The original algorithm was suffering from an ABA race condition:
A: fp = page_stack
B: completely allocates the same page and writes into it some data
A: unsuspecting, loads (invalid) next = fp->next
B: finishes working with the page and returns it back to page_stack
A: compare-exchange page_stack: fp => next succeeds and writes garbage
to page_stack
Fixed this by using an implicit spinlock in hot page allocator.
If a thread encounters timeout == 0 for poll, it considers itself
"busy" and with some hysteresis it tries to drop loops for others to
pick and thus better distribute work between threads.
Memory allocation is a fragile part of BIRD and we need checking that
everybody is using the resource pools in an appropriate way. To assure
this, all the resource pools are associated with locking domains and
every resource manipulation is thoroughly checked whether the
appropriate locking domain is locked.
With transitive resource manipulation like resource dumping or mass free
operations, domains are locked and unlocked on the go, thus we require
pool domains to have higher order than their parent to allow for this
transitive operations.
Adding pool locking revealed some cases of insecure memory manipulation
and this commit fixes that as well.
Now sk_open() requires an explicit IO loop to open the socket in. Also
specific functions for socket RX pause / resume are added to allow for
BGP corking.
And last but not least, socket reloop is now synchronous to resolve
weird cases of the target loop stopping before actually picking up the
relooped socket. Now the caller must ensure that both loops are locked
while relooping, and this way all sockets always have their respective
loop.
If there are lots of loops in a single thread and only some of the loops
are actually active, the other loops are now kept aside and not checked
until they actually get some timers, events or active sockets.
This should help with extreme loads like 100k tables and protocols.
Also ping and loop pickup mechanism was allowing subtle race
conditions. Now properly handling collisions between loop ping and pickup.
Instead of propagating interface updates as they are loaded from kernel,
they are enqueued and all the notifications are called from a
protocol-specific event. This change allows to break the locking loop
between protocols and interfaces.
Anyway, this change is based on v2 branch to keep the changes between v2
and v3 smaller.
The interface list must be flushed when device protocol is stopped. This
was done in a hardcoded specific hook inside generic protocol routines.
The cleanup hook was originally used for table reference counting late
cleanup, yet it can be also simply used for prettier interface list flush.
On large configurations, too many threads would spawn with one thread
per loop. Therefore, threads may now run multiple loops at once. The
thread count is configurable and may be changed during run. All threads
are spawned on startup.
This change helps with memory bloating. BIRD filters need large
temporary memory blocks to store their stack and also memory management
keeps its hot page storage per-thread.
Known bugs:
* Thread autobalancing is not yet implemented.
* Low latency loops are executed together with standard loops.
Some CLI actions, notably "show route", are run by queuing an event
somewhere else. If the user closes the socket, in case such an action is
being executed, the CLI must free the socket immediately from the error
hook but the pool must remain until the asynchronous event finishes and
cleans everything up.
When BIRD has no free memory mapped, it allocates several pages in
advance just to be sure that there is some memory available if needed.
This hysteresis tactics works quite well to reduce memory ping-ping with
kernel.
Yet it had a subtle bug: this pre-allocation didn't take a memory
coldlist into account, therefore requesting new pages from kernel even
in cases when there were other pages available. This led to slow memory
bloating.
To demonstrate this behavior fast enough to be seen well, you may:
* temporarily set the values in sysdep/unix/alloc.c as follows to
exacerbate the issue:
#define KEEP_PAGES_MAIN_MAX 4096
#define KEEP_PAGES_MAIN_MIN 1000
#define CLEANUP_PAGES_BULK 4096
* create a config file with several millions of static routes
* periodically disable all static protocols and then reload config
* log memory consumption
This should give you a steady growth rate of about 16kB per cycle. If
you don't set the values this high, the issue happens much more slowly,
yet after 14 days of running, you are going to see an OOM kill.
After this fix, pre-allocation uses the memory coldlist to get some hot
pages and the same test as described here gets you a perfectly stable
constant memory consumption (after some initial wobbling).
