[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH v3 04/15] argo: init, destroy and soft-reset, with enable command line opt




On Jan 14, 2019, at 06:32, Roger Pau Monné <royger@xxxxxxxxxxx> wrote:

On Mon, Jan 14, 2019 at 9:32 AM Christopher Clark
<christopher.w.clark@xxxxxxxxx> wrote:

I've written a document about the locking to add to the tree with the
series, and a copy is at github here:

https://github.com/dozylynx/xen/blob/0cb95385eba696ecf4856075a524c5e528e60455/docs/misc/argo-locking.md

Thanks. It would have been better to send the contents of the document
to the list, so inline comments can be added. It's hard to comment on
the document now since it's only on github AFAICT.

Here's an inline copy of the doc:

=== begin document ===

# Argo: Locking

## Introduction

Argo is an interdomain communication mechanism. It has requirements for performance isolation between domains, to prevent negative performance impact from malicious or disruptive activity of other domains, or even other vcpus of the same domain operating other rings.

Since Argo operates a data path between domains, sections of this code are *hot* when the communication paths are in use. To encourage high performance, a goal is to limit mutual exclusion to only where required and enable significant concurrency.

Avoidance of deadlock is essential and since state must frequently be updated that pertains to more than one domain, a locking protocol defines which locks are needed and the order of their acquistion.

## Structure

The granular locking structure of Argo enables:

1. Performance isolation of guests
2. Avoidance of DoS of rings by domains that are not authorized to send to them
3. Deadlock-free teardown of state across multiple domains on domain destroy
4. Performance of guests using Argo with concurrent operation of rings.

Argo uses three per-domain locks to protect three separate data structures.  Access to the ring_hash data structure is confined to domains that a ring-registering domain has authorized to send data via the ring.  The complete set of Argo locks is:

* Global : `L1_global_argo_rwlock`
* Per-domain: `rings_L2_rwlock`
* Per-domain: `send_L2_lock`
* Per-domain: `wildcard_L2_lock`
* Per-ring: `L3_lock`

## Protected State

The data structures being protected by the locks are all per-domain. The only global Argo state is the `L1_global_argo_rwlock` used to coordinate access to data structures of other domains.

### State: Rings registered and owned by a domain

This includes the state to run that ring, such as memory frame numbers and established mappings. Per-ring state is protected by its own lock, so that multiple VCPUs of the same domain operating different rings do not inhibit the performance of each other.

The per-domain ring state also includes the list of pending notifications for other domains that are waiting for ring space availability.

### State: Partner rings registered by other domains that this domain is the single allowed sender

This state belonging to the permitted sender is written to when a ring is registered by another domain. The lock that protects this state is subject to locking at arbitrary frequency by those foreign domains when registering rings -- which do not need any permission granted by this domain in order to register a ring to communicate with it --  so it must not inhibit the domain's own ability to use its own rings, to protect them from DoS. For this reason, this state is protected by its own lock.

### State: Pending notifications to this domain about wildcard rings registered by other domains

This data structure is needed when a domain is destroyed, to cancel the outstanding space availability notifications about the wildcard rings of other domains that this domain has queried.

Data is entered into this data structure by the domain that owns it, either by a space-inhibited sendv or a notify operation.

Data is removed from this data structure in one of three cases: when space becomes available in the destination ring and the notification is sent, when the ring is torn down, or when the awaiting domain is destroyed.

In the case where a notification is sent, access to the data structure is triggered by the ring owner domain, rather than the domain waiting for notification. This data structure is protected by its own lock since doing so entails less contention than the alternative of reusing an existing lock owned by the domain.

## Hierarchical Locking Model and Protocol

The locking discipline within the Argo code is heirarchical and utilizes reader/writer locks to enable increased concurrency when operations do not conflict. None of the Argo locks are reentrant.

The hierarchy:

* There is a global rwlock (`L1`) to protect access to all of the per-domain argo data structures.
* There is a rwlock per-domain (`rings_L2`) to protect the hashtable of the per-ring data structures.
* There is a lock per ring (`L3`) to protect the per-ring data structure, `struct argo_ring_info`.

