Adds an API to remove unused entries in the outhash database based on
age and if they are referenced by any unihash
(Bitbake rev: a169ac523d166c6cbba918b152a76782176c3e88)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Extends the get_outhash API with a flag indicating whether to include
the unihash in the output. This is means that the query doesn't require
the unihash entry to be present to return a result
(Bitbake rev: b8d6abfeb4a0765727a62b3d8d83276335c7c7d6)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Adds a `remove` API to the client and server that can be used to remove
hash equivalence entries that match a particular critera
(Bitbake rev: 861d068b3a9fb5e91a01dbec54996a5a6f93ef29)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
We have a choice of policy with hashequivalence - whether to reduce
sstate duplication in the sstate feed to a minimum or have maximal
sstate reuse from the user's perspective.
The challenge is that non-matching outhashes are generated due to
determinism issues, or due to differences in host gcc version,
architecture and so on and the question is how to reconcile then.
The approach before this patch is that any new match is added and
matches can update. This has the side effect that a queried value
from the server can change due to the replacement and you may not
always get the same value from the server. With the client side
caching bitbake has, this can be suboptimal and when using the
autobuilder sstate feed, it results in poor artefact reuse.
This patch switches to the other possible behaviour, once a hash is
assigned, it doesn't change. This means some sstate artefacts may be
duplicated but dependency chains aren't invalidated which I suspect
may give better overall performance.
Update the tests to match the new behaviour.
(Bitbake rev: 20d6ac753efa364349100cdc863e5eabec8e5b78)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Fixes the hashequivalence server to resolve the diverging report race
error. This error occurs when the same task(hash) is run simultaneous on
two different builders, and then the results are reported back but the
hashes diverge (e.g. have different outhashes), and one outhash is
equivalent to a hash and another is not. If taskhash was not originally
in the database, the client will fallback to using the taskhash as the
suggested unihash and the server will see reports come in like:
taskhash: A
unihash: A
outhash: B
taskhash: C
unihash: C
outhash: B
taskhash: C
unihash: C
outhash: D
Note that the second and third reports are the same taskhash, with
diverging outhashes.
Taskhash C should be equivalent to taskhash (and unihash) A because they
share an outhash B, but the server would not do this when tasks were
reported in the order shown.
It became clear while trying to fix this that single large table to
store all reported hashes was going to make these updates difficult
since updating the unihash of all entries would be complex and time
consuming. Instead, it makes more sense to split apart the database into
two tables: One that maps taskhashes to unihashes and one that maps
outhashes to taskhashes. This should hopefully improve the parsing query
times as well since they only care about the taskhashes to unihashes
table, at the cost of more complex INNER JOIN queries on the lesser used
API.
Note this change does delete existing hash equivlance data and starts a
new database table rather than converting existing data.
(Bitbake rev: dff5a17558e2476064e85f35bad1fd65fec23600)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Prevents `ResourceWarning: unclosed event loop` warnings when using the
synchronous client and python exits
(Bitbake rev: 8b95972bc04ce52a98c7780184af15a5e95f987b)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Reworks the async I/O API so that the async loop is only created in the
child process. This requires deferring the creation of the server until
the child process and a queue to transfer the bound address back to the
parent process
(Bitbake rev: 8555869cde39f9e9a9ced5a3e5788209640f6d50)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
[small loop -> self.loop fix in serv.py]
Signed-off-by: Scott Murray <scott.murray@konsulko.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
If the SIGTERM signal is sent to an asyncrpc server before it has
installed the SIGTERM handler in the main loop, it may miss the signal
which will can cause the calling process to wait forever on the join().
To resolve this, the calling process should mask of SIGTERM before
forking the server process and the server should unmask the signal only
after the handler is installed. To simplify the usage of the server, an
new helper function called serve_as_process() is added to do this
automatically and correctly.
