summaryrefslogtreecommitdiff
path: root/doc/threads
diff options
context:
space:
mode:
authorMaria Matejka <mq@ucw.cz>2021-03-04 20:11:26 +0100
committerMaria Matejka <mq@ucw.cz>2021-12-08 12:39:28 +0100
commitf459deee9f494ae8f741cc3b04e0e0db94a77c53 (patch)
tree0770c181ad614d401b8145ac77104327af17bb2b /doc/threads
parent493d45d95076ea1f375346942a1ce4e21dcc55e6 (diff)
Thread documentation: chapters 0, 1 and 2
Diffstat (limited to 'doc/threads')
-rw-r--r--doc/threads/00_filter_structure.pngbin0 -> 307458 bytes
-rw-r--r--doc/threads/00_the_name_of_the_game.md114
-rw-r--r--doc/threads/01_the_route_and_its_attributes.md159
-rw-r--r--doc/threads/02_asynchronous_export.md437
-rw-r--r--doc/threads/Makefile15
5 files changed, 725 insertions, 0 deletions
diff --git a/doc/threads/00_filter_structure.png b/doc/threads/00_filter_structure.png
new file mode 100644
index 00000000..a61603cb
--- /dev/null
+++ b/doc/threads/00_filter_structure.png
Binary files differ
diff --git a/doc/threads/00_the_name_of_the_game.md b/doc/threads/00_the_name_of_the_game.md
new file mode 100644
index 00000000..ddc4f638
--- /dev/null
+++ b/doc/threads/00_the_name_of_the_game.md
@@ -0,0 +1,114 @@
+# BIRD Journey to Threads. Chapter 0: The Reason Why.
+
+BIRD is a fast, robust and memory-efficient routing daemon designed and
+implemented at the end of 20th century. Its concept of multiple routing
+tables with pipes between them, as well as a procedural filtering language,
+has been unique for a long time and is still one of main reasons why people use
+BIRD for big loads of routing data.
+
+## IPv4 / IPv6 duality: Solved
+
+The original design of BIRD has also some drawbacks. One of these was an idea
+of two separate daemons – one BIRD for IPv4 and another BIRD for IPv6, built from the same
+codebase, cleverly using `#ifdef IPV6` constructions to implement the
+common parts of BIRD algorithms and data structures only once.
+If IPv6 adoption went forward as people thought in that time,
+it would work; after finishing the worldwide transition to IPv6, people could
+just stop building BIRD for IPv4 and drop the `#ifdef`-ed code.
+
+The history went other way, however. BIRD developers therefore decided to *integrate*
+these two versions into one daemon capable of any address family, allowing for
+not only IPv6 but for virtually anything. This rework brought quite a lot of
+backward-incompatible changes, therefore we decided to release it as a version 2.0.
+This work was mostly finished in 2018 and as for March 2021, we have already
+switched the 1.6.x branch to a bugfix-only mode.
+
+## BIRD is single-threaded now
+
+The second drawback is a single-threaded design. Looking back to 1998, this was
+a good idea. A common PC had one single core and BIRD was targeting exactly
+this segment. As the years went by, the manufacturers launched multicore x86 chips
+(AMD Opteron in 2004, Intel Pentium D in 2005). This ultimately led to a world
+where as of March 2021, there is virtually no new PC sold with a single-core CPU.
+
+Together with these changes, the speed of one single core has not been growing as fast
+as the Internet is growing. BIRD is still capable to handle the full BGP table
+(868k IPv4 routes in March 2021) with one core, anyway when BIRD starts, it may take
+long minutes to converge.
+
+## Intermezzo: Filters
+
+In 2018, we took some data we had from large internet exchanges and simulated
+a cold start of BIRD as a route server. We used `linux-perf` to find most time-critical
+parts of BIRD and it pointed very clearly to the filtering code. It also showed that the
+IPv4 version of BIRD v1.6.x is substantially faster than the *integrated* version, while
+the IPv6 version was quite as fast as the *integrated* one.
+
+Here we should show a little bit more about how the filters really work. Let's use
+an example of a simple filter:
+
+```
+filter foo {
+ if net ~ [10.0.0.0/8+] then reject;
+ preference = 2 * preference - 41;
+ accept;
+}
+```
+
+This filter gets translated to an infix internal structure.
