Global leader lock was first designated to only acquired once
until the node is killed. However, currently, the code acquires
it repeatedly during the lifetime of the server, now there is a
goroutine leak.
remote error is not required to be passed back to the
client - this is mostly because we have healing that should
eventually, catch up on this and heal the bucket.
Ensure delete marker replication success, especially since the
recent optimizations to heal on HEAD, LIST and GET can force
replication attempts on delete marker before underlying object
version could have synced.
Move to using `xl.meta` data structure to keep temporary partInfo,
this allows for a future change where we move to different parts to
different drives.
PUT shall only proceed if pre-conditions are met, the new
code uses
- x-minio-source-mtime
- x-minio-source-etag
to verify if the object indeed needs to be replicated
or not, allowing us to avoid StatObject() call.
When limiting listing do not count delete, since they may be discarded.
Extend limit, since we may be discarding the forward-to marker.
Fix directories always being sent to resolve, since they didn't return as match.
On occasion this test fails:
```
2022-09-12T17:22:44.6562737Z === RUN TestGetObjectWithOutdatedDisks
2022-09-12T17:22:44.6563751Z erasure-object_test.go:1214: Test 2: Expected data to have md5sum = `c946b71bb69c07daf25470742c967e7c`, found `7d16d23f07072af1a809707ba101ae07`
2
```
Theory: Both objects are written with the same timestamp due to lower timer resolution on Windows. This results in secondary resolution, which is deterministic, but random.
Solution: Instead of hacking in a wait we request the specific version we want. Should still keep the test relevant.
Bonus: Remote action dependency for vulncheck
If replication config could not be read from bucket metadata for some
reason, issue a panic so that unexpected replication outcomes can
be avoided for replicated buckets.
For similar reasons, adding a panic while fetching object-lock config
if it failed for reason other than non-existence of config.
to avoid relying on scanner-calculated replication metrics.
This will improve the accuracy of the replication stats reported.
This PR also adds on to #15556 by handing replication
traffic that could not be queued by available workers to the
MRF queue so that entries in `PENDING` status are healed faster.
500k is a reasonable limit for any single MinIO
cluster deployment, in future we may increase this
value.
However for now we are going to keep this limit.
When healing is parallelized by setting the ` _MINIO_HEAL_WORKERS`
environment variable, multiple goroutines may race while updating the disk's
healing tracker. This change serializes only these concurrent updates using a
channel. Note, the healing tracker is still not concurrency safe in other contexts.
This PR is a continuation of the previous change instead
of returning an error, instead trigger a spot heal on the
'xl.meta' and return only after the healing is complete.
This allows for future GETs on the same resource to be
consistent for any version of the object.
xl.meta gets written and never rolled back, however
we definitely need to validate the state that is
persisted on the disk, if there are inconsistencies
- more than write quorum we should return an error
to the client
- if write quorum was achieved however there are
inconsistent xl.meta's we should simply trigger
an MRF on them
The `clusterInfo` struct in admin-handlers is same as
madmin.ClusterRegistrationInfo, except for small differences in field
names.
Removing this and using madmin.ClusterRegistrationInfo in its place will
help in following ways:
- The JSON payload generated by mc in case of cluster registration will
be consistent (same keys) with cluster.info generated by minio as part
of the profile and inspect zip
- health-analyzer can parse the cluster.info using the same struct and
won't have to define it's own
Currently, there is a short time window where the code is allowed
to save the status of a replication resync. Currently, the window is
`now.Sub(st.EndTime) <= resyncTimeInterval`. Also, any failure to
write in the backend disks is not retried.
Refactor the code a little bit to rely on the last timestamp of a
successful write of the resync status of any given bucket in the
backend disks.
When replication is enabled in a particular bucket, the listing will send
objects to bucket replication, but it is also sending prefixes for non
recursive listing which is useless and shows a lot of error logs.
This commit will ignore prefixes.
under some sequence of events following code would
reach an infinite loop.
```
idx1, idx2 := 0, 1
for ; idx2 != idx1; idx2++ {
fmt.Println(idx2)
}
```
fixes#15639
A lot of warning messages are printed in CI/CD failures generated by go
test. Avoid that by requiring at least Error level for logging when
doing go test.
inlined data often is bigger than the allowed
O_DIRECT alignment, so potentially we can write
'xl.meta' without O_DSYNC instead we can rely on
O_DIRECT + fdatasync() instead.
