osd: implement basic caching policies in ReplicatedPG
Right now these are very basic and aren't as sophisticated as we
want them to end up, but we have a skeleton for where to put the
decision-making logic.
If we get back a redirect reply, we clean up the Op's external references
and re-send using the target_oloc and target_oid. To facilitate this,
recalc_op_target() now only fills them in and overrides them with pool
cache semantics if they're empty.
Keep in mind that this is a pretty simple redirect formula -- the
Objecter will keep following redirects forever if that's what the OSDs
send back. The client is not providing any synchronization right now.
Objecter: write a helper function to clean up ops that need to be retried
We have a little block to clean them up if we get back EAGAIN, but it's
actually leaking map references; we will also use this for redirects
from the OSDs.
When present, clients must send the request to the location specified
by the redirect (by using the combine_with_locator() function on
request_redirect_t).
A separate mechanism must be used to ensure that clients see and respect
the redirect, as we do not bump up the minimum required version to
decode.
Analagous to the oloc->base_oloc rename we did in e2fcad09d94d965867147627b73e99da9454436f, we may specify a different
target name for a redirect. Rename the existing oid to base_oid to
avoid any confusion.
With the introduction of the erasure code pool, arguments to be
interpreted depending on the pool type must be introduced.
For instance the erasure code pool loads a plugin at run time will
use easure-code-k=10 to split each object in 10.
If key=value it is stored in the new properties data member of pg_pool_t
as properties[key] = value, otherwise the value is the empty string.
The pg_pool_t version is bumped to 10 and the encode/decode methods
modified to take the properties into account. The
generate_test_instances method creates a two entries map, one of which
is the empty string to cover the case when no value is specified.
Greg Farnum [Fri, 30 Aug 2013 23:33:31 +0000 (16:33 -0700)]
ReplicatedPG: do not meaninglessly fill in the reqid on make_writeable() cloning
This reqid is used to fill in a map that is used for giving clients
the correct version on replayed ops, unless the reqid is blank (in
which case it doesn't go into the map). Indirect clones are not ops
that the client wants to observe and the version is incorrect right now,
so don't fill it in.
Note that this should not have actually caused any buggy behavior, because
the map would have simply been overwritten by the real requested event
a short time later (while still protected by locks and things). But it's
very confusing.
Loic Dachary [Wed, 28 Aug 2013 15:29:18 +0000 (17:29 +0200)]
ErasureCodePlugin: plugin registry tests and example
libec_example.la is a fully functional plugin based on
ErasureCodeExample to test the ErasureCodePlugin abstract
interface. It is dynamically loaded to test the
ErasureCodePluginRegistry implementation.
Although the plugin is built in the test directory, it will be
installed. noinst_LTLIBRARIES won't build the shared library, only the
static version which is not suitable for testing.
Loic Dachary [Wed, 28 Aug 2013 13:57:54 +0000 (15:57 +0200)]
ErasureCodePlugin: plugin registry
A ErasureCodePluginRegistry singleton holds all erasure plugin objects
derived from ErasureCodePlugin and dlopen(2) handles for the lifetime
of the OSD and is cleaned up by the destructor.
The registry has a single entry point ( method factory ) and should
be used as follows:
If the plugin requested ( "jerasure" in the example above ) is not
found in the *plugins* data member, the load method is called and will:
* dlopen(parameters["erasure-code-directory"] + "jerasure")
* f = dlsym("__erasure_code_init")
* f("jerasure")
* check that it registered "jerasure"
The plugin is expected to do something like
instance.add(plugin_name, new ErasureCodePluginJerasure());
to register itself.
The factory method is protected with a Mutex to avoid race
conditions when using the same plugin from two threads.
