open-nomad/client/node_updater.go

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

452 lines
13 KiB
Go
Raw Permalink Normal View History

// Copyright (c) HashiCorp, Inc.
// SPDX-License-Identifier: MPL-2.0
package client
import (
"context"
"fmt"
"sync"
"time"
"github.com/hashicorp/nomad/client/devicemanager"
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
"github.com/hashicorp/nomad/client/pluginmanager/csimanager"
"github.com/hashicorp/nomad/client/pluginmanager/drivermanager"
"github.com/hashicorp/nomad/nomad/structs"
)
var (
// batchFirstFingerprintsTimeout is the maximum amount of time to wait for
// initial fingerprinting to complete before sending a batched Node update
batchFirstFingerprintsTimeout = 50 * time.Second
)
// batchFirstFingerprints waits for the first fingerprint response from all
// plugin managers and sends a single Node update for all fingerprints. It
// should only ever be called once
func (c *Client) batchFirstFingerprints() {
ctx, cancel := context.WithTimeout(context.Background(), batchFirstFingerprintsTimeout)
defer cancel()
ch, err := c.pluginManagers.WaitForFirstFingerprint(ctx)
if err != nil {
c.logger.Warn("failed to batch initial fingerprint updates, switching to incemental updates")
goto SEND_BATCH
}
// Wait for fingerprinting to complete or timeout before processing batches
select {
case <-ch:
case <-ctx.Done():
}
SEND_BATCH:
c.configLock.Lock()
defer c.configLock.Unlock()
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
newConfig := c.config.Copy()
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
// csi updates
var csiChanged bool
c.batchNodeUpdates.batchCSIUpdates(func(name string, info *structs.CSIInfo) {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
if c.updateNodeFromCSIControllerLocked(name, info, newConfig.Node) {
if newConfig.Node.CSIControllerPlugins[name].UpdateTime.IsZero() {
newConfig.Node.CSIControllerPlugins[name].UpdateTime = time.Now()
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
}
csiChanged = true
}
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
if c.updateNodeFromCSINodeLocked(name, info, newConfig.Node) {
if newConfig.Node.CSINodePlugins[name].UpdateTime.IsZero() {
newConfig.Node.CSINodePlugins[name].UpdateTime = time.Now()
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
}
csiChanged = true
}
})
// driver node updates
var driverChanged bool
c.batchNodeUpdates.batchDriverUpdates(func(driver string, info *structs.DriverInfo) {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
if c.applyNodeUpdatesFromDriver(driver, info, newConfig.Node) {
newConfig.Node.Drivers[driver] = info
if newConfig.Node.Drivers[driver].UpdateTime.IsZero() {
newConfig.Node.Drivers[driver].UpdateTime = time.Now()
}
driverChanged = true
}
})
// device node updates
var devicesChanged bool
c.batchNodeUpdates.batchDevicesUpdates(func(devices []*structs.NodeDeviceResource) {
if c.updateNodeFromDevicesLocked(devices) {
newConfig.Node.NodeResources.Devices = devices
devicesChanged = true
}
})
// only update the node if changes occurred
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
if driverChanged || devicesChanged || csiChanged {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
c.config = newConfig
c.updateNode()
}
close(c.fpInitialized)
}
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
// updateNodeFromCSI receives a CSIInfo struct for the plugin and updates the
// node accordingly
func (c *Client) updateNodeFromCSI(name string, info *structs.CSIInfo) {
c.configLock.Lock()
defer c.configLock.Unlock()
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
newConfig := c.config.Copy()
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
changed := false
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
if c.updateNodeFromCSIControllerLocked(name, info, newConfig.Node) {
if newConfig.Node.CSIControllerPlugins[name].UpdateTime.IsZero() {
newConfig.Node.CSIControllerPlugins[name].UpdateTime = time.Now()
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
}
changed = true
}
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
if c.updateNodeFromCSINodeLocked(name, info, newConfig.Node) {
if newConfig.Node.CSINodePlugins[name].UpdateTime.IsZero() {
newConfig.Node.CSINodePlugins[name].UpdateTime = time.Now()
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
}
changed = true
}
if changed {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
c.config = newConfig
c.updateNode()
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
}
}
// updateNodeFromCSIControllerLocked makes the changes to the node from a csi
// update but does not send the update to the server. c.configLock must be held
// before calling this func.