This PR allows O_DIRECT on inlined data that
would gain the benefits of performing O_DIRECT,
eventually performing an fdatasync() at the end.
Performance boost can be observed here for small
objects < 128KiB. The performance boost is mainly
seen on HDD, and marginal on NVMe setups.
When a node finds a change in the other replication cluster and applies
to itself will already notify other peers. No need for all nodes in a
given cluster to do site replication healing, only one node is
sufficient.
This PR improves the replication failure healing by persisting
most recent failures to disk and re-queuing them until the replication
is successful.
While this does not eliminate the need for healing during a full scan,
queuing MRF vastly improves the ETA to keeping replicated buckets
in sync as it does not wait for the scanner visit to detect unreplicated
object versions.
competing calls on the same object on versioned bucket
mutating calls on the same object may unexpected have
higher delays.
This can be reproduced with a replicated bucket
overwriting the same object writes, deletes repeatedly.
For longer locks like scanner keep the 1sec interval
This PR fixes possible leaks that may emanate from not
listening on context cancelation or timeouts.
```
goroutine 60957610 [chan send, 16 minutes]:
github.com/minio/minio/cmd.(*erasureServerPools).Walk.func1.1.1(...)
github.com/minio/minio/cmd/erasure-server-pool.go:1724 +0x368
github.com/minio/minio/cmd.listPathRaw({0x4a9a740, 0xc0666dffc0},...
github.com/minio/minio/cmd/metacache-set.go:1022 +0xfc4
github.com/minio/minio/cmd.(*erasureServerPools).Walk.func1.1()
github.com/minio/minio/cmd/erasure-server-pool.go:1764 +0x528
created by github.com/minio/minio/cmd.(*erasureServerPools).Walk.func1
github.com/minio/minio/cmd/erasure-server-pool.go:1697 +0x1b7
```
The bottom line is delete markers are a nuisance,
most applications are not version aware and this
has simply complicated the version management.
AWS S3 gave an unnecessary complication overhead
for customers, they need to now manage these
markers by applying ILM settings and clean
them up on a regular basis.
To make matters worse all these delete markers
get replicated as well in a replicated setup,
requiring two ILM settings on each site.
This PR is an attempt to address this inferior
implementation by deviating MinIO towards an
idempotent delete marker implementation i.e
MinIO will never create any more than single
consecutive delete markers.
This significantly reduces operational overhead
by making versioning more useful for real data.
This is an S3 spec deviation for pragmatic reasons.
Queue failed/pending replication for healing during listing and GET/HEAD
API calls. This includes healing of existing objects that were never
replicated or those in the middle of a resync operation.
This PR also fixes a bug in ListObjectVersions where lifecycle filtering
should be done.
when object speedtest is running keep writing
previous speedtest result back to client until
we have a new result - this avoids sending back
blank entries in between the speedtest when it
is running in 'autotune' mode.
```
commit 7bdaf9bc50
Author: Aditya Manthramurthy <donatello@users.noreply.github.com>
Date: Wed Jul 24 17:34:23 2019 -0700
Update on-disk storage format for users system (#7949)
```
Bonus: fixes a bug when etcd keys were being re-encrypted.
Currently, the code doesn't check if the user creating a bucket with
locking feature has bucket locking and versioning permissions enabled,
adding it in accordance with S3 spec.
https://docs.aws.amazon.com/AmazonS3/latest/API/API_CreateBucket.html
Object Lock - If ObjectLockEnabledForBucket is set to true in your CreateBucket request,
s3:PutBucketObjectLockConfiguration and s3:PutBucketVersioning permissions are required.
Capture average, p50, p99, p999 response times
and ttfb values. These are needed for latency
measurements and overall understanding of our
speedtest results.
listConfigItems creates a goroutine but sometimes callers will
exit without properly asking listAllIAMConfigItems() to stop sending
results, hence a goroutine leak.
Create a new context and cancel it for each listAllIAMConfigItems
call.
This commit adds support for automatically reloading
the MinIO client certificate for authentication to KES.
The client certificate will now be reloaded:
- when the private key / certificate file changes
- when a SIGHUP signal is received
- every 15 minutes
Fixes#14869
Signed-off-by: Andreas Auernhammer <hi@aead.dev>
The path is marked dirty automatically when healObject() is called, which is
wrong. HealObject() is called during self-healing and this will lead to
an increase in the false positive result of the bloom filter.