The erasure_codelib_LTLIBRARIES variable is added to the Makefile
and the plugins are expected to add themselves and be installed
in the $(libdir)/erasure-code
Loic Dachary [Wed, 28 Aug 2013 13:46:34 +0000 (15:46 +0200)]
ErasureCodePlugin: plugin interface
When dynamically loaded, a plugin is expected to define
int __erasure_code_init(char *plugin_name);
When called, it is responsible for registering an ErasureCodePlugin
derived object that provides a factory method from which the concrete
implementation of the ErasureCodeInterface object can be generated:
Loic Dachary [Mon, 19 Aug 2013 17:15:07 +0000 (19:15 +0200)]
ErasureCode: example implementation : K=2 M=1
An erasure code implementation designed for tests. Although it is fully
functional and could be used on actual data, it is mainly provided for
testing purposes. It splits data in two, computes an XOR parity and
can sustain the loss of one chunk.
The constructor will usleep(3) for parameters["usleep"] microseconds
so that the caller can create race conditions.
Loic Dachary [Mon, 19 Aug 2013 16:56:56 +0000 (18:56 +0200)]
ErasureCode: abstract interface
The erasure coded pool relies on this abstract interface to encode and
decode the chunks stored in the OSD. It has been designed to be
generic enough to accomodate the libraries and algorithms that are
most likely to be used. It does not claim to be universal.
- In "includes", inttypes.h was cluttering the system's one. This caused
random build errors on some systems/in some conditions. Renaming it.
- Add emergency defs of PRI*64 headers when int_types.h does not define
them (which, unfortunately, can happen on some systems).
Signed-off-by: Roald J. van Loon <roaldvanloon@gmail.com>
Sage Weil [Thu, 5 Sep 2013 04:29:11 +0000 (21:29 -0700)]
common/crc32c_intel_fast: avoid reading partial trailing word
The optimized intel code reads in word-sized chunks, knowing that the
allocator will only hand out memory in word-sized increments. This makes
valgrind unhappy. Whitelisting doesn't work because for some reason there
is no caller context (probably because of some interaction with yasm?).
Instead, just use the baseline code for the last few bytes. This should
not be significant.
Signed-off-by: Sage Weil <sage@inktank.com> Reviewed-by: Greg Farnum <greg@inktank.com>
cleanup: passing context to NamedState for ceph_clock
This makes the constructor call on the subclasses explicit, and passes
the cct to the NamedState constructor. This cct is used by ceph_clock
to set enter_time.
Removes the last reference to g_ceph_context from libosd.
Signed-off-by: Roald J. van Loon <roaldvanloon@gmail.com>
Roald van Loon [Wed, 28 Aug 2013 14:56:23 +0000 (16:56 +0200)]
cleanup: removing globals from common/obj_bencher
This file is in common/ but cant be included in libcommon.la because of
this reference. Removing it, making the binary calling it to pass the
correct cephcontext (rados, rest-bench).
Signed-off-by: Roald J. van Loon <roaldvanloon@gmail.com>
Roald van Loon [Wed, 28 Aug 2013 10:11:08 +0000 (12:11 +0200)]
cleanup: removed last references to globals from client
- There were some refs in SyntheticClient to g_(ceph_context|conf), I
replaced them with client->cct and client->cct->_conf.
- There were some refs in fuse_ll to g_conf, also replaced them with
client->cct or cfuse->client->cct where applicable.
This makes everything in src/client completely independent from globals.
Signed-off-by: Roald J. van Loon <roaldvanloon@gmail.com>
Yehuda Sadeh [Fri, 23 Aug 2013 22:39:20 +0000 (15:39 -0700)]
rgw: flush pending data when completing multipart part upload
Fixes: #6111
Backport: dumpling
When completing the part upload we need to flush any data that we
aggregated and didn't flush yet. With earlier code didn't have to deal
with it as for multipart upload we didn't have any pending data.
What we do now is we call the regular atomic data completion
function that takes care of it.
When posting an object it is possible to provide a key
name that refers to the original filename, however we
need to verify that in the end we don't end up with an
empty object name.
Yehuda Sadeh [Thu, 22 Aug 2013 00:22:46 +0000 (17:22 -0700)]
rgw: OPTIONS request doesn't need to read object info
This is a bucket-only operation, so we shouldn't look at the
object. Object may not exist and we might respond with Not
Exists response which is not what we want.
projects / ceph.git / log