bucket_straw2_choose needs to use weights that may be different from
weight_items. For instance to compensate for an uneven distribution
caused by a low number of values. Or to fix the probability biais
introduced by conditional probabilities (see
http://tracker.ceph.com/issues/15653 for more information).

We introduce a weight_set for each straw2 bucket to set the desired
weight for a given item at a given position. The weight of a given item
when picking the first replica (first position) may be different from
the weight the second replica (second position). For instance the weight
matrix for a given bucket containing items 3, 7 and 13 could be as
follows:

          position 0   position 1

item 3     0x10000      0x100000
item 7     0x40000       0x10000
item 13    0x40000       0x10000

When crush_do_rule picks the first of two replicas (position 0), item 7,
3 are four times more likely to be choosen by bucket_straw2_choose than
item 13. When choosing the second replica (position 1), item 3 is ten
times more likely to be choosen than item 7, 13.

By default the weight_set of each bucket exactly matches the content of
item_weights for each position to ensure backward compatibility.

bucket_straw2_choose compares items by using their id. The same ids are
also used to index buckets and they must be unique. For each item in a
bucket an array of ids can be provided for placement purposes and they
are used instead of the ids. If no replacement ids are provided, the
legacy behavior is preserved.

Reflects ceph.git commit 19537a450fd5c5a0bb8b7830947507a76db2ceca.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

The syntax error was not noticed because dprintk is a macro
and the code is discarded by default.

Reflects ceph.git commit f29b840c64a933b2cb13e3da6f3d785effd73a57.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

The is_out() test may require an additional hashing operation, so we
should skip it whenever possible.

Reflects ceph.git commit db107cc7f15cf2481894add325dc93e33479f529.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Much like Arlo Guthrie, I decided that one big pile is better than two
little piles.

Reflects ceph.git commit 95c2df6c7e0b22d2ea9d91db500cf8b9441c73ba.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Then add it to the working state. It would be very nice if we didn't
have to take a lock to calculate a crush placement. By moving the
permutation array into the working data, we can treat the CRUSH map as
immutable.

Reflects ceph.git commit cbcd039651c0569551cb90d26ce27e1432671f2a.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Include linux/crush/mapper.h in crush/mapper.c to get the prototypes of
crush_find_rule and crush_do_rule which are defined there. This fixes
the following GCC warnings when building with 'W=1':

  net/ceph/crush/mapper.c:40:5: warning: no previous prototype for ‘crush_find_rule’ [-Wmissing-prototypes]
  net/ceph/crush/mapper.c:793:5: warning: no previous prototype for ‘crush_do_rule’ [-Wmissing-prototypes]
Signed-off-by: NTobias Klauser <tklauser@distanz.ch>
[idryomov@gmail.com: corresponding !__KERNEL__ include]
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Remove extra x1 variable, it's just temporary placeholder that
clutters the code unnecessarily.

Reflects ceph.git commit 0d19408d91dd747340d70287b4ef9efd89e95c6b.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Use __builtin_clz() supported by GCC and Clang to figure out
how many bits we should shift instead of shifting by a bit
in a loop until the value gets normalized. Improves performance
of this function by up to 3x in worst-case scenario and overall
straw2 performance by ~10%.

Reflects ceph.git commit 110de33ca497d94fc4737e5154d3fe781fa84a0a.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Add a tunable to fix the bug that chooseleaf may cause unnecessary pg
migrations when some device fails.

Reflects ceph.git commit fdb3f664448e80d984470f32f04e2e6f03ab52ec.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>
Reviewed-by: NSage Weil <sage@redhat.com>

Ensure that the take argument is a valid bucket ID before indexing the
buckets array.

Reflects ceph.git commit 93ec538e8a667699876b72459b8ad78966d89c61.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>
Reviewed-by: NSage Weil <sage@redhat.com>

We were indexing the buckets array without verifying the index was
within the [0,max_buckets) range.  This could happen because
a multistep rule does not have enough buckets and has CRUSH_ITEM_NONE
for an intermediate result, which would feed in CRUSH_ITEM_NONE and
make us crash.

