The __builtin_expect() can be used to provide the compiler with branch
prediction information. The Gcc manual says[1] on the subject:
In general, you should prefer to use actual profile feedback for
this (-fprofile-arcs), as programmers are notoriously bad at
predicting how their programs actually perform.
Stop using __builtin_expect() and ISC_LIKELY() and ISC_UNLIKELY() macros
to provide the branch prediction information as the performance testing
shows that named performs better when the __builtin_expect() is not
being used.
1. https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html#index-_005f_005fbuiltin_005fexpect
Remove the dynamic registration of result codes. Convert isc_result_t
from unsigned + #defines into 32-bit enum type in grand unified
<isc/result.h> header. Keep the existing values of the result codes
even at the expense of the description and identifier tables being
unnecessary large.
Additionally, add couple of:
switch (result) {
[...]
default:
break;
}
statements where compiler now complains about missing enum values in the
switch statement.
Each dns_rpz_zone_t structure keeps a hash table of the names this RPZ
database contains. Here is what happens when an RPZ is updated:
- a new hash table is prepared for the new version of the RPZ by
iterating over it; each name found is added to the summary RPZ
database,
- every name added to the new hash table is searched for in the old
hash table; if found, it is removed from the old hash table,
- the old hash table is iterated over; all names found in it are
removed from the summary RPZ database (because at that point the old
hash table should only contain names which are not present in the
new version of the RPZ),
- the new hash table replaces the old hash table.
When the new version of the RPZ is iterated over, if a given name is
spelled using a different letter case than in the old version of the
RPZ, the new variant will hash to a different value than the old
variant, which means it will not be removed from the old hash table.
When the old hash table is subsequently iterated over to remove
seemingly deleted names, the old variant of the name will still be
there, causing the name to be deleted from the summary RPZ database
(which effectively causes a given rule to be ignored).
The issue can be triggered not just by altering the case of existing
names in an RPZ, but also by adding sibling names spelled with a
different letter case. This is because RBT code preserves case when
node splitting occurs. The end result is that when the RPZ is iterated
over, a given name may be using a different case than in the zone file
(or XFR contents).
Fix by downcasing all names found in the RPZ database before adding them
to the summary RPZ database.
Created isc_refcount_decrement_expect macro to test conditionally
the return value to ensure it is in expected range. Converted
unchecked isc_refcount_decrement to use isc_refcount_decrement_expect.
Converted INSIST(isc_refcount_decrement()...) to isc_refcount_decrement_expect.
Whenever an exact match is found by dns_rbt_findnode(),
the highest level node in the chain will not be put into
chain->levels[] array, but instead the chain->end
pointer will be adjusted to point to that node.
Suppose we have the following entries in a rpz zone:
example.com CNAME rpz-passthru.
*.example.com CNAME rpz-passthru.
A query for www.example.com would result in the
following chain object returned by dns_rbt_findnode():
chain->level_count = 2
chain->level_matches = 2
chain->levels[0] = .
chain->levels[1] = example.com
chain->levels[2] = NULL
chain->end = www
Since exact matches only care for testing rpz set bits,
we need to test for rpz wild bits through iterating the nodechain, and
that includes testing the rpz wild bits in the highest level node found.
In the case of an exact match, chain->levels[chain->level_matches]
will be NULL, to address that we must use chain->end as the start point,
then iterate over the remaining levels in the chain.
this addresses a race that could occur during shutdown or when
reconfiguring to remove RPZ zones.
this change should ensure that the rpzs structure and the incremental
updates don't interfere with each other: rpzs->zones entries cannot
be set to NULL while an update quantum is running, and the
task should be destroyed and its queue purged so that no subsequent
quanta will run.
After an RPZ zone is updated via zone transfer, the RPZ summary
database is updated, inserting the newly added names in the policy
zone and deleting the newly removed ones. The first part of this
was quantized so it would not run too long and starve other tasks
during large updates, but the second part was not quantized, so
that an update in which a large number of records were deleted
could cause named to become briefly unresponsive.
The 3 warnings reported are:
os.c:872:7: warning: Although the value stored to 'ptr' is used in the enclosing expression, the value is never actually read from 'ptr'
if ((ptr = strtok_r(command, " \t", &last)) == NULL) {
^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1 warning generated.
--
rpz.c:1117:10: warning: Although the value stored to 'zbits' is used in the enclosing expression, the value is never actually read from 'zbits'
return (zbits &= x);
^ ~
1 warning generated.