Also move NSUpdated() from renameData() and call it directly in
CompleteMultipart and PutObject, this is not a functional change but
it will make it less prone to errors in the future.
smaller setups may have less drives per server choosing
the concurrency based on number of local drives, and let
the MinIO server change the overall concurrency as
necessary.
It is possible for anyone with admin access to relatively
to get any content of any random OS location by simply
providing the file with 'mc admin update alias/ /etc/passwd`.
Workaround is to disable 'admin:ServiceUpdate' action. Everyone
is advised to upgrade to this patch.
Thanks to @alevsk for finding this bug.
this has been observed in multiple environments
where the setups are small `speedtest` naturally
fails with default '10s' and the concurrency
of '32' is big for such clusters.
choose a smaller value i.e equal to number of
drives in such clusters and let 'autotune'
increase the concurrency instead.
fixes#15334
- re-use net/url parsed value for http.Request{}
- remove gosimple, structcheck and unusued due to https://github.com/golangci/golangci-lint/issues/2649
- unwrapErrs upto leafErr to ensure that we store exactly the correct errors
"consoleAdmin" was used as the policy for root derived accounts, but this
lead to unexpected bugs when an administrator modified the consoleAdmin
policy
This change avoids evaluating a policy for root derived accounts as by
default no policy is mapped to the root user. If a session policy is
attached to a root derived account, it will be evaluated as expected.
This PR changes the handling of bucket deletes for site
replicated setups to hold on to deleted bucket state until
it syncs to all the clusters participating in site replication.
Currently, if one server in a distributed setup fails to upgrade
due to any reasons, it is not possible to upgrade again unless
nodes are restarted.
To fix this, split the upgrade process into two steps :
- download the new binary on all servers
- If successful, overwrite the old binary with the new one
This commit replaces `ioutil.TempDir` with `t.TempDir` in tests. The
directory created by `t.TempDir` is automatically removed when the test
and all its subtests complete.
Prior to this commit, temporary directory created using `ioutil.TempDir`
needs to be removed manually by calling `os.RemoveAll`, which is omitted
in some tests. The error handling boilerplate e.g.
defer func() {
if err := os.RemoveAll(dir); err != nil {
t.Fatal(err)
}
}
is also tedious, but `t.TempDir` handles this for us nicely.
Reference: https://pkg.go.dev/testing#T.TempDir
Signed-off-by: Eng Zer Jun <engzerjun@gmail.com>
Add cluster info to inspect and profiling archive.
In addition to the existing data generation for both inspect and profiling,
cluster.info file is added. This latter contains some info of the cluster.
The generation of cluster.info is is done as the last step and it can fail
if it exceed 10 seconds.
a/b/c/d/ where `a/b/c/` exists results in additional syscalls
such as an Lstat() call to verify if the `a/b/c/` exists
and its a directory.
We do not need to do this on MinIO since the parent prefixes
if exist, we can simply return success without spending
additional syscalls.
Also this implementation attempts to simply use Access() calls
to avoid os.Stat() calls since the latter does memory allocation
for things we do not need to use.
Access() is simpler since we have a predictable structure on
the backend and we know exactly how our path structures are.
A huge number of goroutines would build up from various monitors
When creating test filesystems provide a context so they can shut down when no longer needed.
Do completely independent multipart uploads.
In distributed mode, a lock was held to merge each multipart
upload as it was added. This lock was highly contested and
retries are expensive (timewise) in distributed mode.
Instead, each part adds its metadata information uniquely.
This eliminates the per object lock required for each to merge.
The metadata is read back and merged by "CompleteMultipartUpload"
without locks when constructing final object.
Co-authored-by: Harshavardhana <harsha@minio.io>
This commit adds a `context.Context` to the
the KMS `{Stat, CreateKey, GenerateKey}` API
calls.
The context will be used to terminate external calls
as soon as the client requests gets canceled.
A follow-up PR will add a `context.Context` to
the remaining `DecryptKey` API call.
Signed-off-by: Andreas Auernhammer <hi@aead.dev>
Uploading a part object can leave an inconsistent state inside
.minio.sys/multipart where data are uploaded but xl.meta is not
committed yet.
Do not list upload-ids that have this state in the multipart listing.