+
+![Example of filter internal representation](00_filter_structure.png)
+
+When executing, the filter interpreter just walks the filter internal structure recursively in the
+right order, executes the instructions, collects their results and finishes by
+either rejection or acceptation of the route
+
+## Filter rework
+
+Further analysis of the filter code revealed an absurdly-looking result. The
+most executed parts of the interpreter function were the `push` CPU
+instructions on its very beginning and the `pop` CPU instructions on its very
+end. This came from the fact that the interpreter function was quite long, yet
+most of the filter instructions used an extremely short path, doing all the
+stack manipulation at the beginning, branching by the filter instruction type,
+then it executed just several CPU instructions, popped everything from the
+stack back and returned.
+
+After some thoughts how to minimize stack manipulation when everything you need
+is to take two numbers and multiply them, we decided to preprocess the filter
+internal structure to another structure which is much easier to execute. The
+interpreter is now using a data stack and behaves generally as a
+postfix-ordered language. We also thought about Lua which showed up to be quite
+a lot of work implementing all the glue achieving about the same performance.
+
+After these changes, we managed to reduce the filter execution time by 10–40%,
+depending on how complex the filter is.
+Anyway, even this reduction is quite too little when there is one CPU core
+running for several minutes while others are sleeping.
+
+## We need more threads
+
+As a side effect of the rework, the new filter interpreter is also completely
+thread-safe. It seemed to be the way to go – running the filters in parallel
+while keeping everything else single-threaded. The main problem of this
+solution is a too fine granularity of parallel jobs. We would spend lots of
+time on synchronization overhead.
+
+The only filter parallel execution was also too one-sided, useful only for
+configurations with complex filters. In other cases, the major problem is best
+route recalculation, OSPF recalculation or also kernel synchronization.
+It also turned out to be dirty a lot from the code cleanliness' point of view.
+
+Therefore we chose to make BIRD multithreaded completely. We designed a way how
+to gradually enable parallel computation and best usage of all available CPU
+cores. Our goals are three:
+
+* We want to keep current functionality. Parallel computation should never drop
+ a useful feature.
+* We want to do little steps. No big reworks, even though even the smallest
+ possible step will need quite a lot of refactoring before.
+* We want to be backwards compatible as much as possible.
+
+*It's still a long road to the version 2.1. This series of texts should document
+what is needed to be changed, why we do it and how. In the next chapter, we're
+going to describe the structures for routes and their attributes. Stay tuned!*
diff --git a/doc/threads/01_the_route_and_its_attributes.md b/doc/threads/01_the_route_and_its_attributes.md
new file mode 100644
index 00000000..079eb365
--- /dev/null
+++ b/doc/threads/01_the_route_and_its_attributes.md
@@ -0,0 +1,159 @@
+# BIRD Journey to Threads. Chapter 1: The Route and its Attributes
+
+BIRD is a fast, robust and memory-efficient routing daemon designed and
+implemented at the end of 20th century. We're doing a significant amount of
+BIRD's internal structure changes to make it possible to run in multiple
+threads in parallel. This chapter covers necessary changes of data structures
+which store every single routing data.
+
+*If you want to see the changes in code, look (basically) into the
+`route-storage-updates` branch. Not all of them are already implemented, anyway
+most of them are pretty finished as of end of March, 2021.*
+
+## How routes are stored
+
+BIRD routing table is just a hierarchical noSQL database. On top level, the
+routes are keyed by their destination, called *net*. Due to historic reasons,
+the *net* is not only *IPv4 prefix*, *IPv6 prefix*, *IPv4 VPN prefix* etc.,
+but also *MPLS label*, *ROA information* or *BGP Flowspec record*. As there may
+be several routes for each *net*, an obligatory part of the key is *src* aka.
+*route source*. The route source is a tuple of the originating protocol
+instance and a 32-bit unsigned integer. If a protocol wants to withdraw a route,
+it is enough and necessary to have the *net* and *src* to identify what route
+is to be withdrawn.
+
+The route itself consists of (basically) a list of key-value records, with
+value types ranging from a 16-bit unsigned integer for preference to a complex
+BGP path structure. The keys are pre-defined by protocols (e.g. BGP path or
+OSPF metrics), or by BIRD core itself (preference, route gateway).