Thanks: Scott Murray <scott.murray@konsulko.com> for helping debug
(Bitbake rev: ef2865efa98ad20823267364f2159d8d8c931400)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The asyncrpc module can now be used to provide the json & asyncio based
RPC system used by hashserv.
(Bitbake rev: 5afb9586b0a4a23a05efb0e8ff4a97262631ae4a)
Signed-off-by: Paul Barker <pbarker@konsulko.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The Python built-in ConnectionError type can be used instead of a custom
HashConnectionError type. This will make code refactoring simpler.
(Bitbake rev: 8a796c3d6d99cfa8ef7aff0ae55bb0f23bbbeae1)
Signed-off-by: Paul Barker <pbarker@konsulko.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The get-outhash message can be sent via the get_outhash client method.
This works in a similar way to the get message but looks up a db entry
by outhash rather than by taskhash. It is intended to be used as a
read-only form of the report message.
As both handle_get_outhash and handle_report use the same query string
we can factor this out.
(Bitbake rev: dc19606ada29a4d8afde4fcecd8ec986b47b867e)
Signed-off-by: Paul Barker <pbarker@konsulko.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Use the new get-outhash message to perform a read-only query against an
upstream server (if present) when a reported taskhash/outhash
combination is not found in the current database. If a matching entry is
found upstream it is copied into the current database so it can be found
by future queries.
(Bitbake rev: 2be4f7f0d2ccb09917398289e8140e1467e84bb2)
Signed-off-by: Paul Barker <pbarker@konsulko.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The -r/--readonly argument is added to the bitbake-hashserv app. If this
argument is given then clients may only perform read operations against
the server. The read-only mode is implemented by simply not installing
handlers for write operations, this keeps the permission model simple
and reduces the risk of accidentally allowing write operations.
As a sqlite database can be safely opened by multiple processes in
parallel, it's possible to start two hashserv instances against a single
database if you wish to export both a read-only port and a read-write
port.
(Bitbake rev: 492bb02eb0e071c792407ac3113f92492da1a9cc)
Signed-off-by: Paul Barker <pbarker@konsulko.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
If the server returns an empty response ("null" in json), this cannot
be iterated to check for the presence of the "chunk-stream" key.
(Bitbake rev: bf75370bcd6d02ed08cd959eec6190196b792515)
Signed-off-by: Paul Barker <pbarker@konsulko.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Fixes the bug were long paths would break Unix domain socket clients
(for real this time; the previous attempt was missing os.path.basename).
Adds some tests to prevent regressions
(Bitbake rev: 77790e3656048eff5cb1a086c727d86d32773b68)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Restores a fix for unix domain socket path length limits when using the
synchronous hash equivalence client that was accidentally removed when
the async client was added.
Unfortunately, it's much more difficult to fix the same problem when
using the async client directly due to the interaction of chdir() and
async code, but this will at least restore the old behavior in the
synchronous case.
(Bitbake rev: 53e85022a8b1c8f407c9418260c59beffb96f0f9)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Adds support for an upstream server to be specified. The upstream server
will be queried for equivalent hashes whenever a miss is found in the
local server. If the server returns a match, it is merged into the
local database. In order to keep the get stream queries as fast as
possible since they are the critical path when bitbake is preparing the
run queue, missing tasks provided by the server are not immediately
pulled from the upstream server, but instead are put into a queue to be
backfilled by a worker task later.
(Bitbake rev: e6d6c0b39393e9bdf378c1eba141f815e26b724b)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Adds support for create a client that operates using Python asynchronous
I/O.
(Bitbake rev: cf9bc0310b0092bf52b61057405aeb51c86ba137)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Using localhost for direct builds on host is fine. A case with a
misbehavior has been sighted on a Docker build. Even when the host
supports IPv6, but Docker is not configured correspondingly - some
versions of the asyncio Python module seem to misbehave and try to
use IPv6 where it's not supported in the container. This happens at
least on some Ubuntu 18.04 based containers, resolving the IP
explicitly appears to be the fix.
