libvirt/src/qemu/THREADS.txt
Michal Privoznik bd59e0d7c2 qemu: Introduce APIs for manipulating qemuDomainAgentJob
The point is to break QEMU_JOB_* into smaller pieces which
enables us to achieve higher throughput. For instance, if there
are two threads, one is trying to query something on qemu
monitor while the other is trying to query something on agent
monitor these two threads would serialize. There is not much
reason for that.

Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Jiri Denemark <jdenemar@redhat.com>
2018-06-20 13:24:17 +02:00

406 lines
12 KiB
Plaintext

QEMU Driver Threading: The Rules
=================================
This document describes how thread safety is ensured throughout
the QEMU driver. The criteria for this model are:
- Objects must never be exclusively locked for any prolonged time
- Code which sleeps must be able to time out after suitable period
- Must be safe against dispatch of asynchronous events from monitor
Basic locking primitives
------------------------
There are a number of locks on various objects
* virQEMUDriverPtr
The qemu_conf.h file has inline comments describing the locking
needs for each field. Any field marked immutable, self-locking
can be accessed without the driver lock. For other fields there
are typically helper APIs in qemu_conf.c that provide serialized
access to the data. No code outside qemu_conf.c should ever
acquire this lock
* virDomainObjPtr
Will be locked after calling any of the virDomainObjListFindBy{ID,Name,UUID}
methods. However, preferred method is qemuDomObjFromDomain() that uses
virDomainFindByUUIDRef() which also increases the reference counter and
finds the domain in the domain list without blocking all other lookups.
When the domain is locked and the reference increased, the preferred way of
decrementing the reference counter and unlocking the domain is using the
virDomainObjEndAPI() function.
Lock must be held when changing/reading any variable in the virDomainObjPtr
If the lock needs to be dropped & then re-acquired for a short period of
time, the reference count must be incremented first using virDomainObjRef().
There is no need to increase the reference count if qemuDomObjFromDomain()
was used for looking up the domain. In this case there is one reference
already added by that function.
This lock must not be held for anything which sleeps/waits (i.e. monitor
commands).
* qemuMonitorPrivatePtr: Job conditions
Since virDomainObjPtr lock must not be held during sleeps, the job
conditions provide additional protection for code making updates.
QEMU driver uses three kinds of job conditions: asynchronous, agent
and normal.
Asynchronous job condition is used for long running jobs (such as
migration) that consist of several monitor commands and it is
desirable to allow calling a limited set of other monitor commands
while such job is running. This allows clients to, e.g., query
statistical data, cancel the job, or change parameters of the job.
Normal job condition is used by all other jobs to get exclusive
access to the monitor and also by every monitor command issued by an
asynchronous job. When acquiring normal job condition, the job must
specify what kind of action it is about to take and this is checked
against the allowed set of jobs in case an asynchronous job is
running. If the job is incompatible with current asynchronous job,
it needs to wait until the asynchronous job ends and try to acquire
the job again.
Agent job condition is then used when thread wishes to talk to qemu
agent monitor. It is possible to acquire just agent job
(qemuDomainObjBeginAgentJob), or only normal job
(qemuDomainObjBeginJob) or both at the same time
(qemuDomainObjBeginJobWithAgent). Which type of job to grab depends
whether caller wishes to communicate only with agent socket, or only
with qemu monitor socket or both, respectively.
Immediately after acquiring the virDomainObjPtr lock, any method
which intends to update state must acquire asynchronous, normal or
agent job . The virDomainObjPtr lock is released while blocking on
these condition variables. Once the job condition is acquired, a
method can safely release the virDomainObjPtr lock whenever it hits
a piece of code which may sleep/wait, and re-acquire it after the
sleep/wait. Whenever an asynchronous job wants to talk to the
monitor, it needs to acquire nested job (a special kind of normal
job) to obtain exclusive access to the monitor.
Since the virDomainObjPtr lock was dropped while waiting for the
job condition, it is possible that the domain is no longer active
when the condition is finally obtained. The monitor lock is only
safe to grab after verifying that the domain is still active.
