev Stipakov <[email protected]...>
NCP negotiation replaces worst cast crypto overhead with actual one in data channel frame. That frame params are used by mssfix.
Fragment frame still contains worst case overhead. Because of that TCP packets are fragmented, since MSS value exceeds max fragment size.
Quoted 7 lines
Quoted 3 lines
Fix by replacing worst case crypto overhead with actual one for fragment frame, as it is done for data channel frame.
The change looks good and is tested.
Acked-by Arne Schwabe <[email protected]...>
Since Arne agrees, I'm completely convinced now too:
Your patch has been applied to the master branch.
I have not tried to understand the actual change, but the general explanation makes sense and with two ACKs already... I did run t_client tests, though :-) (and there is a fragment test).
I did try to apply it to 2.4, because it's a bugfix (so "can I have it for release/2.4, please!") and while it *applied* fine, it did not compile :-( - the function signature for crypto_adjust_frame_parameters() is different in 2.4 - "use_iv" still exists. Can you have a look, please, and send a 2.4 version?
commit d22ba6b2c551fa83d23b5cf668e08a08fde446bc (master)
Author: Lev Stipakov Date: Mon Jan 21 22:04:54 2019 +0200
Fix broken fragment/mssfix with NCP
Signed-off-by: Lev Stipakov <[email protected]...>
Acked-by: Arne Schwabe <[email protected]...>
Acked-by: Steffan Karger <[email protected]...>
Message-Id: <[email protected]...>
URL: https://www.mail-archive.com/[email protected]/msg18135.html
Signed-off-by: Gert Doering <[email protected]...>--
kind regards,
Gert Doering
PATCH v2 1/1 Fix broken fragment/mssfix with NCP
From: Lev Stipakov <[email protected]...>
This is the 2.4 backport of master patch (commit d22ba6b).
NCP negotiation replaces worst cast crypto overhead with actual one in data channel frame. That frame params are used by mssfix.
Fragment frame still contains worst case overhead. Because of that TCP packets are fragmented, since MSS value exceeds max fragment size.
Fix by replacing worst case crypto overhead with actual one for fragment frame, as it is done for data channel frame.
Trac #1140
---
v2: fix --disable-fragment build
src/openvpn/forward.c | 3 +++ src/openvpn/init.c | 12 +++++++++++- src/openvpn/openvpn.h | 1 + src/openvpn/push.c | 9 ++++++++- src/openvpn/ssl.c | 19 ++++++++++++++++++- src/openvpn/ssl.h | 13 ++++++++----- 6 files changed, 49 insertions(+), 8 deletions(-)
src/openvpn/forward.c +3 -0
@@ -873,6 +873,9 @@ process_incoming_link_part1(struct context *c, struct link_socket_info *lsi, booif (is_hard_reset (opcode, c->options.key_method)) { c->c2.frame = c->c2.frame_initial; +#ifdef ENABLE_FRAGMENT + c->c2.frame_fragment = c->c2.frame_fragment_initial; +#endif }interval_action (&c->c2.tmp_int);
src/openvpn/init.c +11 -1
@@ -2294,9 +2294,18 @@ do_deferred_options(struct context *c, const unsigned int found) {tls_poor_mans_ncp (&c->options, c->c2.tls_multi->remote_ciphername); } +struct frame *frame_fragment =NULL ; +#ifdef ENABLE_FRAGMENT +if (c->options.ce.fragment) + { + frame_fragment = &c->c2.frame_fragment; + } +#endif +/* Do not regenerate keys if server sends an extra push reply */ if (!session->key[KS_PRIMARY ].crypto_options.key_ctx_bi.initialized - && !tls_session_update_crypto_params (session, &c->options, &c->c2.frame)) + && !tls_session_update_crypto_params (session, &c->options, &c->c2.frame, + frame_fragment)) {msg (D_TLS_ERRORS ,"OPTIONS ERROR: failed to import crypto options" );return false;
@@ -3035,6 +3044,7 @@ do_init_frame(struct context *c)*/ c->c2.frame_fragment = c->c2.frame;frame_subtract_extra (&c->c2.frame_fragment, &c->c2.frame_fragment_omit); + c->c2.frame_fragment_initial = c->c2.frame_fragment;#endif #if defined(ENABLE_FRAGMENT ) && defined(ENABLE_OCC )
src/openvpn/openvpn.h +1 -0
@@ -269,6 +269,7 @@ struct context_2/* Object to handle advanced MTU negotiation and datagram fragmentation */ struct fragment_master *fragment;struct frame frame_fragment; +struct frame frame_fragment_initial;struct frame frame_fragment_omit;#endif
src/openvpn/push.c +8 -1