(Bitbake rev: 0e20f91c11afdc17ea776aa02e0cc8b0d59a23d4)
Signed-off-by: Anatol Belski <anbelski@linux.microsoft.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Correctly import, and inherit functions, and variables.
Also fix some typos and remove some Python 2 code that isn't recognised.
(Bitbake rev: b0c807be5c2170c9481c1a04d4c11972135d7dc5)
Signed-off-by: Frazer Clews <frazerleslieclews@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The hash equivalence client and server can occasionally send messages
that are too large for the server to fit in the receive buffer (64 KB).
To prevent this, support is added to the protocol to "chunkify" the
stream and break it up into manageable pieces that the server can each
side can back together.
Ideally, this would be negotiated by the client and server, but it's
currently hard coded to 32 KB to prevent the round-trip delay.
(Bitbake rev: e27a28c1e40e886ee68ba4b99b537ffc9c3577d4)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
removed unused imports which made the code harder to read, and slightly
but less efficient
(Bitbake rev: 4367692a932ac135c5aa4f9f2a4e4f0150f76697)
Signed-off-by: Frazer Clews <frazer.clews@codethink.co.uk>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The reason for this should be recorded in the commit logs. Imagine
you have a target recipe (e.g. meta-extsdk-toolchain) which depends on
gdb-cross. sstate in OE-Core allows gdb-cross to have the same hash
regardless of whether its built on x86 or arm. The outhash will be
different.
We need hashequiv to be able to adapt to the prescence of sstate artefacts
for meta-extsdk-toolchain and allow the hashes to re-intersect, rather than
trying to force a rebuild of meta-extsdk-toolchain. By this point in the build,
it would have already been installed from sstate so the build needs to adapt.
Equivalent hashes should be reported to the server as a taskhash that
needs to map to an specific unihash. This patch adds API to the hashserv
client/server to allow this.
[Thanks to Joshua Watt for help with this patch]
(Bitbake rev: 674692fd46a7691a1de59ace6af0556cc5dd6a71)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The hash server process is terminated and waited on with join(), so it
should not be a daemon. Daemonizing it cause races with the server
cleanup, especially in the selftest because the process may not have
terminated and cleanup up its socket before the test cleanup runs and
tries to do it.
[YOCTO #13542]
(Bitbake rev: 7c829675581818f92d57056b57fbd3880829b6bd)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The os module is required to connect to a unix domain socket
(Bitbake rev: 31a5111bcd0080a583d0d95fad3e09ae78bdf0fa)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Reworks the hash equivalence server to address performance issues that
were encountered with the REST mechanism used previously, particularly
during the heavy request load encountered during signature generation.
Notable changes are:
1) The server protocol is no longer HTTP based. Instead, it uses a
simpler JSON over a streaming protocol link. This protocol has much
lower overhead than HTTP since it eliminates the HTTP headers.
2) The hash equivalence server can either bind to a TCP port, or a Unix
domain socket. Unix domain sockets are more efficient for local
communication, and so are preferred if the user enables hash
equivalence only for the local build. The arguments to the
'bitbake-hashserve' command have been updated accordingly.
3) The value to which BB_HASHSERVE should be set to enable a local hash
equivalence server is changed to "auto" instead of "localhost:0". The
latter didn't make sense when the local server was using a Unix
domain socket.
4) Clients are expected to keep a persistent connection to the server
instead of creating a new connection each time a request is made for
optimal performance.
5) Most of the client logic has been moved to the hashserve module in
bitbake. This makes it easier to share the client code.
6) A new bitbake command has been added called 'bitbake-hashclient'.
This command can be used to query a hash equivalence server, including
fetching the statistics and running a performance stress test.
7) The table indexes in the SQLite database have been updated to
optimize hash lookups. This change is backward compatible, as the
database will delete the old indexes first if they exist.
8) The server has been reworked to use python async to maximize
performance with persistently connected clients. This requires Python
3.5 or later.