//
// It is safe to call for all CSI Updates, but will only perform changes when
// a ControllerInfo field is present.
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
func (c *Client) updateNodeFromCSIControllerLocked(name string, info *structs.CSIInfo, node *structs.Node) bool {
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
var changed bool
if info.ControllerInfo == nil {
return false
}
i := info.Copy()
i.NodeInfo = nil
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
oldController, hadController := node.CSIControllerPlugins[name]
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
if !hadController {
// If the controller info has not yet been set, do that here
changed = true
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
node.CSIControllerPlugins[name] = i
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
} else {
// The controller info has already been set, fix it up
if !oldController.Equal(i) {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
node.CSIControllerPlugins[name] = i
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
changed = true
}
// If health state has changed, trigger node event
if oldController.Healthy != i.Healthy || oldController.HealthDescription != i.HealthDescription {
changed = true
if i.HealthDescription != "" {
event := structs.NewNodeEvent().
SetSubsystem("CSI").
SetMessage(i.HealthDescription).
AddDetail("plugin", name).
AddDetail("type", "controller")
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
c.triggerNodeEvent(event)
}
}
}
return changed
}
// updateNodeFromCSINodeLocked makes the changes to the node from a csi
// update but does not send the update to the server. c.configLock must be hel
// before calling this func.
//
// It is safe to call for all CSI Updates, but will only perform changes when
// a NodeInfo field is present.
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
func (c *Client) updateNodeFromCSINodeLocked(name string, info *structs.CSIInfo, node *structs.Node) bool {
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
var changed bool
if info.NodeInfo == nil {
return false
}
i := info.Copy()
i.ControllerInfo = nil
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
oldNode, hadNode := node.CSINodePlugins[name]
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
if !hadNode {
// If the Node info has not yet been set, do that here
changed = true
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
node.CSINodePlugins[name] = i
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
} else {
// The node info has already been set, fix it up
if !oldNode.Equal(info) {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
node.CSINodePlugins[name] = i
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
changed = true
}
// If health state has changed, trigger node event
if oldNode.Healthy != i.Healthy || oldNode.HealthDescription != i.HealthDescription {
changed = true
if i.HealthDescription != "" {
event := structs.NewNodeEvent().
SetSubsystem("CSI").
SetMessage(i.HealthDescription).
AddDetail("plugin", name).
AddDetail("type", "node")
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
c.triggerNodeEvent(event)
}
}
}
return changed
}
// updateNodeFromDriver receives a DriverInfo struct for the driver and updates
// the node accordingly
func (c *Client) updateNodeFromDriver(name string, info *structs.DriverInfo) {
c.configLock.Lock()
defer c.configLock.Unlock()
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
newConfig := c.config.Copy()
if c.applyNodeUpdatesFromDriver(name, info, newConfig.Node) {
newConfig.Node.Drivers[name] = info
if newConfig.Node.Drivers[name].UpdateTime.IsZero() {
newConfig.Node.Drivers[name].UpdateTime = time.Now()
}
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
c.config = newConfig
c.updateNode()
}
}
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
// applyNodeUpdatesFromDriver applies changes to the passed in node. true is
// returned if the node has changed.
func (c *Client) applyNodeUpdatesFromDriver(name string, info *structs.DriverInfo, node *structs.Node) bool {
var hasChanged bool
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
hadDriver := node.Drivers[name] != nil
if !hadDriver {
// If the driver info has not yet been set, do that here
hasChanged = true
for attrName, newVal := range info.Attributes {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
node.Attributes[attrName] = newVal
}
} else {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
oldVal := node.Drivers[name]
// The driver info has already been set, fix it up
if oldVal.Detected != info.Detected {
hasChanged = true
}
// If health state has change, trigger node event
if oldVal.Healthy != info.Healthy || oldVal.HealthDescription != info.HealthDescription {
hasChanged = true
if info.HealthDescription != "" {
event := structs.NewNodeEvent().