* qemuMonitorPtr: Mutex
Lock to be used when invoking any monitor command to ensure safety
wrt any asynchronous events that may be dispatched from the monitor.
It should be acquired before running a command.
The job condition *MUST* be held before acquiring the monitor lock
The virDomainObjPtr lock *MUST* be held before acquiring the monitor
lock.
The virDomainObjPtr lock *MUST* then be released when invoking the
monitor command.
Helper methods
--------------
To lock the virDomainObjPtr
virObjectLock()
- Acquires the virDomainObjPtr lock
virObjectUnlock()
- Releases the virDomainObjPtr lock
To acquire the normal job condition
qemuDomainObjBeginJob()
- Waits until the job is compatible with current async job or no
async job is running
- Waits for job.cond condition 'job.active != 0' using virDomainObjPtr
mutex
- Rechecks if the job is still compatible and repeats waiting if it
isn't
- Sets job.active to the job type
qemuDomainObjEndJob()
- Sets job.active to 0
- Signals on job.cond condition
To acquire the agent job condition
qemuDomainObjBeginAgentJob()
- Waits until there is no other agent job set
- Sets job.agentActive tp the job type
qemuDomainObjEndAgentJob()
- Sets job.agentActive to 0
- Signals on job.cond condition
To acquire both normal and agent job condition
qemuDomainObjBeginJobWithAgent()
- Waits until there is no normal and no agent job set
- Sets both job.active and job.agentActive with required job types
qemuDomainObjEndJobWithAgent()
- Sets both job.active and job.agentActive to 0
- Signals on job.cond condition
To acquire the asynchronous job condition
qemuDomainObjBeginAsyncJob()
- Waits until no async job is running
- Waits for job.cond condition 'job.active != 0' using virDomainObjPtr
mutex
- Rechecks if any async job was started while waiting on job.cond
and repeats waiting in that case
- Sets job.asyncJob to the asynchronous job type
qemuDomainObjEndAsyncJob()
- Sets job.asyncJob to 0
- Broadcasts on job.asyncCond condition
To acquire the QEMU monitor lock
qemuDomainObjEnterMonitor()
- Acquires the qemuMonitorObjPtr lock
- Releases the virDomainObjPtr lock
qemuDomainObjExitMonitor()
- Releases the qemuMonitorObjPtr lock
- Acquires the virDomainObjPtr lock
These functions must not be used by an asynchronous job.
Note that the virDomainObj is unlocked during the time in
monitor and it can be changed, e.g. if QEMU dies, qemuProcessStop
may free the live domain definition and put the persistent
definition back in vm->def. The callers should check the return
value of ExitMonitor to see if the domain is still alive.
To acquire the QEMU monitor lock as part of an asynchronous job
qemuDomainObjEnterMonitorAsync()
- Validates that the right async job is still running
- Acquires the qemuMonitorObjPtr lock
- Releases the virDomainObjPtr lock
- Validates that the VM is still active
qemuDomainObjExitMonitor()
- Releases the qemuMonitorObjPtr lock
- Acquires the virDomainObjPtr lock
These functions are for use inside an asynchronous job; the caller
must check for a return of -1 (VM not running, so nothing to exit).
Helper functions may also call this with QEMU_ASYNC_JOB_NONE when
used from a sync job (such as when first starting a domain).
To keep a domain alive while waiting on a remote command
qemuDomainObjEnterRemote()
- Releases the virDomainObjPtr lock
qemuDomainObjExitRemote()
- Acquires the virDomainObjPtr lock
Design patterns
---------------
* Accessing something directly to do with a virDomainObjPtr
virDomainObjPtr obj;
obj = qemuDomObjFromDomain(dom);
...do work...
virDomainObjEndAPI(&obj);
* Updating something directly to do with a virDomainObjPtr
virDomainObjPtr obj;
obj = qemuDomObjFromDomain(dom);
qemuDomainObjBeginJob(obj, QEMU_JOB_TYPE);
...do work...