Add up to 256 bytes of padding for compressed+encrypted files.
This will obscure the obvious cases of extremely compressible content
and leave a similar output size for a very wide variety of inputs.
This does *not* mean the compression ratio doesn't leak information
about the content, but the outcome space is much smaller,
so often *less* information is leaked.
Make bucket requests sent after decommissioning is started are not
created in a suspended pool. Therefore listing buckets should avoid
suspended pools as well.
Rename Trigger -> Event to be a more appropriate
name for the audit event.
Bonus: fixes a bug in AddMRFWorker() it did not
cancel the waitgroup, leading to waitgroup leaks.
There is no point in compressing very small files.
Typically the effective size on disk will be the same due to disk blocks.
So don't waste resources on extremely small files.
We don't check on multipart. 1) because we don't know and 2) this is very likely a big object anyway.
This commit adds a minimal set of KMS-related metrics:
```
# HELP minio_cluster_kms_online Reports whether the KMS is online (1) or offline (0)
# TYPE minio_cluster_kms_online gauge
minio_cluster_kms_online{server="127.0.0.1:9000"} 1
# HELP minio_cluster_kms_request_error Number of KMS requests that failed with a well-defined error
# TYPE minio_cluster_kms_request_error counter
minio_cluster_kms_request_error{server="127.0.0.1:9000"} 16790
# HELP minio_cluster_kms_request_success Number of KMS requests that succeeded
# TYPE minio_cluster_kms_request_success counter
minio_cluster_kms_request_success{server="127.0.0.1:9000"} 348031
```
Currently, we report whether the KMS is available and how many requests
succeeded/failed. However, KES exposes much more metrics that can be
exposed if necessary. See: https://pkg.go.dev/github.com/minio/kes#Metric
Signed-off-by: Andreas Auernhammer <hi@aead.dev>
If more than 1M folders (objects or prefixes) are found at the top level in a bucket allow it to be compacted.
While very suboptimal structure we should limit memory usage at some point.
GetDiskInfo() uses timedValue to cache the disk info for one second.
timedValue behavior was recently changed to return an old cached value
when calculating a new value returns an error.
When a mount point is empty, GetDiskInfo() will return errUnformattedDisk,
timedValue will return cached disk info with unexpected IsRootDisk value,
e.g. false if the mount point belongs to a root disk. Therefore, the mount
point will be considered a valid disk and will be formatted as well.
This commit will also add more defensive code when marking root disks:
always mark a disk offline for any GetDiskInfo() error except
errUnformattedDisk. The server will try anyway to reconnect to those
disks every 10 seconds.
it is not safe to pass around sync.Map
through pointers, as it may be concurrently
updated by different callers.
this PR simplifies by avoiding sync.Map
altogether, we do not need sync.Map
to keep object->erasureMap association.
This PR fixes a crash when concurrently
using this value when audit logs are
configured.
```
fatal error: concurrent map iteration and map write
goroutine 247651580 [running]:
runtime.throw({0x277a6c1?, 0xc002381400?})
runtime/panic.go:992 +0x71 fp=0xc004d29b20 sp=0xc004d29af0 pc=0x438671
runtime.mapiternext(0xc0d6e87f18?)
runtime/map.go:871 +0x4eb fp=0xc004d29b90 sp=0xc004d29b20 pc=0x41002b
```
The current code uses approximation using a ratio. The approximation
can skew if we have multiple pools with different disk capacities.
Replace the algorithm with a simpler one which counts data
disks and ignore parity disks.
fix: allow certain mutation on objects during decommission
currently by mistake deletion of objects was skipped,
if the object resided on the pool being decommissioned.
delete's are okay to be allowed since decommission is
designed to run on a cluster with active I/O.
Small uploads spend a significant amount of time (~5%) fetching disk info metrics. Also maps are allocated for each call.
Add a 100ms cache to disk metrics.
versioned buckets were not creating the delete markers
present in the versioned stack of an object, this essentially
would stop decommission to succeed.
This PR fixes creating such delete markers properly during
a decommissioning process, adds tests as well.
Current code incorrectly passed the
config asset object name while decommissioning,
make sure that we pass the right object name
to be hashed on the newer set of pools.
This PR fixes situations after a successful
decommission, the users and policies might go
missing due to wrong hashed set.
also use designated names for internal
calls
- storageREST calls are storageR
- lockREST calls are lockR
- peerREST calls are just peer
Named in this fashion to facilitate wildcard matches
by having prefixes of the same name.