SetSubsystem("Driver").
SetMessage(info.HealthDescription).
AddDetail("driver", name)
c.triggerNodeEvent(event)
}
}
for attrName, newVal := range info.Attributes {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
oldVal := node.Drivers[name].Attributes[attrName]
if oldVal == newVal {
continue
}
hasChanged = true
if newVal == "" {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
delete(node.Attributes, attrName)
} else {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
node.Attributes[attrName] = newVal
}
}
}
// COMPAT Remove in Nomad 0.10
// We maintain the driver enabled attribute until all drivers expose
// their attributes as DriverInfo
driverName := fmt.Sprintf("driver.%s", name)
if info.Detected {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
node.Attributes[driverName] = "1"
} else {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
delete(node.Attributes, driverName)
}
return hasChanged
}
func (c *Client) updateNodeFromDevices(devices []*structs.NodeDeviceResource) {
c.configLock.Lock()
defer c.configLock.Unlock()
// Not updating node.Resources: the field is deprecated and includes
// dispatched task resources and not appropriate for expressing
// node available device resources
if c.updateNodeFromDevicesLocked(devices) {
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
c.updateNode()
}
}
// updateNodeFromDevicesLocked updates the node with the results of devices,
// but does send the update to the server. c.configLock must be held before
// calling this func
func (c *Client) updateNodeFromDevicesLocked(devices []*structs.NodeDeviceResource) bool {
if !structs.DevicesEquals(c.config.Node.NodeResources.Devices, devices) {
c.logger.Debug("new devices detected", "devices", len(devices))
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
newConfig := c.config.Copy()
newConfig.Node.NodeResources.Devices = devices
c.config = newConfig
return true
}
return false
}
// batchNodeUpdates allows for batching multiple Node updates from fingerprinting.
// Once ready, the batches can be flushed and toggled to stop batching and forward
// all updates to a configured callback to be performed incrementally
type batchNodeUpdates struct {
// access to driver fields must hold driversMu lock
drivers map[string]*structs.DriverInfo
driversBatched bool
driverCB drivermanager.UpdateNodeDriverInfoFn
driversMu sync.Mutex
// access to devices fields must hold devicesMu lock
devices []*structs.NodeDeviceResource
devicesBatched bool
devicesCB devicemanager.UpdateNodeDevicesFn
devicesMu sync.Mutex
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
// access to csi fields must hold csiMu lock
csiNodePlugins map[string]*structs.CSIInfo
csiControllerPlugins map[string]*structs.CSIInfo
csiBatched bool
csiCB csimanager.UpdateNodeCSIInfoFunc
csiMu sync.Mutex
}
func newBatchNodeUpdates(
driverCB drivermanager.UpdateNodeDriverInfoFn,
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
devicesCB devicemanager.UpdateNodeDevicesFn,
csiCB csimanager.UpdateNodeCSIInfoFunc) *batchNodeUpdates {
return &batchNodeUpdates{
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
drivers: make(map[string]*structs.DriverInfo),
driverCB: driverCB,
devices: []*structs.NodeDeviceResource{},
devicesCB: devicesCB,
csiNodePlugins: make(map[string]*structs.CSIInfo),
csiControllerPlugins: make(map[string]*structs.CSIInfo),
csiCB: csiCB,
}
}
// updateNodeFromCSI implements csimanager.UpdateNodeCSIInfoFunc and is used in
// the csi manager to send csi fingerprints to the server.