(Bitbake rev: 2124eec3a5830afe8e07ffb6f2a0df6a417ac973)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
This can cause a huge backlog of closing sockets on the server and
in our case we don't really want/need the protection TCP is trying to
give us so work around it.
(Bitbake rev: 7bc79fdf60519231da7c0c7b5b6143ce090ed830)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
At exit the hashserv code was causing tracebacks as join() wasn't
being called from the thread that started the process. Ensure that
the hashserver is started from the pre_serve hook which is the
final thread the cooker runs in. This avoids the traceback at the
expense of some horrific poking into data stores which will ultimately
need improving through a proper API.
(Bitbake rev: 05888700e5f6cba48a26c8a4c447634a28e3baa6)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
There were hard to debug lockups when trying to use threading to start
hashserv as a thread. Switch to multiprocessing which doesn't show the
same locking problems.
(Bitbake rev: be23d887c8e244f1ef961298fbc9214d0fd0968a)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Experience with the prserv shows that having two threads, one accepting
and queueing connections and one handling the requests leads to much
more reliable behaviour than having everything in a single thread.
(Bitbake rev: a03d60671a53d9ff70e07cc42fe35f6f8776dac2)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
We're seeing performance problems with hashserv running on a normal build
system. The cause seems to be the large amounts of file IO that builds involve
blocking writes to the database. Since sqlite blocks on the sync calls, this
causes a significant problem.
Since if we lose power we have bigger problems, run with synchronous=off
to avoid locking and put the jounral into memory to avoid any write issues
there too.
This took writes from 120s down to negligible in my tests, which means
hashserv then responds promptly to requests.
(Bitbake rev: 7ae56a4d4fcf66e1da1581c70f75e30bfdf3ed83)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Its useful, particularly in the local developer model of usage, for
bitbake to start and stop a hash equivalence server on local port,
rather than relying on one being started by the user before the build.
The new BB_HASHSERVE variable supports this.
The database handling is moved internally into the hashserv code so that
different threads/processes can be used for the server without errors.
(Bitbake rev: a4fa8f1bd88995ae60e10430316fbed63d478587)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Implements a number of optimizations to the SQL used in the hash
equivalence server:
1) Two indexes are created for the two methods (method, taskhash and
method outhash) by which rows are found in order to speed up the
lookup
2) An extra SELECT to lookup the just inserted row was removed. This
SELECT is unnecessary since all of the information about the newly
inserted row is already available.
3) A uniqueness constraint was added to the table. This should allow
the server to be multithreaded in the future since duplicate inserts
can be detected (and ignored). This change requires bumping the
database version to '2', since a uniqueness constraint can't be
added to an existing table.
4) Some comments are added to clarify the trick SELECT statement used
when inserting new equivalent hashes
(Bitbake rev: 7aec8632e67b4f0ab7b72692c40a42f6926608c3)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
With the introduction of SPDX-License-Identifier headers, we don't need a ton
of header boilerplate in every file. Simplify the files and rely on the top
level for the full licence text.
(Bitbake rev: 695d84397b68cc003186e22f395caa378b06bc75)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
This adds the SPDX-License-Identifier license headers to the majority of
our source files to make it clearer exactly which license files are under.
The bulk of the files are under GPL v2.0 with one found to be under V2.0
or later, some under MIT and some have dual license. There are some files
which are potentially harder to classify where we've imported upstream code
and those can be handled specifically in later commits.
The COPYING file is replaced with LICENSE.X files which contain the full
license texts.
(Bitbake rev: ff237c33337f4da2ca06c3a2c49699bc26608a6b)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Implements a reference implementation of the hash equivalence server.
This server has minimal dependencies (and no dependencies outside of the
standard Python library), and implements the minimum required to be a
conforming hash equivalence server.
[YOCTO #13030]
(Bitbake rev: 1bb2ad0b44b94ee04870bf3f7dac4e663bed6e4d)
Signed-off-by: Joshua Watt <JPEWhacker@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