+Finally, the user can declare their own attribute keys using the keyword
+`attribute` in config.
+
+## Attribute list implementation
+
+Currently, there are three layers of route attributes. We call them *route*
+(*rte*), *attributes* (*rta*) and *extended attributes* (*ea*, *eattr*).
+
+The first layer, *rte*, contains the *net* pointer, several fixed-size route
+attributes (mostly preference and protocol-specific metrics), flags, lastmod
+time and a pointer to *rta*.
+
+The second layer, *rta*, contains the *src* (a pointer to a singleton instance),
+a route gateway, several other fixed-size route attributes and a pointer to
+*ea* list.
+
+The third layer, *ea* list, is a variable-length list of key-value attributes,
+containing all the remaining route attributes.
+
+Distribution of the route attributes between the attribute layers is somehow
+arbitrary. Mostly, in the first and second layer, there are attributes that
+were thought to be accessed frequently (e.g. in best route selection) and
+filled in in most routes, while the third layer is for infrequently used
+and/or infrequently accessed route attributes.
+
+## Attribute list deduplication
+
+When protocols originate routes, there are commonly more routes with the
+same attribute list. BIRD could ignore this fact, anyway if you have several
+tables connected with pipes, it is more memory-efficient to store the same
+attribute lists only once.
+
+Therefore, the two lower layers (*rta* and *ea*) are hashed and stored in a
+BIRD-global database. Routes (*rte*) contain a pointer to *rta* in this
+database, maintaining a use-count of each *rta*. Attributes (*rta*) contain
+a pointer to normalized (sorted by numerical key ID) *ea*.
+
+## Attribute list rework
+
+The first thing to change is the distribution of route attributes between
+attribute list layers. We decided to make the first layer (*rte*) only the key
+and other per-record internal technical information. Therefore we move *src* to
+*rte* and preference to *rta* (beside other things). *This is already done.*
+
+We also found out that the nexthop (gateway), originally one single IP address
+and an interface, has evolved to a complex attribute with several sub-attributes;
+not only considering multipath routing but also MPLS stacks and other per-route
+attributes. This has led to a too complex data structure holding the nexthop set.
+
+We decided finally to squash *rta* and *ea* to one type of data structure,
+allowing for completely dynamic route attribute lists. This is also supported
+by adding other *net* types (BGP FlowSpec or ROA) where lots of the fields make
+no sense at all, yet we still want to use the same data structures and implementation
+as we don't like duplicating code. *Multithreading doesn't depend on this change,
+anyway this change is going to happen soon anyway.*
+
+## Route storage
+
+The process of route import from protocol into a table can be divided into several phases:
+
+1. (In protocol code.) Create the route itself (typically from
+ protocol-internal data) and choose the right channel to use.
+2. (In protocol code.) Create the *rta* and *ea* and obtain an appropriate
+ hashed pointer. Allocate the *rte* structure and fill it in.
+3. (Optionally.) Store the route to the *import table*.
+4. Run filters. If reject, free everything.
+5. Check whether this is a real change (it may be idempotent). If not, free everything and do nothing more.
+6. Run the best route selection algorithm.
+7. Execute exports if needed.
+
+We found out that the *rte* structure allocation is done too early. BIRD uses
+global optimized allocators for fixed-size blocks (which *rte* is) to reduce
+its memory footprint, therefore the allocation of *rte* structure would be a
+synchronization point in multithreaded environment.
+
+The common code is also much more complicated when we have to track whether the
+current *rte* has to be freed or not. This is more a problem in export than in
+import as the export filter can also change the route (and therefore allocate
+another *rte*). The changed route must be therefore freed after use. All the
+route changing code must also track whether this route is writable or
+read-only.
+
+We therefore introduce a variant of *rte* called *rte_storage*. Both of these
+hold the same, the layer-1 route information (destination, author, cached
+attribute pointer, flags etc.), anyway *rte* is always local and *rte_storage*
+is intended to be put in global data structures.
+
+This change allows us to remove lots of the code which only tracks whether any
+*rte* is to be freed as *rte*'s are almost always allocated on-stack, naturally
+limiting their lifetime. If not on-stack, it's the responsibility of the owner
+to free the *rte* after import is done.