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
func (b *batchNodeUpdates) updateNodeFromCSI(plugin string, info *structs.CSIInfo) {
b.csiMu.Lock()
defer b.csiMu.Unlock()
if b.csiBatched {
b.csiCB(plugin, info)
return
}
// Only one of these is expected to be set, but a future implementation that
// explicitly models monolith plugins with a single fingerprinter may set both
if info.ControllerInfo != nil {
b.csiControllerPlugins[plugin] = info
}
if info.NodeInfo != nil {
b.csiNodePlugins[plugin] = info
}
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
}
// batchCSIUpdates sends all of the batched CSI updates by calling f for each
// plugin batched
func (b *batchNodeUpdates) batchCSIUpdates(f csimanager.UpdateNodeCSIInfoFunc) error {
b.csiMu.Lock()
defer b.csiMu.Unlock()
if b.csiBatched {
return fmt.Errorf("csi updates already batched")
}
b.csiBatched = true
for plugin, info := range b.csiNodePlugins {
f(plugin, info)
}
for plugin, info := range b.csiControllerPlugins {
f(plugin, info)
}
CSI Plugin Registration (#6555) This changeset implements the initial registration and fingerprinting of CSI Plugins as part of #5378. At a high level, it introduces the following: * A `csi_plugin` stanza as part of a Nomad task configuration, to allow a task to expose that it is a plugin. * A new task runner hook: `csi_plugin_supervisor`. This hook does two things. When the `csi_plugin` stanza is detected, it will automatically configure the plugin task to receive bidirectional mounts to the CSI intermediary directory. At runtime, it will then perform an initial heartbeat of the plugin and handle submitting it to the new `dynamicplugins.Registry` for further use by the client, and then run a lightweight heartbeat loop that will emit task events when health changes. * The `dynamicplugins.Registry` for handling plugins that run as Nomad tasks, in contrast to the existing catalog that requires `go-plugin` type plugins and to know the plugin configuration in advance. * The `csimanager` which fingerprints CSI plugins, in a similar way to `drivermanager` and `devicemanager`. It currently only fingerprints the NodeID from the plugin, and assumes that all plugins are monolithic. Missing features * We do not use the live updates of the `dynamicplugin` registry in the `csimanager` yet. * We do not deregister the plugins from the client when they shutdown yet, they just become indefinitely marked as unhealthy. This is deliberate until we figure out how we should manage deploying new versions of plugins/transitioning them.
2019-10-22 13:20:26 +00:00
return nil
}
// updateNodeFromDriver implements drivermanager.UpdateNodeDriverInfoFn and is
// used in the driver manager to send driver fingerprints to
func (b *batchNodeUpdates) updateNodeFromDriver(driver string, info *structs.DriverInfo) {
b.driversMu.Lock()
defer b.driversMu.Unlock()
if b.driversBatched {
b.driverCB(driver, info)
return
}
b.drivers[driver] = info
}
// batchDriverUpdates sends all of the batched driver node updates by calling f
// for each driver batched
func (b *batchNodeUpdates) batchDriverUpdates(f drivermanager.UpdateNodeDriverInfoFn) error {
b.driversMu.Lock()
defer b.driversMu.Unlock()
if b.driversBatched {
return fmt.Errorf("driver updates already batched")
}
b.driversBatched = true
for driver, info := range b.drivers {
f(driver, info)
}
return nil
}
// updateNodeFromDevices implements devicemanager.UpdateNodeDevicesFn and is
// used in the device manager to send device fingerprints to
func (b *batchNodeUpdates) updateNodeFromDevices(devices []*structs.NodeDeviceResource) {
b.devicesMu.Lock()
defer b.devicesMu.Unlock()
if b.devicesBatched {
b.devicesCB(devices)
return
}
b.devices = devices
}
// batchDevicesUpdates sends the batched device node updates by calling f with
// the devices
func (b *batchNodeUpdates) batchDevicesUpdates(f devicemanager.UpdateNodeDevicesFn) error {
b.devicesMu.Lock()
defer b.devicesMu.Unlock()
if b.devicesBatched {
return fmt.Errorf("devices updates already batched")
}
b.devicesBatched = true
f(b.devices)
return nil
}