@@ -287,11 +287,18 @@ incoming_push_message(struct context *c, const struct buffer *buffer) {if (c->options.mode ==MODE_SERVER ) { +struct frame *frame_fragment =NULL ; +#ifdef ENABLE_FRAGMENT +if (c->options.ce.fragment) + { + frame_fragment = &c->c2.frame_fragment; + } +#endif struct tls_session *session = &c->c2.tls_multi->session[TM_ACTIVE ];/* Do not regenerate keys if client send a second push request */ if (!session->key[KS_PRIMARY ].crypto_options.key_ctx_bi.initialized && !tls_session_update_crypto_params (session, &c->options, - &c->c2.frame)) + &c->c2.frame, frame_fragment)) {msg (D_TLS_ERRORS ,"TLS Error: initializing data channel failed" ); goto error;
src/openvpn/ssl.c +18 -1
@@ -1962,7 +1962,8 @@ cleanup:bool tls_session_update_crypto_params (struct tls_session *session, -struct options *options,struct frame *frame) +struct options *options,struct frame *frame, +struct frame *frame_fragment) {if (!session->opt->server && 0 != strcmp(options->ciphername, session->opt->config_ciphername)
@@ -2006,6 +2007,22 @@ tls_session_update_crypto_params(struct tls_session *session,frame_init_mssfix (frame, options);frame_print (frame,D_MTU_INFO ,"Data Channel MTU parms" ); +/* +* mssfix uses data channel framing, which at this point contains +* actual overhead. Fragmentation logic uses frame_fragment, which +* still contains worst case overhead. Replace it with actual overhead +* to prevent unneeded fragmentation. +*/ + +if (frame_fragment) + { +frame_remove_from_extra_frame (frame_fragment,crypto_max_overhead ()); +crypto_adjust_frame_parameters (frame_fragment, &session->opt->key_type, + options->use_iv, options->replay, packet_id_long_form); +frame_set_mtu_dynamic (frame_fragment, options->ce.fragment,SET_MTU_UPPER_BOUND ); +frame_print (frame_fragment,D_MTU_INFO ,"Fragmentation MTU parms" ); + } +return tls_session_generate_data_channel_keys (session); }
src/openvpn/ssl.h +8 -5
@@ -475,15 +475,18 @@ void tls_update_remote_addr(struct tls_multi *multi,* Update TLS session crypto parameters (cipher and auth) and derive data * channel keys based on the supplied options. * -* @param session The TLS session to update. -* @param options The options to use when updating session. -* @param frame The frame options for this session (frame overhead is -* adjusted based on the selected cipher/auth). +* @param session The TLS session to update. +* @param options The options to use when updating session. +* @param frame The frame options for this session (frame overhead is +* adjusted based on the selected cipher/auth). +* @param frame_fragment The fragment frame options. * * @return true if updating succeeded, false otherwise. */ bool tls_session_update_crypto_params (struct tls_session *session, -struct options *options,struct frame *frame); +struct options *options, +struct frame *frame, +struct frame *frame_fragment);/** * "Poor man's NCP": Use peer cipher if it is an allowed (NCP) cipher.
Signature · 1 line
2.7.4
Hi,
Quoted 8 lines · Lev Stipakov
From: Lev Stipakov <[email protected]...>
This is the 2.4 backport of master patch (commit d22ba6b).
NCP negotiation replaces worst cast crypto overhead with actual one in data channel frame. That frame params are used by mssfix.
Quoted 7 lines
Fragment frame still contains worst case overhead. Because of that TCP packets are fragmented, since MSS value exceeds max fragment size.
Fix by replacing worst case crypto overhead with actual one for fragment frame, as it is done for data channel frame.
Thanks for the backport.
Unfortunately, I can not reproduce the problem *without* the patch, so I wonder if the issue got fixed by something else as side-effect?
If I do what you describe in the ticket:
run
openvpn --fragment 1300 --mssfix
and grep out the MSS lines, I get (for IPv4, IPv6 numbers are -20 bytes):
with NCP: MSS: 1460 -> 1203
--ncp-disable: MSS: 1284 -> 1211
... and the very same numbers from master...
OpenVPN versions tested:
OpenVPN 2.4.7 [git:release/2.4/416532f8e4125adb+]
OpenVPN 2.5_git [git:master/b5fe104ddbbdf59b+]
So, I'm wondering how to proceed here...
gert
Signature · 6 lines
"If was one thing all people took for granted, was conviction that if you
feed honest figures into a computer, honest figures come out. Never doubted
it myself till I met a computer with a sense of humor."