Additionally, also enable funcNames for generic handlers
that return errors, currently we disable '<unknown>'
In a replicated setup, when an object is updated in one cluster but
still waiting to be replicated to the other cluster, GET requests with
if-match, and range headers will likely fail. It is better to proxy
requests instead.
Also, this commit avoids printing verbose logs about precondition &
range errors.
fix: change timedvalue to return previous cached value
caller can interpret the underlying error and decide
accordingly, places where we do not interpret the
errors upon timedValue.Get() - we should simply use
the previously cached value instead of returning "empty".
Bonus: remove some unused code
Add a generic handler that adds a new tracing context to the request if
tracing is enabled. Other handlers are free to modify the tracing
context to update information on the fly, such as, func name, enable
body logging etc..
With this commit, requests like this
```
curl -H "Host: ::1:3000" http://localhost:9000/
```
will be traced as well.
Directories markers are not healed when healing a new fresh disk. A
a proper fix would be moving object names encoding/decoding to erasure
object level but it is too late now since the object to set distribution is
calculated at a higher level.
It is observed in a local 8 drive system the CPU seems to be
bottlenecked at
```
(pprof) top
Showing nodes accounting for 1385.31s, 88.47% of 1565.88s total
Dropped 1304 nodes (cum <= 7.83s)
Showing top 10 nodes out of 159
flat flat% sum% cum cum%
724s 46.24% 46.24% 724s 46.24% crypto/sha256.block
219.04s 13.99% 60.22% 226.63s 14.47% syscall.Syscall
158.04s 10.09% 70.32% 158.04s 10.09% runtime.memmove
127.58s 8.15% 78.46% 127.58s 8.15% crypto/md5.block
58.67s 3.75% 82.21% 58.67s 3.75% github.com/minio/highwayhash.updateAVX2
40.07s 2.56% 84.77% 40.07s 2.56% runtime.epollwait
33.76s 2.16% 86.93% 33.76s 2.16% github.com/klauspost/reedsolomon._galMulAVX512Parallel84
8.88s 0.57% 87.49% 11.56s 0.74% runtime.step
7.84s 0.5% 87.99% 7.84s 0.5% runtime.memclrNoHeapPointers
7.43s 0.47% 88.47% 22.18s 1.42% runtime.pcvalue
```
Bonus changes:
- re-use transport for bucket replication clients, also site replication clients.
- use 32KiB buffer for all read and writes at transport layer seems to help
TLS read connections.
- Do not have 'MaxConnsPerHost' this is problematic to be used with net/http
connection pooling 'MaxIdleConnsPerHost' is enough.
- Always reformat all disks when a new disk is detected, this will
ensure new uploads to be written in new fresh disks
- Always heal all buckets first when an erasure set started to be healed
- Use a lock to prevent two disks belonging to different nodes but in
the same erasure set to be healed in parallel
- Heal different sets in parallel
Bonus:
- Avoid logging errUnformattedDisk when a new fresh disk is inserted but
not detected by healing mechanism yet (10 seconds lag)
site replication errors were printed at
various random locations, repeatedly - this
PR attempts to remove double logging and
capture all of them at a common place.
This PR also enhances the code to show
partial success and errors as well.
`mc admin heal -r <alias>` in a multi setup pools returns incorrectly
grey objects. The reason is that erasure-server-pools.HealObject() runs
HealObject in all pools and returns the result of the first nil
error. However, in the lower erasureObject level, HealObject() returns
nil if an object does not exist + missing error in each disk of the object
in that pool, therefore confusing mc.
Make erasureObject.HealObject() to return not found error in the lower
level, so at least erasureServerPools will know what pools to ignore.
Harshavardhana
Anis Elleuch
Krishnan Parthasarathi
Poorna
Andreas Auernhammer
jiuker
Klaus Post
Kaan Kabalak
Florian Schwab
Shireesh Anjal
Aditya Manthramurthy
Taran Pelkey
Taran Pelkey
Javier Adriel
Lenin Alevski
Daryl White
Minio Trusted
yudoutingle
Abirdcfly
Febriananda Wida Pramudita
ebozduman
Cesar Celis Hernandez
Eng Zer Jun
LHHDZ
Praveen raj Mani
haslersn
sota