scanner: disable most rules when we expect literal symbol
Florian Westphal
maintainerPATCH 1/1 scanner: disable most rules when we expect literal symbol
nft fails to parse certain corner-cases, for example:
nft add rule filter input meta rtclassid daddr
... as it finds DADDR token. However, 'daddr' might be a valid routing realm listed in iproute2/rt_realms, so this should be allowed.
Pablo suggested to change the start conditions in the scanner accordingly.
After this patch, the following rule works:
ct label & (foobar | saddr) == saddr ip saddr 1.2.3.4 rtclassid { 42, cosmos, rule}
[ assuming saddr and rule are also valid label/rt_realms ].
Unfortunately this patch is a lot more ugly than strictly needed -- problem is that we have to end the literal mode from the parser since we can only leave literal mode once RHS is complete.
Unfortunately, BEGIN cannot be used outside of actions in a reentrant scanner, so we have to resort to %option stack + push/pop.
Transitions are:
INITIAL -> LITERAL (got 'label' or 'rtclassid' keywords) LITERAL -> INITIAL (complete RHS)
INITIAL -> INITIAL (complete RHS w.o. previous start condition change)
---
... on top of Pablos 'parser: restrict relational rhs expression recursion'.
Let me know if you have any other suggestions or see problems with this approach.
Thanks!
include/parser.h | 3 +++ src/parser_bison.y | 2 +- src/scanner.l | 65 ++++++++++++++++++++++++++++++++++++++++++------------ 3 files changed, 55 insertions(+), 15 deletions(-)
include/parser.h +3 -0
@@ -27,6 +27,8 @@ struct parser_state {struct list_head cmds;struct eval_ctx ectx; + +unsigned int curr_yy_state; };extern void parser_init (struct parser_state *state,struct list_head *msgs);
@@ -42,5 +44,6 @@ extern int scanner_include_file(void *scanner, const char *filename,extern void scanner_push_buffer (void *scanner,const struct input_descriptor *indesc,const char *buffer); +extern void scanner_rhs_done (void *scanner);#endif /* NFTABLES_PARSER_H */
src/parser_bison.y +1 -1
@@ -1996,7 +1996,7 @@ concat_rhs_expr : basic_rhs_expr { $$ = $1; } } ; -basic_rhs_expr : primary_rhs_expr { $$ = $1; } +basic_rhs_expr : primary_rhs_expr { $$ = $1; scanner_rhs_done(scanner); } ; primary_rhs_expr : symbol_expr { $$ = $1; }
src/scanner.l +51 -14
130 lines not on this page · raw message
Signature · 1 line
2.4.10
Florian Westphal
maintainerQuoted 7 lines · Florian Westphal
nft fails to parse certain corner-cases, for example:
nft add rule filter input meta rtclassid daddr
... as it finds DADDR token. However, 'daddr' might be a valid routing realm listed in iproute2/rt_realms, so this should be allowed.
Quoted 6 lines
Pablo suggested to change the start conditions in the scanner accordingly.
After this patch, the following rule works:
ct label & (foobar | saddr) == saddr ip saddr 1.2.3.4 rtclassid { 42, cosmos, rule}
Note that this will not work:
ct label eq foobar
(we disabled eq token, eq is expected to be name of label).
&, ==, !=, etc. will continue to work.
Not sure if thats a bug or feature -- it would be easy to just remove the <INITIAL> from "eq" so that we continue to recognize it as "==", but it means that its not possible to use eq, lt, gt, and so on as usernames, rtclassids, etc etc.
Florian Westphal
maintainerQuoted 11 lines · Florian Westphal
Quoted 10 lines · Florian Westphal
nft fails to parse certain corner-cases, for example:
nft add rule filter input meta rtclassid daddr
... as it finds DADDR token. However, 'daddr' might be a valid routing realm listed in iproute2/rt_realms, so this should be allowed.
Pablo suggested to change the start conditions in the scanner accordingly.
Quoted 7 lines
Quoted 3 lines
After this patch, the following rule works:
ct label & (foobar | saddr) == saddr ip saddr 1.2.3.4 rtclassid { 42, cosmos, rule}
Note that this will not work:
ct label eq foobar
This patch doesn't work, we can return from literal mode too early.
rtclassid { rule, saddr }
we will leave literal mode after 'rule' so we choke on the SADDR token.
I'm flagging this as 'rejected' and will see if this can be fixed somehow.