Reflects ceph.git commit 976a24a326da8931e689ee22fce35feab5b67b76.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>
Reviewed-by: NSage Weil <sage@redhat.com>

.. up to ceph.git commit 1db1abc8328d ("crush: eliminate ad hoc diff
between kernel and userspace").  This fixes a bunch of recently pulled
coding style issues and makes includes a bit cleaner.

A patch "crush:Make the function crush_ln static" from Nicholas Krause
<xerofoify@gmail.com> is folded in as crush_ln() has been made static
in userspace as well.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Verify that the 'take' argument is a valid device or bucket.
Otherwise ignore it (do not add the value to the working vector).

Reflects ceph.git commit 9324d0a1af61e1c234cc48e2175b4e6320fff8f4.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

This is an improved straw bucket that correctly avoids any data movement
between items A and B when neither A nor B's weights are changed.  Said
differently, if we adjust the weight of item C (including adding it anew
or removing it completely), we will only see inputs move to or from C,
never between other items in the bucket.

Notably, there is not intermediate scaling factor that needs to be
calculated.  The mapping function is a simple function of the item weights.

The below commits were squashed together into this one (mostly to avoid
adding and then yanking a ~6000 lines worth of crush_ln_table):

- crush: add a straw2 bucket type
- crush: add crush_ln to calculate nature log efficently
- crush: improve straw2 adjustment slightly
- crush: change crush_ln to provide 32 more digits
- crush: fix crush_get_bucket_item_weight and bucket destroy for straw2
- crush/mapper: fix divide-by-0 in straw2
  (with div64_s64() for draw = ln / w and INT64_MIN -> S64_MIN - need
   to create a proper compat.h in ceph.git)

Reflects ceph.git commits 242293c908e923d474910f2b8203fa3b41eb5a53,
                          32a1ead92efcd351822d22a5fc37d159c65c1338,
                          6289912418c4a3597a11778bcf29ed5415117ad9,
                          35fcb04e2945717cf5cfe150b9fa89cb3d2303a1,
                          6445d9ee7290938de1e4ee9563912a6ab6d8ee5f,
                          b5921d55d16796e12d66ad2c4add7305f9ce2353.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Crush temporary buffers are allocated as per replica size configured
by the user.  When there are more final osds (to be selected as per
rule) than the replicas, buffer overlaps and it causes crash.  Now, it
ensures that at most num-rep osds are selected even if more number of
osds are allowed by the rule.

Reflects ceph.git commits 6b4d1aa99718e3b367496326c1e64551330fabc0,
                          234b066ba04976783d15ff2abc3e81b6cc06fb10.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

This lets you adjust the vary_r tunable on a per-rule basis.

Reflects ceph.git commit f944ccc20aee60a7d8da7e405ec75ad1cd449fac.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NJosh Durgin <josh.durgin@inktank.com>

The current crush_choose_firstn code will re-use the same 'r' value for
the recursive call.  That means that if we are hitting a collision or
rejection for some reason (say, an OSD that is marked out) and need to
retry, we will keep making the same (bad) choice in that recursive
selection.

Introduce a tunable that fixes that behavior by incorporating the parent
'r' value into the recursive starting point, so that a different path
will be taken in subsequent placement attempts.

Note that this was done from the get-go for the new crush_choose_indep
algorithm.

This was exposed by a user who was seeing PGs stuck in active+remapped
after reweight-by-utilization because the up set mapped to a single OSD.

Reflects ceph.git commit a8e6c9fbf88bad056dd05d3eb790e98a5e43451a.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NJosh Durgin <josh.durgin@inktank.com>

These two fields are misnomers; they are *retry* counts.

