This change should make it impossible for the monotonic time to roll
over from one EWMA tick to the next during this test, and make it so
that this test never invokes scale_active_circuits() (which it
doesn't test).
(Earlier changes during the 0.3.4 series should make this call even
rarer than it was before, since we fixed#25927 and removed
cached_gettimeofday. Because this test didn't update
cached_gettimeofday, the chance of rolling over a 10-second interval
was much higher.)
Closes ticket 25994; bugfix on 0.3.3.1-alpha when this test was
introduced.
This part of the code was the only part that used "cached
getttimeofday" feature, which wasn't monotonic, which we updated at
slight expense, and which I'd rather not maintain.
* ADD new /src/common/crypto_rand.[ch] module.
* ADD new /src/common/crypto_util.[ch] module (contains the memwipe()
function, since all crypto_* modules need this).
* FIXES part of #24658: https://bugs.torproject.org/24658
Since 0.2.4, tor uses EWMA circuit policy to prioritize. The previous
algorithm, round-robin, hasn't been used since then but was still used as a
fallback.
Now that EWMA is mandatory, remove that code entirely and enforce a cmux
policy to be set.
This is part of a circuitmux cleanup to improve performance and reduce
complexity in the code. We'll be able to address future optimization with this
work.
Closes#25268
Signed-off-by: David Goulet <dgoulet@torproject.org>
To achieve this, a default value for the CircuitPriorityHalflife option was
needed. We still look in the options and then the consensus but in case no
value can be found, the default CircuitPriorityHalflifeMsec=30000 is used. It
it the value we've been using since 0.2.4.4-alpha.
This means that EWMA, our only policy, can not be disabled anymore fallbacking
to the round robin algorithm. Unneeded code to control that is removed in this
commit.
Part of #25268
Signed-off-by: David Goulet <dgoulet@torproject.org>
Using absolute_msec requires a 64-bit division operation every time
we calculate it, which gets expensive on 32-bit architectures.
Instead, just use the lazy "monotime_coarse_get()" operation, and
don't convert to milliseconds until we absolutely must.
In this case, it seemed fine to use a full monotime_coarse_t rather
than a truncated "stamp" as we did to solve this problem for the
timerstamps in buf_t and packed_cell_t: There are vastly more cells
and buffer chunks than there are channels, and using 16 bytes per
channel in the worst case is not a big deal.
There are still more millisecond operations here than strictly
necessary; let's see any divisions show up in profiles.
Couple things happen in this commit. First, we do not re-queue a cell back in
the circuit queue if the write packed cell failed. Currently, it is close to
impossible to have it failed but just in case, the channel is mark as closed
and we move on.
The second thing is that the channel_write_packed_cell() always took ownership
of the cell whatever the outcome. This means, on success or failure, it needs
to free it.
It turns out that that we were using the wrong free function in one case and
not freeing it in an other possible code path. So, this commit makes sure we
only free it in one place that is at the very end of
channel_write_packed_cell() which is the top layer of the channel abstraction.
This makes also channel_tls_write_packed_cell_method() return a negative value
on error.
Two unit tests had to be fixed (quite trivial) due to a double free of the
packed cell in the test since now we do free it in all cases correctly.
Part of #23709
Signed-off-by: David Goulet <dgoulet@torproject.org>
First, that test was broken from the previous commit because the
channel_queue_cell() has been removed. This now tests the
channel_process_cell() directly.
Second, it wasn't testing much except if the channel subsystem actually went
through the cell handler. This commit adds more checks on the state of a
channel going from open, receiving a cell and closing.
Third, this and the id_map unit test are working, not the others so they've
been marked as not working and future commit will improve and fix those.
Signed-off-by: David Goulet <dgoulet@torproject.org>
The channel_write_cell() and channel_write_var_cell() can't be possibly called
nor are used by tor. We only write on the connection outbuf packed cell coming
from the scheduler that takes them from the circuit queue.
This makes channel_write_packed_cell() the only usable function. It is
simplify and now returns a code value. The reason for this is that in the next
commit(s), we'll re-queue the cell onto the circuit queue if the write fails.
Finally, channel unit tests are being removed with this commit because they do
not match the new semantic. They will be re-written in future commits.
Signed-off-by: David Goulet <dgoulet@torproject.org>
The channel subsystem was doing a whole lot to track and try to predict the
channel queue size but they are gone due to previous commit.
Signed-off-by: David Goulet <dgoulet@torproject.org>
For the rationale, see ticket #23709.
This is a pretty massive commit. Those queues were everywhere in channel.c and
it turns out that it was used by lots of dead code.
The channel subsystem *never* handles variable size cell (var_cell_t) or
unpacked cells (cell_t). The variable ones are only handled in channeltls and
outbound cells are always packed from the circuit queue so this commit removes
code related to variable and unpacked cells.
However, inbound cells are unpacked (cell_t), that is untouched and is handled
via channel_process_cell() function.
In order to make the commit compile, test have been modified but not passing
at this commit. Also, many tests have been removed but better improved ones
get added in future commits.
This commit also adds a XXX: which indicates that the handling process of
outbound cells isn't fully working. This as well is fixed in a future commit.
Finally, at this commit, more dead code remains, it will be cleanup in future
commits.
Fixes#23709
Signed-off-by: David Goulet <dgoulet@torproject.org>
The `test-operator-cleanup` patch, and related coccinelle patches,
don't do any checks for line length. This patch fixes the line
length issues caused by the previous commits.
This patch fixes the operator usage in src/test/*.c to use the symbolic
operators instead of the normal C comparison operators.
This patch was generated using:
./scripts/coccinelle/test-operator-cleanup src/test/*.[ch]
This was breaking the build on debian precise, since it thought that
using a 'const int' to dimension an array made that array
variable-size, and made us not get protection.
Bug not in any released version of Tor.
I will insist that this one wasn't my fault.
"Variables won't. Constants aren't." -- Osborn's Law
This is a big-ish patch, but it's very straightforward. Under this
clang warning, we're not actually allowed to have a global variable
without a previous extern declaration for it. The cases where we
violated this rule fall into three roughly equal groups:
* Stuff that should have been static.
* Stuff that was global but where the extern was local to some
other C file.
* Stuff that was only global when built for the unit tests, that
needed a conditional extern in the headers.
The first two were IMO genuine problems; the last is a wart of how
we build tests.
This warning triggers on silently promoting a float to a double. In
our code, it's just a sign that somebody used a float by mistake,
since we always prefer double.