qemuDomainObjEndJob(obj);
virDomainObjEndAPI(&obj);
* Invoking a monitor command on a virDomainObjPtr
virDomainObjPtr obj;
qemuDomainObjPrivatePtr priv;
obj = qemuDomObjFromDomain(dom);
qemuDomainObjBeginJob(obj, QEMU_JOB_TYPE);
...do prep work...
if (virDomainObjIsActive(vm)) {
qemuDomainObjEnterMonitor(obj);
qemuMonitorXXXX(priv->mon);
qemuDomainObjExitMonitor(obj);
}
...do final work...
qemuDomainObjEndJob(obj);
virDomainObjEndAPI(&obj);
* Invoking an agent command on a virDomainObjPtr
virDomainObjPtr obj;
qemuAgentPtr agent;
obj = qemuDomObjFromDomain(dom);
qemuDomainObjBeginAgentJob(obj, QEMU_AGENT_JOB_TYPE);
...do prep work...
if (!qemuDomainAgentAvailable(obj, true))
goto cleanup;
agent = qemuDomainObjEnterAgent(obj);
qemuAgentXXXX(agent, ..);
qemuDomainObjExitAgent(obj, agent);
...do final work...
qemuDomainObjEndAgentJob(obj);
virDomainObjEndAPI(&obj);
* Invoking both monitor and agent commands on a virDomainObjPtr
virDomainObjPtr obj;
qemuAgentPtr agent;
obj = qemuDomObjFromDomain(dom);
qemuDomainObjBeginJobWithAgent(obj, QEMU_JOB_TYPE, QEMU_AGENT_JOB_TYPE);
if (!virDomainObjIsActive(dom))
goto cleanup;
...do prep work...
if (!qemuDomainAgentAvailable(obj, true))
goto cleanup;
agent = qemuDomainObjEnterAgent(obj);
qemuAgentXXXX(agent, ..);
qemuDomainObjExitAgent(obj, agent);
...
qemuDomainObjEnterMonitor(obj);
qemuMonitorXXXX(priv->mon);
qemuDomainObjExitMonitor(obj);
/* Alternatively, talk to the monitor first and then talk to the agent. */
...do final work...
qemuDomainObjEndJobWithAgent(obj);
virDomainObjEndAPI(&obj);
* Running asynchronous job
virDomainObjPtr obj;
qemuDomainObjPrivatePtr priv;
obj = qemuDomObjFromDomain(dom);
qemuDomainObjBeginAsyncJob(obj, QEMU_ASYNC_JOB_TYPE);
qemuDomainObjSetAsyncJobMask(obj, allowedJobs);
...do prep work...
if (qemuDomainObjEnterMonitorAsync(driver, obj,
QEMU_ASYNC_JOB_TYPE) < 0) {
/* domain died in the meantime */
goto error;
}
...start qemu job...
qemuDomainObjExitMonitor(driver, obj);
while (!finished) {
if (qemuDomainObjEnterMonitorAsync(driver, obj,
QEMU_ASYNC_JOB_TYPE) < 0) {
/* domain died in the meantime */
goto error;
}
...monitor job progress...
qemuDomainObjExitMonitor(driver, obj);
virObjectUnlock(obj);
sleep(aWhile);
virObjectLock(obj);
}
...do final work...
qemuDomainObjEndAsyncJob(obj);
virDomainObjEndAPI(&obj);
* Coordinating with a remote server for migration
virDomainObjPtr obj;
qemuDomainObjPrivatePtr priv;
obj = qemuDomObjFromDomain(dom);
qemuDomainObjBeginAsyncJob(obj, QEMU_ASYNC_JOB_TYPE);
...do prep work...
if (virDomainObjIsActive(vm)) {
qemuDomainObjEnterRemote(obj);
...communicate with remote...
qemuDomainObjExitRemote(obj);
/* domain may have been stopped while we were talking to remote */
if (!virDomainObjIsActive(vm)) {
qemuReportError(VIR_ERR_INTERNAL_ERROR, "%s",
_("guest unexpectedly quit"));
}
}
...do final work...
qemuDomainObjEndAsyncJob(obj);
virDomainObjEndAPI(&obj);