Robert A. Heinlein, The Moon is a Harsh Mistress
Gert Doering - Munich, Germany [email protected]...Hi,
Quoted 9 lines
with NCP: MSS: 1460 -> 1203
--ncp-disable: MSS: 1284 -> 1211... and the very same numbers from master...
So, I'm wondering how to proceed here...
As it turns out, the fix is correct but explanation is not quite clear.
The problem is not in incorrect MSS value (it is correct), but inside openvpn's fragmentation logic, which decides "to fragment or not to fragment".
Without this patch, fragmentation logic incorrectly uses max crypto overhead when calculating packet size. It exceeds fragment size and openvpn peforms fragmentation:
Quoted 1 line
sudo tcpdump port 1194
13:59:06.956394 IP server.fi.openvpn > nat2.panoulu.net.openvpn: UDP, length 652
13:59:06.956489 IP server.fi.openvpn > nat2.panoulu.net.openvpn: UDP, length 648
With this patch, an actual crypto overhead is used and no unnecessary fragmentation is performed:
Quoted 1 line
sudo tcpdump port 1194
13:58:08.685915 IP server.fi.openvpn > nat2.panoulu.net.openvpn: UDP, length 1272
13:58:08.686007 IP server.fi.openvpn > nat2.panoulu.net.openvpn: UDP, length 1272
I will send v3 with a better explanation.
Signature · 1 line
-Lev
PATCH v3 1/1 Fix broken fragmentation logic when using NCP
From: Lev Stipakov <[email protected]...>
This is the 2.4 backport of master patch (commit d22ba6b).
NCP negotiation replaces worst case crypto overhead with actual one in data channel frame. That frame params are used by mssfix. Fragment frame still contains worst case overhead.
Without this patch, fragmentation logic incorrectly uses max crypto overhead when calculating packet size. It exceeds fragment size and openvpn peforms fragmentation:
Quoted 1 line
sudo tcpdump port 1194
13:59:06.956394 IP server.fi.openvpn > nat2.panoulu.net.openvpn: UDP, length 652 13:59:06.956489 IP server.fi.openvpn > nat2.panoulu.net.openvpn: UDP, length 648
This patch fixes fragmentation calculation by setting actual crypto overhead, and no unnecessary fragmentation is performed:
Quoted 1 line
sudo tcpdump port 1194
13:58:08.685915 IP server.fi.openvpn > nat2.panoulu.net.openvpn: UDP, length 1272 13:58:08.686007 IP server.fi.openvpn > nat2.panoulu.net.openvpn: UDP, length 1272
Trac #1140
---
v3:
Refine a bit misleading commit message. Before one might think that
problem is in incorrect MSS value being set - it is not. The problem is in fragment calculation.
v2:
Fix --disable-fragment build
src/openvpn/forward.c | 3 +++ src/openvpn/init.c | 12 +++++++++++- src/openvpn/openvpn.h | 1 + src/openvpn/push.c | 9 ++++++++- src/openvpn/ssl.c | 19 ++++++++++++++++++- src/openvpn/ssl.h | 13 ++++++++----- 6 files changed, 49 insertions(+), 8 deletions(-)
src/openvpn/forward.c +3 -0
@@ -873,6 +873,9 @@ process_incoming_link_part1(struct context *c, struct link_socket_info *lsi, booif (is_hard_reset (opcode, c->options.key_method)) { c->c2.frame = c->c2.frame_initial; +#ifdef ENABLE_FRAGMENT + c->c2.frame_fragment = c->c2.frame_fragment_initial; +#endif }interval_action (&c->c2.tmp_int);
src/openvpn/init.c +11 -1
@@ -2294,9 +2294,18 @@ do_deferred_options(struct context *c, const unsigned int found) {tls_poor_mans_ncp (&c->options, c->c2.tls_multi->remote_ciphername); } +struct frame *frame_fragment =NULL ; +#ifdef ENABLE_FRAGMENT +if (c->options.ce.fragment) + { + frame_fragment = &c->c2.frame_fragment; + } +#endif +/* Do not regenerate keys if server sends an extra push reply */ if (!session->key[KS_PRIMARY ].crypto_options.key_ctx_bi.initialized - && !tls_session_update_crypto_params (session, &c->options, &c->c2.frame)) + && !tls_session_update_crypto_params (session, &c->options, &c->c2.frame, + frame_fragment)) {msg (D_TLS_ERRORS ,"OPTIONS ERROR: failed to import crypto options" );return false;
@@ -3035,6 +3044,7 @@ do_init_frame(struct context *c)*/ c->c2.frame_fragment = c->c2.frame;frame_subtract_extra (&c->c2.frame_fragment, &c->c2.frame_fragment_omit); + c->c2.frame_fragment_initial = c->c2.frame_fragment;#endif #if defined(ENABLE_FRAGMENT ) && defined(ENABLE_OCC )
src/openvpn/openvpn.h +1 -0