Thanks to NIX-CZ for reporting and helping to investigate this issue.
Thanks to Santiago for finding the cause in the code.
The usage pattern implemented in allocator seems to be incompatible with
transparent huge pages, as memory released using madvise(MADV_DONTNEED)
with regular page size and alignment does not seem to trigger demotion
of huge pages back to regular pages, even when significant number of
pages is released. Even if demotion is triggered when system memory
is low, it still breaks memory accounting.
Log message before aborting due to watchdog timeout. We have to use
async-safe write to debug log, as it is done in signal handler.
Minor changes from committer.
When there is a continuos stream of CLI commands, cli_get_command()
always returns 1 (there is a new command). Anyway, the socket receive
buffer was reset only when there was no command at all, leading to a
strange behavior: after a while, the CLI receive buffer came to its end,
then read() was called with zero size buffer, it returned 0 which was
interpreted as EOF.
The patch fixes that by resetting the buffer position after each command
and moving remaining data at the beginning of buffer.
Thanks to Maria Matejka for examining the bug and for the original bugfix.
BIRD keeps a previous (old) configuration for the purpose of undo. The
existing code frees it after a new configuration is successfully parsed
during reconfiguration. That causes memory usage spikes as there are
temporarily three configurations (old, current, and new). The patch
changes it to free the old one before parsing the new one (as user
already requested a new config). The disadvantage is that undo is
not available after failed reconfiguration.
Memory unmapping causes slow address space fragmentation, leading in
extreme cases to failing to allocate pages at all. Removing this problem
by keeping all the pages allocated to us, yet calling madvise() to let
kernel dispose of them.
This adds a little complexity and overhead as we have to keep the
pointers to the free pages, therefore to hold e.g. 1 GB of 4K pages with
8B pointers, we have to store 2 MB of data.
This is a reimplementation of commit 0f2be469f8
by Alexander Zubkov. In the master branch, changes in commit eb937358
broke setting of channel preference for alien routes learned during
scan. The preference was set only for async routes.
The original solution is extended here to accomodate for v3 specifics.
Changes in commit eb937358 broke setting of channel preference for alien
routes learned during scan. The preference was set only for async routes.
Move common attribute processing part of functions krt_learn_async() and
krt_learn_async() to a separate function to have only one place for such
changes.
Remove compile-time sysdep option CONFIG_ALL_TABLES_AT_ONCE, replace it
with runtime ability to run either separate table scans or shared scan.
On Linux, use separate table scans by default when the netlink socket
option NETLINK_GET_STRICT_CHK is available, but retreat to shared scan
when it fails.
Running separate table scans has advantages where some routing tables are
managed independently, e.g. when multiple routing daemons are running on
the same machine, as kernel routing table modification performance is
significantly reduced when the table is modified while it is being
scanned.
Thanks Daniel Gröber for the original patch and Toke Høiland-Jørgensen
for suggestions.
The learnt routes are now pushed all into the connected table, not only
the best one. This shouldn't do any damage in well managed setups, yet
it should be noted that it is a change of behavior.
If anybody misses a feature which they implemented by misusing this
internal learn table, let us know, we'll consider implementing it in a
better way.
Passing protocol to preexport was in fact a historical relic from the
old times when channels weren't a thing. Refactoring that to match
current extensibility needs.
There were quite a lot of conflicts in flowspec validation code which
ultimately led to some code being a bit rewritten, not only adapted from
this or that branch, yet it is still in a limit of a merge.
For now, all route attributes are stored as eattrs in ea_list. This
should make route manipulation easier and it also allows for a layered
approach of route attributes where updates from filters will be stored
as an overlay over the previous version.
As there is either a nexthop or another destination specification
(or othing in case of ROAs and Flowspec), it may be merged together.
This code is somehow quirky and should be replaced in future by better
implementation of nexthop.
Also flowspec validation result has its own attribute now as it doesn't
have anything to do with route nexthop.
This doesn't do anything more than to put the whole structure inside
adata. The overall performance is certainly going downhill; we'll
optimize this later.