There are a two other per-domain L2 locks; their operation is similar and they are described later.

The protocol to safely acquire write access to the per-ring data structure, `struct argo_ring_info`, is:

1) Acquire a Read lock on L1.
2) Acquire a Read lock on L2.
3) Acquire L3.

An alternative valid sequence is:

1) Acquire a Read lock on L1.
2) Acquire a Write lock on L2.

This second sequence grants write access to _all_ of the `argo_ring_info` structs belonging to the domain, but at the expense of less concurrency: no other operation can access those structs while the locks are held, which will inhibit operations on those rings until the locks are released.

Another alternative valid sequence is:

1) Acquire a Write lock on L1.

This grants write access to _all_ of the `argo_ring_info` structs belonging to _all domains_, but again at the expense of far less concurrency: no other operation can operate on Argo rings until the locks are released.

## Lock Definitions

The full set of locks that are directly operated upon by the Argo code are described in the following section.

### The global singleton lock:

* `L1_global_argo_rwlock`

The rationale for having a global lock is to be able to enforce system-wide exclusion for a critical region and simplify the logic required to avoid deadlock, for teardown of state across multiple domains when a domain is destroyed.

The majority of operations take a read-lock on this lock, allowing concurrent Argo operations by many domains.

The pointer d->argo on every domain is protected by this lock. A set of more granular per-domain locks could be used to do that, but since domain start and stop is expected to be a far less frequent operation than the other argo operations, acquiring a single read lock to enable access to all the argo structs of all domains simplifies the protocol.

Points of write-locking on this lock:

* `argo_destroy`, where:
  * All of the domain's own rings are destroyed.
      * All of the notifications pending for other domains are cancelled.
   * All of the unicast partner rings owned by other domains for this domain to send to, are destroyed.
      * All of the notifications pending on those rings are cancelled.
   * All of the notifications pending for this domain on wildcard rings owned by other domains are cancelled.
* `argo_soft_reset`, for similar teardown operations as argo_destroy.
* `argo_init`, where the `d->argo` pointer is first populated.
  * Since the write lock is taken here, there is serialization all concurrent Argo operations around this single pointer write; this is the cost of using the simpler one global lock approach.

Enforcing that the write_lock is acquired on `L1_global_argo_rwlock` before executing teardown, ensures that no teardown operations act concurrently and no other Argo operations happen concurrently with a teardown. The teardown logic is free to safely modify the Argo state across all domains without having to acquire per-domain locks and deadlock cannot occur.

### Per-Domain: Ring hash lock

`rings_L2_rwlock`

Protects: the per-domain ring hash table of `argo_ring_info` structs.

Holding a read lock on `rings_L2` protects the ring hash table and the elements in the hash table `d->argo->ring_hash`, and the `node` and `id` fields in struct `argo_ring_info` in the hash table.

Holding a write lock on `rings_L2` protects all of the elements of all the struct `argo_ring_info` belonging to this domain.

To take `rings_L2` you must already have `R(L1)`. `W(L1)` implies `W(rings_L2)` and `L3`.

Prerequisites:

* `R(L1_global_argo_rwlock)` must be acquired before taking either read or write on `rings_L2_rwlock`.
* `W(L1_global_argo_rwlock)` implies `W(rings_L2_rwlock)`, so if `W(L1_global_argo_rwlock)` is held, then `rings_L2_rwlock` does not need to be acquired, and all the data structures that `rings_L2_rwlock` protects can be accessed as if `W(ring_L2_rwlock)` was held.

Is accessed by the hypervisor on behalf of:

* The domain that registered the ring.
* Any domain that is allowed to send to the ring -- so that's the partner domain, for unicast rings, or any domain, for wildcard rings.

### Send hash lock

`send_L2_lock`

Protects: the per-domain send hash table of `argo_send_info` structs.

