- apply both pg_upmap and pg_upmap_items
- allow bidirectional swap of pg-upmap-items
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.com>
There is now a fallback to a choose_arg index of -1 if there isn't
a pool-specific choose_arg set. If you create a per-pool weight-set,
that works for that pool. Otherwise we try the compat/default one. If
that doesn't exist either, then we use the normal CRUSH weights.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.com>
... otherwise we die in insert_pg_mapping(), which wants pg->node to be
empty, i.e. initialized with RB_CLEAR_NODE.
Fixes: 6f428df47d ("libceph: pg_upmap[_items] infrastructure")
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
No sooner than Dan had fixed this issue in commit 293dffaad8
("libceph: NULL deref on crush_decode() error path"), I brought it
back. Add a new label and set -EINVAL once, right before failing.
Fixes: 278b1d709c ("libceph: ceph_decode_skip_* helpers")
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
There are hidden gotos in the ceph_decode_* macros. We need to set the
"err" variable on these error paths otherwise we end up returning
ERR_PTR(0) which is NULL. It causes NULL dereferences in the callers.
Fixes: 6f428df47d ("libceph: pg_upmap[_items] infrastructure")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
[idryomov@gmail.com: similar bug in osdmap_decode(), changelog tweak]
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
If there is no crush_choose_arg_map for a given pool, a NULL pointer is
passed to preserve existing crush_do_rule() behavior.
Reflects ceph.git commits 55fb91d64071552ea1bc65ab4ea84d3c8b73ab4b,
dbe36e08be00c6519a8c89718dd47b0219c20516.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
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: Ilya Dryomov <idryomov@gmail.com>
Previously, pg_to_raw_osds() didn't filter for existent OSDs because
raw_to_up_osds() would filter for "up" ("up" is predicated on "exists")
and raw_to_up_osds() was called directly after pg_to_raw_osds(). Now,
with apply_upmap() call in there, nonexistent OSDs in pg_to_raw_osds()
output can affect apply_upmap(). Introduce remove_nonexistent_osds()
to deal with that.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Move raw_pg_to_pg() call out of get_temp_osds() and into
ceph_pg_to_up_acting_osds(), for upcoming apply_upmap().
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
pg_temp and pg_upmap encodings are the same (PG -> array of osds),
except for the incremental remove: it's an empty mapping in new_pg_temp
for pg_temp and a separate old_pg_upmap set for pg_upmap. (This isn't
to allow for empty pg_upmap mappings -- apparently, pg_temp just wasn't
looked at as an example for pg_upmap encoding.)
Reuse __decode_pg_temp() for decoding pg_upmap and new_pg_upmap.
__decode_pg_temp() stores into pg_temp union member, but since pg_upmap
union member is identical, reading through pg_upmap later is OK.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Some of these won't be as efficient as they could be (e.g.
ceph_decode_skip_set(... 32 ...) could advance by len * sizeof(u32)
once instead of advancing by sizeof(u32) len times), but that's fine
and not worth a bunch of extra macro code.
Replace skip_name_map() with ceph_decode_skip_map as an example.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Make __{lookup,remove}_pg_mapping() look like their ceph_spg_mapping
counterparts: take const struct ceph_pg *.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
The old (v15) pi->last_force_request_resend has been repurposed to
make pre-RESEND_ON_SPLIT clients that don't check for PG splits but do
obey pi->last_force_request_resend resend on splits. See ceph.git
commit 189ca7ec6420 ("mon/OSDMonitor: make pre-luminous clients resend
ops on split").
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
If there is not enough space then ceph_decode_32_safe() does a goto bad.
We need to return an error code in that situation. The current code
returns ERR_PTR(0) which is NULL. The callers are not expecting that
and it results in a NULL dereference.
Fixes: f24e9980eb ("ceph: OSD client")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Ilya Dryomov <idryomov@gmail.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Since ceph.git commit 4e28f9e63644 ("osd/OSDMap: clear osd_info,
osd_xinfo on osd deletion"), weight is set to IN when OSD is deleted.