Anyway, this is one of the latest items inside rta and in several
commits we may drop rta completely and move to eattrs-only routes.
Some tokens are both keywords and symbols. For now, we allow only
specific keywords to be redefined; in future, more of the keywords may
be added to this category.
The redefinable keywords must be specified in any .Y file as follows:
toksym: THE_KEYWORD ;
See proto/bgp/config.Y for an example.
Also dropped a lot of unused terminals.
Changes in internal API:
* Every route attribute must be defined as struct ea_class somewhere.
* Registration of route attributes known at startup must be done by
ea_register_init() from protocol build functions.
* Every attribute has now its symbol registered in a global symbol table
defined as SYM_ATTRIBUTE
* All attribute ID's are dynamically allocated.
* Attribute value custom formatting hook is defined in the ea_class.
* Attribute names are the same for display and filters, always prefixed
by protocol name.
Also added some unit testing code for filters with route attributes.
This commit removes the EAF_TYPE_* namespace completely and also for
route attributes, filter-based types T_* are used. This simplifies
fetching and setting route attributes from filters.
Also, there is now union bval which serves as an universal value holder
instead of private unions held separately by eattr and filter code.
When BIRD was munmapping too many pages, it sometimes aborted, saying
that munmap failed with "Not enough memory" as the address space was
getting more and more fragmented.
There is a workaround in place, simply keeping that page for future use,
yet it has never been compiled in because I somehow forgot to include
errno.h. And because I also thought that somebody may have ENOMEM not
defined (why?!), there was a check which quietly omitted that
workaround.
Anyway, ENOMEM is POSIX. It's an utter nonsense to check for its
existence. If it doesn't exist, something is broken.
Add option to socket interface for nonlocal binding, i.e. binding to an
IP address that is not present on interfaces. This behaviour is enabled
when SKF_FREEBIND socket flag is set. For Linux systems, it is
implemented by IP_FREEBIND socket flag.
Minor changes done by commiter.
This commit prevents use-after-free of routes belonging to protocols
which have been already destroyed, delaying also all the protocols'
shutdown until all of their routes have been finally propagated through
all the pipes down to the appropriate exports.
The use-after-free was somehow hypothetic yet theoretically possible in
rare conditions, when one BGP protocol authors a lot of routes and the
user deletes that protocol by reconfiguring in the same time as next hop
update is requested, causing rte_better() to be called on a
not-yet-pruned network prefix while the owner protocol has been already
freed.
In parallel execution environments, this would happen an inter-thread
use-after-free, causing possible heisenbugs or other nasty problems.
There is a simple universal IO loop, taking care of events, timers and
sockets. Primarily, one instance of a protocol should use exactly one IO
loop to do all its work, as is now done in BFD.
Contrary to previous versions, the loop is now launched and cleaned by
the nest/proto.c code, allowing for a protocol to just request its own
loop by setting the loop's lock order in config higher than the_bird.
It is not supported nor checked if any protocol changed the requested
lock order in reconfigure. No protocol should do it at all.
In previous versions, every thread used its own time structures,
effectively leading to different time in every thread and strange
logging messages.
The time processing code now uses global atomic variables to keep
current time available for fast concurrent reading and safe updates.
* internal tables are now more standalone, having their own import and
export hooks
* route refresh/reload uses stale counter instead of stale flag,
allowing to drop walking the table at the beginning
* route modify (by BGP LLGR) is now done by a special refeed hook,
reimporting the modified routes directly without filters
Channels have now included rt_import_req and rt_export_req to hook into
the table instead of just one list node. This will (in future) allow for:
* channel import and export bound to different tables
* more efficient pipe code (dropping most of the channel code)
* conversion of 'show route' to a special kind of export
* temporary static routes from CLI
The import / export states are also updated to the new algorithms.
Routes are now allocated only when they are just to be inserted to the
table. Updating a route needs a locally allocated route structure.
Ownership of the attributes is also now not transfered from protocols to
tables and vice versa but just borrowed which should be easier to handle
in a multithreaded environment.