+
+This change also removes the need for *rte* allocation in protocol code and
+also *rta* can be safely allocated on-stack. As a result, protocols can simply
+allocate all the data on stack, call the update routine and the common code in
+BIRD's *nest* does all the storage for them.
+
+Allocating *rta* on-stack is however not required. BGP and OSPF use this to
+import several routes with the same attribute list. In BGP, this is due to the
+format of BGP update messages containing first the attributes and then the
+destinations (BGP NLRI's). In OSPF, in addition to *rta* deduplication, it is
+also presumed that no import filter (or at most some trivial changes) is applied
+as OSPF would typically not work well when filtered.
+
+*This change is already done.*
+
+## Route cleanup and table maintenance
+
+In some cases, the route update is not originated by a protocol/channel code.
+When the channel shuts down, all routes originated by that channel are simply
+cleaned up. Also routes with recursive routes may get changed without import,
+simply by changing the IGP route.
+
+This is currently done by a `rt_event` (see `nest/rt-table.c` for source code)
+which is to be converted to a parallel thread, running when nobody imports any
+route. *This change is freshly done in branch `guernsey`.*
+
+## Parallel protocol execution
+
+The long-term goal of these reworks is to allow for completely independent
+execution of all the protocols. Typically, there is no direct interaction
+between protocols; everything is done thought BIRD's *nest*. Protocols should
+therefore run in parallel in future and wait/lock only when something is needed
+to do externally.
+
+We also aim for a clean and documented protocol API.
+
+*It's still a long road to the version 2.1. This series of texts should document
+what is needed to be changed, why we do it and how. In the next chapter, we're
+going to describe how the route is exported from table to protocols and how this
+process is changing. Stay tuned!*
diff --git a/doc/threads/02_asynchronous_export.md b/doc/threads/02_asynchronous_export.md
new file mode 100644
index 00000000..5a10ce5a
--- /dev/null
+++ b/doc/threads/02_asynchronous_export.md
@@ -0,0 +1,437 @@
+# BIRD Journey to Threads. Chapter 2: Asynchronous route export
+
+Route export is a core algorithm of BIRD. This chapter covers how we are making
+this procedure multithreaded. Desired outcomes are mostly lower latency of
+route import, flap dampening and also faster route processing in large
+configurations with lots of export from one table.
+
+BIRD is a fast, robust and memory-efficient routing daemon designed and
+implemented at the end of 20th century. We're doing a significant amount of
+BIRD's internal structure changes to make it possible to run in multiple
+threads in parallel.
+
+## How routes are propagated through BIRD
+
+In the [previous chapter](https://en.blog.nic.cz/2021/03/23/bird-journey-to-threads-chapter-1-the-route-and-its-attributes/), you could learn how the route import works. We should
+now extend that process by the route export.
+
+1. (In protocol code.) Create the route itself and propagate it through the
+ right channel by calling `rte_update`.
+2. The channel runs its import filter.
+3. New best route is selected.
+4. For each channel:
+ 1. The channel runs its preexport hook and export filter.
+ 2. (Optionally.) The channel merges the nexthops to create an ECMP route.
+ 3. The channel calls the protocol's `rt_notify` hook.
+5. After all exports are finished, the `rte_update` call finally returns and
+ the source protocol may do anything else.
+
+Let's imagine that all the protocols are running in parallel. There are two
+protocols with a route prepared to import. One of those wins the table lock,
+does the import and then the export touches the other protocol which must
+either:
+
+* store the route export until it finishes its own imports, or
+* have independent import and export parts.
+
+Both of these conditions are infeasible for common use. Implementing them would
+make protocols much more complicated with lots of new code to test and release
+at once and also quite a lot of corner cases. Risk of deadlocks is also worth
+mentioning.
+
+## Asynchronous route export
+
+We decided to make it easier for protocols and decouple the import and export
+this way:
+
+1. The import is done.
+2. Best route is selected.
+3. Resulting changes are stored.
+
+Then, after the importing protocol returns, the exports are processed for each
+exporting channel. In future, this will be possible in parallel: Some protocols
+may process the export directly after it is stored, other protocols will wait
+until they finish another job.
+
+This eliminates the risk of deadlocks and all protocols' `rt_notify` hooks can
+rely on their independence. There is only one question. How to store the changes?
+
+## Route export modes
+
+To find a good data structure for route export storage, we shall first know the
+readers. The exporters may request different modes of route export.