This changes the result of applying an incremental for clients, not
just OSDs. Because CRUSH computations are obviously affected,
pre-4e28f9e63644 servers disagree with post-4e28f9e63644 clients on
object placement, resulting in misdirected requests.
Mirrors ceph.git commit a6009d1039a55e2c77f431662b3d6cc5a8e8e63f.
Fixes: 930c532869 ("libceph: apply new_state before new_up_client on incrementals")
Link: http://tracker.ceph.com/issues/19122
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.com>
Older (shorter) CRUSH maps too need to be finalized.
Fixes: 66a0e2d579 ("crush: remove mutable part of CRUSH map")
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
With EC overwrites maturing, the kernel client will be getting exposed
to potentially very wide EC pools. While "min(pi->size, X)" works fine
when the cluster is stable and happy, truncating OSD sets interferes
with resend logic (ceph_is_new_interval(), etc). Abort the mapping if
the pool is too wide, assigning the request to the homeless session.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.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: Ilya 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: Ilya Dryomov <idryomov@gmail.com>
Add pool namesapce pointer to struct ceph_file_layout and struct
ceph_object_locator. Pool namespace is used by when mapping object
to PG, it's also used when composing OSD request.
The namespace pointer in struct ceph_file_layout is RCU protected.
So libceph can read namespace without taking lock.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
[idryomov@gmail.com: ceph_oloc_destroy(), misc minor changes]
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Define new ceph_file_layout structure and rename old ceph_file_layout
to ceph_file_layout_legacy. This is preparation for adding namespace
to ceph_file_layout structure.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
Currently, osd_weight and osd_state fields are updated in the encoding
order. This is wrong, because an incremental map may look like e.g.
new_up_client: { osd=6, addr=... } # set osd_state and addr
new_state: { osd=6, xorstate=EXISTS } # clear osd_state
Suppose osd6's current osd_state is EXISTS (i.e. osd6 is down). After
applying new_up_client, osd_state is changed to EXISTS | UP. Carrying
on with the new_state update, we flip EXISTS and leave osd6 in a weird
"!EXISTS but UP" state. A non-existent OSD is considered down by the
mapping code
2087 for (i = 0; i < pg->pg_temp.len; i++) {
2088 if (ceph_osd_is_down(osdmap, pg->pg_temp.osds[i])) {
2089 if (ceph_can_shift_osds(pi))
2090 continue;
2091
2092 temp->osds[temp->size++] = CRUSH_ITEM_NONE;
and so requests get directed to the second OSD in the set instead of
the first, resulting in OSD-side errors like:
[WRN] : client.4239 192.168.122.21:0/2444980242 misdirected client.4239.1:2827 pg 2.5df899f2 to osd.4 not [1,4,6] in e680/680
and hung rbds on the client:
[ 493.566367] rbd: rbd0: write 400000 at 11cc00000 (0)
[ 493.566805] rbd: rbd0: result -6 xferred 400000
[ 493.567011] blk_update_request: I/O error, dev rbd0, sector 9330688
The fix is to decouple application from the decoding and:
- apply new_weight first
- apply new_state before new_up_client
- twiddle osd_state flags if marking in
- clear out some of the state if osd is destroyed
Fixes: http://tracker.ceph.com/issues/14901
Cc: stable@vger.kernel.org # 3.15+: 6dd74e44dc: libceph: set 'exists' flag for newly up osd
Cc: stable@vger.kernel.org # 3.15+
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Josh Durgin <jdurgin@redhat.com>
Commit d30291b985 ("libceph: variable-sized ceph_object_id") changed
dout()s in what is now encode_request() and ceph_object_locator_to_pg()
to use %pE, mostly to document that, although all rbd and cephfs object
names are NULL-terminated strings, ceph_object_id will handle any RADOS
object name, including the one containing NULs, just fine.