Is accessed by the hypervisor on behalf of:

* Any domain that registers a ring that specifies the domain as the unicast sender.
* The domain that has been allowed to send, as part of teardown when the domain is being destroyed.


### Wildcard pending list lock

`wildcard_L2_lock`

Protects: the per-domain list of pending notifications to the domain from wildcard rings owned by other domains.

Is accessed by the hypervisor on behalf of:

* The domain that issued a query to another about space availability in one of its wildcard rings - this can be done by attempting a send operation when there is insufficient ring space available at the time.
* Any domain that the domain has issued a query to about space availability in one of their wildcard rings.

### Per-Ring locks:

* `L3_lock`

This lock protects the members of a `struct ring_info` which is the primary state for a domain's own registered ring.


## Reasoning Model

A common model for reasoning about concurrent code focusses on accesses to individual variables: if code touches this variable, see that it first acquires the corresponding lock and then drops it afterwards. A challenge with this model is in ensuring that the sequence of locks acquired within nested functions, when operating on data from multiple domains with concurrent operations, is safe from deadlock.

An alternative method that is better suited to the Argo software is to consider the execution path, the full sequence of locks acquired, accesses performed, and locks released, from entering an operation, to the completion of the work.

An example code path for an operation:

`[entry] > -- [ take R(L1) ] -- [ take R(L2) ] -- loop [ take a L3 / drop L3 ] --  [ drop R(L2) ] -- [ drop R(L1)] -- > [exit]`

If a function implements a section of the path, it is important to know not only what variables the function itself operates upon, but also the locking state that will already have been established at the point when the function is invoked, since this will affect what data the function can access. For this reason, comments in the code, or ASSERTs that explicitly check lock state, communicate what the locking state is expected and intended to be when that code is invoked. See the macros defined to support this for Argo later in this document.


## Macros to Validate and Document Lock State

These macros encode the logic to verify that the locking has adhered to the locking discipline.

eg. On entry to logic that requires holding at least `R(rings_L2)`, this:

`ASSERT(LOCKING_Read_rings_L2(d));`

checks that the lock state is sufficient, validating that one of the following must be true when executed:

`R(rings_L2) && R(L1)`
or:  `W(rings_L2) && R(L1)`
or:  `W(L1)`

The macros are defined thus:

```
/* RAW macros here are only used to assist defining the other macros below */
#define RAW_LOCKING_Read_L1 (rw_is_locked(&L1_global_argo_rwlock))
#define RAW_LOCKING_Read_rings_L2(d) \
  (rw_is_locked(&d->argo->rings_L2_rwlock) && RAW_LOCKING_Read_L1)

/* The LOCKING macros defined below here are for use at verification points */
#define LOCKING_Write_L1 (rw_is_write_locked(&L1_global_argo_rwlock))
#define LOCKING_Read_L1 (RAW_LOCKING_Read_L1 || LOCKING_Write_L1)

#define LOCKING_Write_rings_L2(d) \
  ((RAW_LOCKING_Read_L1 && rw_is_write_locked(&d->argo->rings_L2_rwlock)) || \
   LOCKING_Write_L1)

#define LOCKING_Read_rings_L2(d) \
  ((RAW_LOCKING_Read_L1 && rw_is_locked(&d->argo->rings_L2_rwlock)) || \
   LOCKING_Write_rings_L2(d) || LOCKING_Write_L1)

#define LOCKING_L3(d, r) \
  ((RAW_LOCKING_Read_rings_L2(d) && spin_is_locked(&r->L3_lock)) || \
   LOCKING_Write_rings_L2(d) || LOCKING_Write_L1)

#define LOCKING_send_L2(d) \
  ((RAW_LOCKING_Read_L1 && spin_is_locked(&d->argo->send_L2_lock)) || \
   LOCKING_Write_L1)
```

Here is an example of a macro in use:

```
static void
notify_ring(const struct domain *d, struct argo_ring_info *ring_info,
          struct hlist_head *to_notify)
{
  uint32_t space;

  ASSERT(LOCKING_Read_rings_L2(d));

  spin_lock(&ring_info->L3_lock);

  if ( ring_info->len )
      space = ringbuf_payload_space(d, ring_info);
  else
      space = 0;

  spin_unlock(&ring_info->L3_lock);

  if ( space )
      pending_find(d, ring_info, space, to_notify);
}

```

In the above example, it can be seen that it is safe to acquire the `L3` lock because _at least_ `R(rings_L2)` is already held, as documented and verified by the macro.