+
+### Export everything
+
+This is the most simple route export mode. The exporter wants to know about all
+the routes as they're changing. We therefore simply store the old route until
+the change is fully exported and then we free the old stored route.
+
+To manage this, we can simply queue the changes one after another and postpone
+old route cleanup after all channels have exported the change. The queue member
+would look like this:
+
+```
+struct {
+ struct rte_storage *new;
+ struct rte_storage *old;
+};
+```
+
+### Export best
+
+This is another simple route export mode. We check whether the best route has
+changed; if not, no export happens. Otherwise, the export is propagated as the
+old best route changing to the new best route.
+
+To manage this, we could use the queue from the previous point by adding new
+best and old best pointers. It is guaranteed that both the old best and new
+best pointers are always valid in time of export as all the changes in them
+must be stored in future changes which have not been exported yet by this
+channel and therefore not freed yet.
+
+```
+struct {
+ struct rte_storage *new;
+ struct rte_storage *new_best;
+ struct rte_storage *old;
+ struct rte_storage *old_best;
+};
+```
+
+Anyway, we're getting to the complicated export modes where this simple
+structure is simply not enough.
+
+### Export merged
+
+Here we're getting to some kind of problems. The exporting channel requests not
+only the best route but also all routes that are good enough to be considered
+ECMP-eligible (we call these routes *mergable*). The export is then just one
+route with just the nexthops merged. Export filters are executed before
+merging and if the best route is rejected, nothing is exported at all.
+
+To achieve this, we have to re-evaluate export filters any time the best route
+or any mergable route changes. Until now, the export could just do what it wanted
+as there was only one thread working. To change this, we need to access the
+whole route list and process it.
+
+### Export first accepted
+
+In this mode, the channel runs export filters on a sorted list of routes, best first.
+If the best route gets rejected, it asks for the next one until it finds an
+acceptable route or exhausts the list. This export mode requires a sorted table.
+BIRD users will know this export mode as `secondary` in BGP.
+
+For now, BIRD stores two bits per route for each channel. The *export bit* is set
+if the route has been really exported to that channel. The *reject bit* is set
+if the route was rejected by the export filter.
+
+When processing a route change for accepted, the algorithm first checks the
+export bit for the old route. If this bit is set, the old route is that one
+exported so we have to find the right one to export. Therefore the sorted route
+list is walked best to worst to find a new route to export, using the reject
+bit to evaluate only routes which weren't rejected in previous runs of this
+algorithm.
+
+If the old route bit is not set, the algorithm walks the sorted route list best
+to worst, checking the position of new route with respect to the exported route.
+If the new route is worse, nothing happens, otherwise the new route is sent to
+filters and finally exported if passes.
+
+### Export by feed
+
+To resolve problems arising from previous two export modes (merged and first accepted),
+we introduce a way to process a whole route list without locking the table
+while export filters are running. To achieve this, we follow this algorithm:
+
+1. The exporting channel sees a pending export.
+2. *The table is locked.*
+3. The channel counts all the routes for the given destination.
+4. The channel stores pointers to that routes to a local array.
+5. *The table is unlocked.*
+6. The channel processes the local array of route pointers.
+7. The pending export is marked as processed by this channel.
+
+After unlocking the table, the pointed-to routes are implicitly guarded by the
+sole fact that the pending export has not yet been processed by all channels
+and the cleanup routine never frees any resource related to a pending export.
+
+## Pending export data structure
+
+As the two complicated export modes use the export-by-feed algorithm, the
+pending export data structure may be quite minimalistic.
+
+```
+struct rt_pending_export {
+ struct rt_pending_export * _Atomic next; /* Next export for the same destination */
+ struct rte_storage *new; /* New route */
+ struct rte_storage *new_best; /* New best route in unsorted table */
+ struct rte_storage *old; /* Old route */
+ struct rte_storage *old_best; /* Old best route in unsorted table */
+ _Atomic u64 seq; /* Sequential ID (table-local) of the pending export */
+};
+```
+
+To allow for squashing outdated pending exports (e.g. for flap dampening
+purposes), there is a `next` pointer to the next export for the same
+destination. This is also needed for the export-by-feed algorithm: if there are
+several exports for one net at once, all of them are processed by one
+export-by-feed automatically marked as done.