However, it turns out that vbin_printf() can't handle anything but ints
and %s - all %p suffixes are ignored. The buffer %p** points to isn't
recorded, resulting in trash in the messages if the buffer had been
reused by the time bstr_printf() got to it.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
This leads to a simpler osdmap handling code, particularly when dealing
with pi->was_full, which is introduced in a later commit.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Introduce ceph_osd_request_target, containing all mapping-related
fields of ceph_osd_request and calc_target() for calculating mappings
and populating it.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Add and decode pi->min_size and pi->last_force_request_resend. These
are going to be used by calc_target().
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
calc_target() code is going to need to know how to compare PGs. Take
lhs and rhs pgid by const * while at it.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Rename ceph_calc_pg_primary() to ceph_pg_to_acting_primary() to
emphasise that it returns acting primary.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Knowning just acting set isn't enough, we need to be able to record up
set as well to detect interval changes. This means returning (up[],
up_len, up_primary, acting[], acting_len, acting_primary) and passing
it around. Introduce and switch to ceph_osds to help with that.
Rename ceph_calc_pg_acting() to ceph_pg_to_up_acting_osds() and return
both up and acting sets from it.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Rename ceph_oloc_oid_to_pg() to ceph_object_locator_to_pg(). Emphasise
that returned is raw PG and return -ENOENT instead of -EIO if the pool
doesn't exist.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Currently ceph_object_id can hold object names of up to 100
(CEPH_MAX_OID_NAME_LEN) characters. This is enough for all use cases,
expect one - long rbd image names:
- a format 1 header is named "<imgname>.rbd"
- an object that points to a format 2 header is named "rbd_id.<imgname>"
We operate on these potentially long-named objects during rbd map, and,
for format 1 images, during header refresh. (A format 2 header name is
a small system-generated string.)
Lift this 100 character limit by making ceph_object_id be able to point
to an externally-allocated string. Apart from being able to work with
almost arbitrarily-long named objects, this allows us to reduce the
size of ceph_object_id from >100 bytes to 64 bytes.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
struct crush_bucket_tree::num_nodes is u8, so ceph_decode_8_safe()
should be used. -Wconversion catches this, but I guess it went
unnoticed in all the noise it spews. The actual problem (at least for
common crushmaps) isn't the u32 -> u8 truncation though - it's the
advancement by 4 bytes instead of 1 in the crushmap buffer.
Fixes: http://tracker.ceph.com/issues/2759
Cc: stable@vger.kernel.org
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Josh Durgin <jdurgin@redhat.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: Ilya Dryomov <idryomov@gmail.com>
Use the more common pr_warn.
Other miscellanea:
o Coalesce formats
o Realign arguments
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
If the state variable is krealloced successfully, map->osd_state will be
freed, once following two reallocation failed, and exit the function
without resetting map->osd_state, map->osd_state become a wild pointer.
fix it by resetting them after krealloc successfully.
Signed-off-by: Li RongQing <roy.qing.li@gmail.com>
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Commit e2b149cc4b ("crush: add chooseleaf_vary_r tunable") added the
crush_map::chooseleaf_vary_r field but missed the decode part. This
lead to misdirected requests caused by incorrect raw crush mapping
sets.
Fixes: http://tracker.ceph.com/issues/8226
Reported-and-Tested-by: Dmitry Smirnov <onlyjob@member.fsf.org>
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
osd_primary_affinity array is indexed into incorrectly when checking
for non-default primary-affinity values. This nullifies the impact of
the rest of the apply_primary_affinity() and results in misdirected
requests.
if (osds[i] != CRUSH_ITEM_NONE &&
osdmap->osd_primary_affinity[i] !=
^^^
CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) {
For a pool with size 2, this always ends up checking osd0 and osd1
primary_affinity values, instead of the values that correspond to the
osds in question. E.g., given a [2,3] up set and a [max,max,0,max]
primary affinity vector, requests are still sent to osd2, because both
osd0 and osd1 happen to have max primary_affinity values and therefore
we return from apply_primary_affinity() early on the premise that all
osds in the given set have max (default) values. Fix it.
Fixes: http://tracker.ceph.com/issues/7954
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