--
openssleddsa_link.c:532:10: warning: Although the value stored to 'err' is used in the enclosing expression, the value is never actually read from 'err'
while ((err = ERR_get_error()) != 0) {
^ ~~~~~~~~~~~~~~~
1 warning generated.
This commits removes superfluous checks when using the isc_refcount API.
Examples of superfluous checks:
1. The isc_refcount_decrement function ensures there was not underflow,
so this check is superfluous:
INSIST(isc_refcount_decrement(&r) > 0);
2 .The isc_refcount_destroy() includes check whether the counter
is zero, therefore this is superfluous:
INSIST(isc_refcount_decrement(&r) == 1 && isc_refcount_destroy(&r));
The isc_refcount_decrement() was either used as:
if (isc_refcount_decrement() == 1) { destroy(); }
or
if (isc_refcount_decrement() != 1) { return; } destroy();
This commits eradicates the last usage of the later, so the code is unified to
use the former.
when looking for a possible wildcard match in the RPZ summary database,
use an rbtnodechain to walk up label by label, rather than using the
node's parent pointer.
In dns_rpz_update_from_db we call setup_update which creates the db
iterator and calls dns_dbiterator_first. This unpauses the iterator and
might cause db->tree_lock to be acquired. We then do isc_task_send(...)
on an event to do quantum_update, which (correctly) after each iteration
calls dns_dbiterator_pause, and re-isc_task_sends itself.
That's an obvious bug, as we're holding a lock over an async task send -
if a task requesting write (e.g. prune_tree) is scheduled on the same
workers queue as update_quantum but before it, it will wait for the
write lock indefinitely, resulting in a deadlock.
To fix it we have to pause dbiterator in setup_update.
As pointed out in !813 db_registered is sort of redundant. It is
set to `true` only in `dns_zone_rpz_enable_db()` right before the
`dns_rpz_dbupdate_callback()` callback is registered. It is only
required in that callback and it is the only place that the callback
is registered. Therefore there is no path that that `REQUIRE` can
fail.
The `db_registered` variable is only set to `false` in
`dns_rpz_new_zone`, so it is not like the variable is unset again
later.
The only other place where `db_registered` is checked is in
`rpz_detach()`. If `true`, it will call
`dns_db_updatenotify_unregister()`. However if that happens, the
`db_registered` is not set back to `false` thus this implies that
this may happen multiple times. If called a second time, most
likely the unregister function will return `ISC_R_NOTFOUND`, but
the return value is not checked anyway. So it can do without the
`db_registered` check.
This may happen when loading an RPZ failed and the code path skips
calling dns_db_endload(). The dns_rpz_zone_t object is still kept
marked as having registered db. So when this object is finally
destroyed in rpz_detach(), this code will incorrectly call
`dns_db_updatenotify_unregister()`:
if (rpz->db_registered)
dns_db_updatenotify_unregister(rpz->db,
dns_rpz_dbupdate_callback, rpz);
and trigger this assertion failure:
REQUIRE(db != NULL);
To fix this, only call `dns_db_updatenotify_unregister()` when
`rpz->db` is not NULL.
While implementing the new unit testing framework cmocka, it was found that the
BIND 9 code doesn't compile when assertions are disabled or replaced with any
function (such as mock_assert() from cmocka unit testing framework) that's not
directly recognized as assertion by the compiler.
This made the compiler to complain about blocks of code that was recognized as
unreachable before, but now it isn't.
The changes in this commit include:
* assigns default values to couple of local variables,
* moves some return statements around INSIST assertions,
* adds __builtin_unreachable(); annotations after some INSIST assertions,
* fixes one broken assertion (= instead of ==)
If an RPZ zone is to be freed during an update, canceling the
update_quantum() event is not enough because the resources released when
an update completes also need to be accounted for. Failure to do this
results in a hang upon shutdown. Fix by copying cleanup code from the
end of update_quantum() to rpz_detach().
If another RPZ update is pending when processing the previous one nears
completion and min-update-interval is set to 0, isc_timer_reset() gets
called with 'interval' set to 0, which triggers an assertion failure.
To prevent such a scenario from causing a crash, queue the update event
directly instead of asking the timer thread to do it.
This properly orders clearing the freed pointer and calling isc_refcount_destroy
as early as possible to have ability to put proper memory barrier when cleaning
up reference counting.