+
+We should also add several items into `struct channel`.
+
+```
+ struct coroutine *export_coro; /* Exporter and feeder coroutine */
+ struct bsem *export_sem; /* Exporter and feeder semaphore */
+ struct rt_pending_export * _Atomic last_export; /* Last export processed */
+ struct bmap export_seen_map; /* Keeps track which exports were already processed */
+ u64 flush_seq; /* Table export seq when the channel announced flushing */
+```
+
+To run the exports in parallel, `export_coro` is run and `export_sem` is
+used for signalling new exports to it. The exporter coroutine also marks all
+seen sequential IDs in its `export_seen_map` to make it possible to skip over
+them if seen again.
+
+There is also a table cleaner routine
+(see [previous chapter](https://en.blog.nic.cz/2021/03/23/bird-journey-to-threads-chapter-1-the-route-and-its-attributes/))
+which must cleanup also the pending exports after all the channels are finished with them.
+To signal that, there is `last_export` working as a release point: the channel
+guarantees that it doesn't touch the pointed-to pending export, nor any data
+from it.
+
+The last tricky point is channel flushing. When any channel stops, all its
+routes are automatically freed and withdrawals are exported if appropriate.
+Until now, the routes could be flushed synchronously, anyway now flush has
+several phases:
+
+1. Flush started (here the channel stores the `seq` of last current pending export).
+2. All routes unlinked yet not freed, withdrawals pending.
+3. Pending withdrawals processed and cleaned up. Channel may safely stop and free its structures.
+
+Finally, some additional information has to be stored in tables:
+
+```
+ _Atomic byte export_used; /* Export journal cleanup scheduled */ \
+ struct rt_pending_export * _Atomic first_export; /* First export to announce */ \
+ byte export_scheduled; /* Export is scheduled */
+ list pending_exports; /* List of packed struct rt_pending_export */
+ struct fib export_fib; /* Auxiliary fib for storing pending exports */
+ u64 next_export_seq; /* The next export will have this ID */
+```
+
+The exports are:
+1. Assigned the `next_export_seq` sequential ID, incrementing this item by one.
+2. Put into `pending_exports` and `export_fib` for both sequential and by-destination access.
+3. Signalled by setting `export_scheduled` and also `first_export` if not `NULL`.
+
+After processing several exports, `export_used` is set and route table maintenance
+coroutine is woken up to possibly do cleanup.
+
+The `struct rt_pending_export` seems to be best allocated by requesting a whole
+memory page, containing a common list node, a simple header and packed all the
+structures in the rest of the page. This may save a significant amount of memory.
+In case of congestion, there will be lots of exports and every spare kilobyte
+counts. If BIRD is almost idle, the optimization does nothing on the overall performance.
+
+## Export algorithm
+
+As we have explained at the beginning, the current export algorithm is
+table-driven. The table walks the channel list and propagates the update.
+The now export algorithm is channel-driven. The table just indicates that it
+has something new in export queue and the channel decides what to do with that and when.
+
+### Pushing an export
+
+When a table has something to export, it enqueues an appropriate instance of
+`struct rt_pending_export`. Then it pings its maintenance coroutine
+(`rt_event`) to notify the exporting channels about a new route. This coroutine
+runs with only the table locked so the protocol may e.g. prepare the next route inbetween.
+The maintenance coroutine, when it wakes up, walks the list of channels and
+wakes their export coroutines.
+
+These two levels of asynchronicity are here for two reasons.
+
+1. There may be lots of channels (hundreds of them) and we don't want to
+ inefficiently iterate the list for each export.
+2. The notification is going to wait until the route author finishes, hopefully
+ importing more routes and therefore allowing more exports to be processed at once.
+
+The table also stores the first and last exports for every destination. These
+come handy mostly for cleanup purposes and for channel feeding.
+
+### Processing an export
+
+After these two pings, the channel finally knows that there is an export pending.
+The channel checks whether there is a `last_export` stored. If yes, it proceeds
+with the next one, otherwise it takes `first_export` from the table (which is
+atomic for this reason).
+
+1. The channel checks its `export_seen_map` whether this export has been
+ already processed. If so, it skips to the next export. No action is needed.
+2. The export chain is scanned for the current first and last export. This is
+ done by following the `next` pointer in the exports as accessing the
+ auxiliary export fib needs locking.