## Appendix:  FAQ / Other Considerations

### Why not have a single per-domain lock?

Due to performance isolation / DoS avoidance: if there is a single per-domain lock, acquiring this lock will stall operations on other active rings owned by the domain. A malicious domain can loop registering and unregistering rings, without any consent by the targetted domain, which would experience decreased throughput due to the contention on the single per-domain lock. The granular locking structure of Argo prevents this. It also allows concurrent operation of different rings by multiple VCPUs of the same domain without contention, to avoid negative application performance interaction.

## Rationale for Using a Singleton Global Lock: L1

### Teardown on domain destroy

The single global lock enables exclusive access to the argo data structures across domains when a domain is destroyed. Every unicast ring that the dying domain is the authorized sender is torn down and any pending space-available notifications in other domain's wildcard rings are cancelled. This requires gaining safe access to the data structures on each of the domains involved.

The 'send hashtable' data structure is needed in order to perform the teardown of rings when a domain is destroyed. To populate it, whenever a unicast ring is registered, the lock that protects that data structure must be taken exclusively.

There are granular per-domain locks which protect the per-domain data structures. The global singleton L1 lock operates with-and-above the per-domain locks and is used to obtain exclusive access to multiple domain's argo data structures in the infrequent case where it is used -- for domain destroy -- whilst otherwise allowing concurrent access, via acquiring it with 'read' access, for the majority of the time.

To perform the required state teardown on domain destruction, which can require removing state from the data structures of multiple domains, a locking protocol to obtain mutual exclusion and safe access to the state is required, without deadlocking.

Using the single global lock avoids the need for sequencing the acquisition of multiple individual per-domain locks (and lower level data structure locks) to prevent deadlock: taking W(L1) grants access to all and taking R(L1) ensures that teardown of any domain will not interfere with any Argo hypercall operation. It enables introducing granular locking without complex or error-prone lock acquisition logic.

=== end document ===

There are also several claims that fine-grainer locking provides
better performance in order to justify the need of such locks. IMO
without providing any evidence of such performance benefit it's hard
to be convinced so many locks are needed.

Benchmarks would be useful for regression testing.  We can investigate resourcing for Argo synthetic benchmarks in the Xen 4.13 release cycle.  In the meantime, we can cite the shipment of Citrix XenClient in 2011, followed by customer production deployments of v4v in OpenXT and Bromium uXen (including HP business laptops).  Argo is derived from v4v.

Ian Pratt's PSEC 2018 presentation [1] on hypervisor security indirectly referenced v4v and uXen isolation/performance requirements, it may illustrate the scope of possible Argo use cases.  Click the "uXen" button below the video to navigate to the clip.  There's also a uXen source code link.  An excerpt (my annotations in []):

"PV device interfaces all built on simple hypervisor copy-based primitive 

We came up with a very simple primitive [v4v] for communication between VMs, between VMs and the host, and then just built everything on that primitive.  It's a simple copy-based primitive.  We didn't want any memory sharing of any kind.  Anything involving memory sharing ends up being complex ... Grant tables being a huge example ... a copy-based approach is just much simpler, and actually, performance-wise, it turns out being equivalent from a performance point of view ... 

Getting rid of other terrible ideas, like Xenstore ... we still want the primitive [v4v] interface to be able to support things like device reconnection.  That was a good idea, this idea that you can restart these driver domains, have VMs reconnect and be able to continue, that's very useful.  We want to enable this simple primitive [v4v] to do that.  Then just build everything using very simple, narrow interfaces."



Regards,
Rich
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.