Reflects ceph.git commit f17caba8ae0cad7b6f8f35e53e5f73b444696835.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NJosh Durgin <josh.durgin@inktank.com>

Back in 27f4d1f6bc32c2ed7b2c5080cbd58b14df622607 we refactored the CRUSH
code to allow adjustment of the retry counts on a per-pool basis.  That
commit had an off-by-one bug: the previous "tries" counter was a *retry*
count, not a *try* count, but the new code was passing in 1 meaning
there should be no retries.

Fix the ftotal vs tries comparison to use < instead of <= to fix the
problem.  Note that the original code used <= here, which means the
global "choose_total_tries" tunable is actually counting retries.
Compensate for that by adding 1 in crush_do_rule when we pull the tunable
into the local variable.

This was noticed looking at output from a user provided osdmap.
Unfortunately the map doesn't illustrate the change in mapping behavior
and I haven't managed to construct one yet that does.  Inspection of the
crush debug output now aligns with prior versions, though.

Reflects ceph.git commit 795704fd615f0b008dcc81aa088a859b2d075138.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NJosh Durgin <josh.durgin@inktank.com>

Reflects ceph.git commit 8b38f10bc2ee3643a33ea5f9545ad5c00e4ac5b4.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Reflects ceph.git commit ea3a0bb8b773360d73b8b77fa32115ef091c9857.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

This allows all of the tunables to be overridden by a specific rule.

Reflects ceph.git commits d129e09e57fbc61cfd4f492e3ee77d0750c9d292,
                          0497db49e5973b50df26251ed0e3f4ac7578e66e.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

The legacy behavior is to make the normal number of tries for the
recursive chooseleaf call.  The descend_once tunable changed this to
making a single try and bail if we get a reject (note that it is
impossible to collide in the recursive case).

The new set_chooseleaf_tries lets you select the number of recursive
chooseleaf attempts for indep mode, or default to 1.  Use the same
behavior for firstn, except default to total_tries when the legacy
tunables are set (for compatibility).  This makes the rule step
override the (new) default of 1 recursive attempt, keeping behavior
consistent with indep mode.

Reflects ceph.git commit 685c6950ef3df325ef04ce7c986e36ca2514c5f1.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

This aligns the internal identifier names with the user-visible names in
the decompiled crush map language.

Reflects ceph.git commit caa0e22e15e4226c3671318ba1f61314bf6da2a6.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Since we can specify the recursive retries in a rule, we may as well also
specify the non-recursive tries too for completeness.

Reflects ceph.git commit d1b97462cffccc871914859eaee562f2786abfd1.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Parameterize the attempts for the _firstn choose method, and apply the
rule-specified tries count to firstn mode as well.  Note that we have
slightly different behavior here than with indep:

 If the firstn value is not specified for firstn, we pass through the
 normal attempt count.  This maintains compatibility with legacy behavior.
 Note that this is usually *not* actually N^2 work, though, because of the
 descend_once tunable.  However, descend_once is unfortunately *not* the
 same thing as 1 chooseleaf try because it is only checked on a reject but
 not on a collision.  Sigh.

 In contrast, for indep, if tries is not specified we default to 1
 recursive attempt, because that is simply more sane, and we have the
 option to do so.  The descend_once tunable has no effect for indep.

Reflects ceph.git commit 64aeded50d80942d66a5ec7b604ff2fcbf5d7b63.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Explicitly control the number of sample attempts, and allow the number of
tries in the recursive call to be explicitly controlled via the rule. This
is important because the amount of time we want to spend looking for a
solution may be rule dependent (e.g., higher for the wide indep pool than
the rep pools).

(We should do the same for the other tunables, by the way!)

Reflects ceph.git commit c43c893be872f709c787bc57f46c0e97876ff681.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Pass down the parent's 'r' value so that we will sample different values in
the recursive call when the parent tries multiple times.  This avoids doing
useless work (calling multiple times and trying the same values).

Reflects ceph.git commit 2731d3030d7a3e80922b7f1b7756f9a4a124bac5.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Pass numrep (the width of the result) separately from the number of results
we want *this* iteration.  This makes things less awkward when we do a
recursive call (for chooseleaf) and want only one item.