+3. In the best-only and all-routes export mode, the changes are processed
+ without further table locking.
+4. If export-by-feed is used, the current state of routes in table are fetched.
+ Then the table is unlocked and the changes are processed without further locking.
+5. All processed exports are marked as seen.
+6. The channel stores the `last_export` and returns to beginning.to wait for next export.
+
+## The full route life-cycle
+
+Until now, we're always assuming that the channels *just exist*. In real life,
+any channel may go up or down and we must handle it, flushing the routes
+appropriately and freeing all the memory just in time to avoid both
+use-after-free and memory leaks. It should be noted here explicitly that BIRD
+is written in C which has no garbage collector or other modern features alike.
+
+### Protocols and channels as viewed from a route
+
+BIRD consists effectively of protocols and tables. **Protocols** are active parts,
+kind-of subprocesses manipulating routes and other data. **Tables** are passive,
+serving as a database of routes. To connect a protocol to a table, a
+**channel** is created.
+
+Every route has its `sender` storing the channel which has put the route into
+the current table. This comes handy when the channel goes down to know which
+routes to flush.
+
+Every route also has its `src`, a route source allocated by the protocol which
+originated it first. This is kept when a route is passed through a *pipe*.
+
+Both `src` and `sender` must point to active protocols and channels as inactive
+protocols and channels may be deleted in any time.
+
+### Protocol and channel lifecycle
+
+In the beginning, all channels and protocols are down. Until they fully start,
+no route from them is allowed to any table. When the protocol and channel is up,
+they may originate routes freely. However, the transitions are worth mentioning.
+
+### Channel startup and feed
+
+When protocols and channels start, they need to get the current state of the
+appropriate table. Therefore, after a protocol and channel start, also the
+export-feed coroutine is initiated.
+
+Tables can contain millions of routes. It may lead to long import latency if a channel
+was feeding itself in one step. The table structure is (at least for now) too
+complicated to be implemented as lockless, thus even read access needs locking.
+To mitigate this, the feeds are split to allow for regular route propagation
+with a reasonable latency.
+
+When the exports were synchronous, we simply didn't care and just announced the
+exports to the channels from the time they started feeding. When making exports
+asynchronous, it was crucial to avoid most of the possible race conditions
+which could arise from simultaneous feed and export. As the feeder routines had
+to be rewritten, it was a good opportunity to make this precise.
+
+Therefore, when a channel goes up, it also starts exports:
+
+1. Start the feed-export coroutine.
+2. *Lock the table.*
+3. Store the last export in queue.
+4. Read a limited number of routes to local memory together with their pending exports.
+5. If there are some routes to process:
+ 1. *Unlock the table.*
+ 2. Process the loaded routes.
+ 3. Set the appropriate pending exports as seen.
+ 4. *Lock the table*
+ 5. Go to 4. to continue feeding.
+6. If there was a last export stored, load the next one to be processed. Otherwise take the table's `first_export`.
+7. *Unlock the table.*
+8. Run the exporter loop.
+
+*Note: There are some nuances not mentioned here how to do things in right
+order to avoid missing some events while changing state. Precise description
+would be too detailed for this text.*
+
+When the feeder loop finishes, it continues smoothly to process all the exports
+that have been queued while the feed was running. Step 5.3 ensures that already
+seen exports are skipped, steps 3 and 6 ensure that no export is missed.
+
+### Channel flush
+
+Protocols and channels need to stop for a handful of reasons, All of these
+cases follow the same routine.
+
+1. (Maybe.) The protocol requests to go down or restart.
+2. The channel requests to go down or restart.
+3. The channel requests to stop export.
+4. In the feed-export coroutine:
+ 1. At a designated cancellation point, check cancellation.
+ 2. Clean up local data.
+ 3. *Lock main BIRD context locked*
+ 4. If shutdown requested, switch the channel to *flushing* state and request table maintenance.
+5. In the table maintenance coroutine:
+ 1. Walk across all channels and check them for *flushing* state.
+ 2. Walk across the table (split to allow for low latency updates) and
+ generate a withdrawal for each route sent by the flushing channels.
+ 3. Wait until all the withdrawals are processed.
+ 4. Mark the flushing channels as *down* and eventually proceed to the protocol shutdown or restart.