@@ -269,6 +269,7 @@ struct context_2/* Object to handle advanced MTU negotiation and datagram fragmentation */ struct fragment_master *fragment;struct frame frame_fragment; +struct frame frame_fragment_initial;struct frame frame_fragment_omit;#endif
src/openvpn/push.c +8 -1
@@ -287,11 +287,18 @@ incoming_push_message(struct context *c, const struct buffer *buffer) {if (c->options.mode ==MODE_SERVER ) { +struct frame *frame_fragment =NULL ; +#ifdef ENABLE_FRAGMENT +if (c->options.ce.fragment) + { + frame_fragment = &c->c2.frame_fragment; + } +#endif struct tls_session *session = &c->c2.tls_multi->session[TM_ACTIVE ];/* Do not regenerate keys if client send a second push request */ if (!session->key[KS_PRIMARY ].crypto_options.key_ctx_bi.initialized && !tls_session_update_crypto_params (session, &c->options, - &c->c2.frame)) + &c->c2.frame, frame_fragment)) {msg (D_TLS_ERRORS ,"TLS Error: initializing data channel failed" ); goto error;
src/openvpn/ssl.c +18 -1
@@ -1962,7 +1962,8 @@ cleanup:bool tls_session_update_crypto_params (struct tls_session *session, -struct options *options,struct frame *frame) +struct options *options,struct frame *frame, +struct frame *frame_fragment) {if (!session->opt->server && 0 != strcmp(options->ciphername, session->opt->config_ciphername)
@@ -2006,6 +2007,22 @@ tls_session_update_crypto_params(struct tls_session *session,frame_init_mssfix (frame, options);frame_print (frame,D_MTU_INFO ,"Data Channel MTU parms" ); +/* +* mssfix uses data channel framing, which at this point contains +* actual overhead. Fragmentation logic uses frame_fragment, which +* still contains worst case overhead. Replace it with actual overhead +* to prevent unneeded fragmentation. +*/ + +if (frame_fragment) + { +frame_remove_from_extra_frame (frame_fragment,crypto_max_overhead ()); +crypto_adjust_frame_parameters (frame_fragment, &session->opt->key_type, + options->use_iv, options->replay, packet_id_long_form); +frame_set_mtu_dynamic (frame_fragment, options->ce.fragment,SET_MTU_UPPER_BOUND ); +frame_print (frame_fragment,D_MTU_INFO ,"Fragmentation MTU parms" ); + } +return tls_session_generate_data_channel_keys (session); }
src/openvpn/ssl.h +8 -5
@@ -475,15 +475,18 @@ void tls_update_remote_addr(struct tls_multi *multi,* Update TLS session crypto parameters (cipher and auth) and derive data * channel keys based on the supplied options. * -* @param session The TLS session to update. -* @param options The options to use when updating session. -* @param frame The frame options for this session (frame overhead is -* adjusted based on the selected cipher/auth). +* @param session The TLS session to update. +* @param options The options to use when updating session. +* @param frame The frame options for this session (frame overhead is +* adjusted based on the selected cipher/auth). +* @param frame_fragment The fragment frame options. * * @return true if updating succeeded, false otherwise. */ bool tls_session_update_crypto_params (struct tls_session *session, -struct options *options,struct frame *frame); +struct options *options, +struct frame *frame, +struct frame *frame_fragment);/** * "Poor man's NCP": Use peer cipher if it is an allowed (NCP) cipher.
Signature · 1 line
2.7.4
[Openvpn-devel] [PATCH applied] Re: Fix broken fragmentation logic when using NCP
Thanks for clarifying what the actual problem is, so we have it nicely in the logs. I have diff'ed this patch vs. d22ba6b2c551f, and the only difference is related to "use_iv" which got kicked out from master.
So I'm mostly relying on the ACks from Arne and Steffan and not doing any review myself :-) - I have done a full client test, of course.
Your patch has been applied to the release/2.4 branch.
commit 3bd91cd0e68762b861c57cf37f144d8a11704e9d
Author: Lev Stipakov Date: Wed Oct 30 14:44:59 2019 +0200
Fix broken fragmentation logic when using NCP
Signed-off-by: Lev Stipakov <[email protected]...>
Acked-by: Gert Doering <[email protected]...>
Message-Id: <[email protected]...>
URL: https://www.mail-archive.com/[email protected]/msg18975.html
Signed-off-by: Gert Doering <[email protected]...>--
kind regards,
Gert Doering