Reflects ceph.git commit 1b567ee08972f268c11b43fc881e57b5984dd08b.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Now that indep is handled by crush_choose_indep, rename crush_choose to
crush_choose_firstn and remove all the conditionals.  This ends up
stripping out *lots* of code.

Note that it *also* makes it obvious that the shenanigans we were playing
with r' for uniform buckets were broken for firstn mode.  This appears to
have happened waaaay back in commit dae8bec9 (or earlier)... 2007.

Reflects ceph.git commit 94350996cb2035850bcbece6a77a9b0394177ec9.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Reflects ceph.git commit 4551fee9ad89d0427ed865d766d0d44004d3e3e1.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Reflects ceph.git commit 86e978036a4ecbac4c875e7c00f6c5bbe37282d3.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

For firstn mode, if we fail to make a valid placement choice, we just
continue and return a short result to the caller.  For indep mode, however,
we need to make the position stable, and return an undefined value on
failed placements to avoid shifting later results to the left.

Reflects ceph.git commit b1d4dd4eb044875874a1d01c01c7d766db5d0a80.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

This is only present to size the temporary scratch arrays that we put on
the stack.  Let the caller allocate them as they wish and remove the
limitation.

Reflects ceph.git commit 1cfe140bf2dab99517589a82a916f4c75b9492d1.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Reflects ceph.git commit 3cef755428761f2481b1dd0e0fbd0464ac483fc5.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Reflects ceph.git commit e7d47827f0333c96ad43d257607fb92ed4176550.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

Pass the size of the weight vector into crush_do_rule() to ensure that we
don't access values past the end.  This can happen if the caller misbehaves
and passes a weight vector that is smaller than max_devices.

Currently the monitor tries to prevent that from happening, but this will
gracefully tolerate previous bad osdmaps that got into this state.  It's
also a bit more defensive.

Reflects ceph.git commit 5922e2c2b8335b5e46c9504349c3a55b7434c01a.
Signed-off-by: NIlya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: NSage Weil <sage@inktank.com>

This saves us some cycles, but does not affect the placement result at
all.

This corresponds to ceph.git commit 4abb53d4f.
Signed-off-by: NSage Weil <sage@inktank.com>

Add libceph support for a new CRUSH tunable recently added to Ceph servers.

Consider the CRUSH rule
  step chooseleaf firstn 0 type <node_type>

This rule means that <n> replicas will be chosen in a manner such that
each chosen leaf's branch will contain a unique instance of <node_type>.

When an object is re-replicated after a leaf failure, if the CRUSH map uses
a chooseleaf rule the remapped replica ends up under the <node_type> bucket
that held the failed leaf.  This causes uneven data distribution across the
storage cluster, to the point that when all the leaves but one fail under a
particular <node_type> bucket, that remaining leaf holds all the data from
its failed peers.

This behavior also limits the number of peers that can participate in the
re-replication of the data held by the failed leaf, which increases the
time required to re-replicate after a failure.

For a chooseleaf CRUSH rule, the tree descent has two steps: call them the
inner and outer descents.

If the tree descent down to <node_type> is the outer descent, and the descent
from <node_type> down to a leaf is the inner descent, the issue is that a
down leaf is detected on the inner descent, so only the inner descent is
retried.

In order to disperse re-replicated data as widely as possible across a
storage cluster after a failure, we want to retry the outer descent. So,
fix up crush_choose() to allow the inner descent to return immediately on
choosing a failed leaf.  Wire this up as a new CRUSH tunable.

Note that after this change, for a chooseleaf rule, if the primary OSD
in a placement group has failed, choosing a replacement may result in
one of the other OSDs in the PG colliding with the new primary.  This
requires that OSD's data for that PG to need moving as well.  This
seems unavoidable but should be relatively rare.

This corresponds to ceph.git commit 88f218181a9e6d2292e2697fc93797d0f6d6e5dc.
Signed-off-by: NJim Schutt <jaschut@sandia.gov>
Reviewed-by: NSage Weil <sage@inktank.com>