+
+There is also a separate routine that handles bulk cleanup of `src`'s which
+contain a pointer to the originating protocol. This routine may get reworked in
+future, yet for now it is good enough.
+
+### Route export cleanup
+
+Last but not least is the export cleanup routine. Until now, the withdrawn
+routes were exported synchronously and freed directly after the import was
+done. This is not possible anymore. The export is stored and the import returns
+to let the importing protocol continue its work. We therefore need a routine to
+cleanup the withdrawn routes and also the processed exports.
+
+First of all, this routine refuses to cleanup when any export is feeding or
+shutting down. This may change in future, anyway for now we aren't sure about
+possible race conditions.
+
+Anyway, when all the exports are in a steady state, the routine works as follows:
+
+1. Walk the active exports and find a minimum (oldest export) between their `last_export` values.
+2. If there is nothing to clear between the actual oldest export and channels' oldest export, do nothing.
+3. Find the table's new `first_export` and set it. Now there is nobody pointing to the old exports.
+4. Free the withdrawn routes.
+5. Free the old exports, removing them also from the first-last list of exports for the same destination.
+
+## Results of these changes
+
+This step is a first major step to move forward. Using just this version may be
+still as slow as the single-threaded version, at least if your export filters are trivial.
+Anyway, the main purpose of this step is not an immediate speedup. It is more
+of a base for the next steps:
+
+* Unlocking of pipes should enable parallel execution of all the filters on
+ pipes, limited solely by the principle *one thread for every direction of
+ pipe*.
+* Conversion of CLI's `show route` to the new feed-export coroutines should
+ enable faster table queries. Moreover, this approach will allow for
+ better splitting of model and view in CLI with a good opportunity to
+ implement more output formats, e.g. JSON.
+* Unlocking of kernel route synchronization should fix latency issues induced
+ by long-lasting kernel queries.
+* Partial unlocking of BGP packet processing should finally allow for parallel
+ execution in almost all phases of BGP route propagation.
+
+The development is now being done mostly in the branch `alderney`. If you asked
+why such strange branch names like `jersey`, `guernsey` and `alderney`, here is
+a kind-of reason. Yes, these names could be named `mq-async-export`,
+`mq-async-export-new`, `mq-async-export-new-new`, `mq-another-async-export` and
+so on. That's so ugly, isn't it?
+
+Also why so many branches? The development process is quite messy. BIRD's code
+heavily depends on single-threaded approach. This is exceptionally good for
+performance, as long as you have one thread only. On the other hand, lots of
+these assumptions are not documented so in many cases one desired
+change yields a chain of other unforeseen changes which must precede. This brings
+lots of backtracking, branch rebasing and other Git magic. There is always a
+can of worms somewhere in the code.
+
+*It's still a long road to the version 2.1. This series of texts should document
+what is needed to be changed, why we do it and how. The
+[previous chapter](https://en.blog.nic.cz/2021/03/23/bird-journey-to-threads-chapter-1-the-route-and-its-attributes/)
+showed the necessary changes in route storage. In the next chapter, we're going
+to describe how the coroutines are implemented and what kind of locking system
+are we employing to prevent deadlocks. Stay tuned!*
diff --git a/doc/threads/Makefile b/doc/threads/Makefile
new file mode 100644
index 00000000..f5198be7
--- /dev/null
+++ b/doc/threads/Makefile
@@ -0,0 +1,15 @@
+SUFFICES := .pdf -wordpress.html
+CHAPTERS := 00_the_name_of_the_game 01_the_route_and_its_attributes 02_asynchronous_export
+
+all: $(foreach ch,$(CHAPTERS),$(addprefix $(ch),$(SUFFICES)))
+
+00_the_name_of_the_game.pdf: 00_filter_structure.png
+
+%.pdf: %.md
+ pandoc -f markdown -t latex -o $@ $<
+
+%.html: %.md
+ pandoc -f markdown -t html5 -o $@ $<
+
+%-wordpress.html: %.html Makefile
+ sed -r 's#</p>#\n#g; s#<p>##g; s#<(/?)code>#<\1tt>#g; s#<pre><tt>#<code>#g; s#</tt></pre>#</code>#g; s#</?figure>##g; s#<figcaption>#<p style="text-align: center">#; s#</figcaption>#</p>#; ' $< > $@