Libor Peltan authored 3 years ago and Daniel Salzman committed 2 years ago

...as dynarray itself had been moved from contrib

JP Mens authored 2 years ago and Daniel Salzman committed 2 years ago

closes #801

The number of UMEM frames have been decreased as it doesn't seem it helped the performance.
As a consequence the overall memory consumption decreases.

Also a hint on the fill ring size from the xdpsock_user.c sample has been considered.

After the change the following assert sometimes occured when running
kxdpgun in the emulated XDP mode.

knot_xdp_recv_finish():
  const uint32_t reserved = xsk_ring_prod__reserve(fq, count, &idx);
  assert(reserved == count);

This reverts commit f5f4ee9f.

JP Mens authored 2 years ago and Daniel Salzman committed 2 years ago

closes #798

If a packet was successfully alloced by knot_xdp_send_alloc(),
there must be a room in the TX ring for it.

events: improve mutexing

See merge request !1454

Libor Peltan authored 2 years ago and Daniel Salzman committed 2 years ago

This fixes a bug when zone_reload_modules only waits for
an event to be finished, but following reschedule is not
synchronized and might access memory freed by zone_free.

tests-extra XFR refactoring

See merge request !1453

This is a slight simplification of the test.

This is needed even without Valgrind when running massively in parallel

On slaves one common ACL for all actions has been replaced with a sequence
of ACLs for notify, one rule for one master.

the heap-element-replace bug fix

See merge request !1452

David Vasek authored 2 years ago and Libor Peltan committed 2 years ago

This avoids the timeouts in master and as a result the "TCP, terminated inactive client"
messages in the master's log are eliminated. Slave seems to be busy for too long when signing
the zones under Valgrind.

David Vasek authored 2 years ago and Libor Peltan committed 2 years ago

In contrast with Knot coding style, most single line "if" constructions retained.

David Vasek authored 2 years ago and Libor Peltan committed 2 years ago

Possible as a result of modified heap_replace().

David Vasek authored 2 years ago and Libor Peltan committed 2 years ago

Current code as used by evsched_schedule() has led to a corrupted heap
(with its elements out of order), resulting in unpredictable event scheduling
behaviour, including events being processed out of order or already scheduled
events being postponed, sometimes for several days, weeks, or maybe longer.

Solving it by modifying the heap.c library a little -- the new code properly
handles a situation when the original element and the element it's being replaced by
are identical (that is, the same memory object), but already containing a modified
value than what originally was in the heap. Instead of comparing the new element
with the one it replaces, we now start comparing it with its parent.

This commit fixes an old bug introduced in:
925ad42d

It's said timercmp() == is broken in the same broken implementations too.
Furthemore, this expression is simpler and faster.
While there, fix some coding style.

MSAN is unable to correctly evaluate explicit_bzero() for memory zeroing in
wire_ctx operations. Then OSS-Fuzz creates false-positive issues. Using bzero()
seems to be a reasonable workaround.

zonedb: replan events for new zone_t instance no sooner...

See merge request !1451