Add support for indirect descriptors in VirtQueue with alloc feature

[qwandor]

This is based on some of the work in #96, but fixes some issues there with how allocation and DMA memory is handled, adds tests for the indirect virtqueue itself, and also adds support to the fake device support so that device tests will work with indirect queues.

Devices can choose when constructing a VirtQueue whether it should support indirect descriptors or not, based on their preference and the device features. For now only the VirtIOBlk driver uses indirect descriptors (if the underlying device supports it). If a VirtQueue is constructed with indirect descriptor support then it will use them whenever more than one buffer is added to the queue. There's no point using indirect descriptors for a single buffer so we don't bother.

Space for the indirect descriptor lists is allocated (and shared with the device via the HAL) dynamically. This means that it depends on the alloc feature, which is enabled by default. If the alloc feature is disabled, then indirect descriptors will never be used.

[equation314]

LGTM

[equation314]

Maybe we should add a new feature "indirect-desc"?

[qwandor]

Maybe we should add a new feature "indirect-desc"?

You mean a cargo feature? Is there any case where you have an allocator available, the VirtIO device implementation supports indirect descriptors, but you wouldn't want to use them? I can't think of one, but if so we could add a flag to the device driver constructor so it could be decided at runtime.

[equation314]

The conditional compilation attribute #[cfg(feature = "alloc")] is confused, I think use #[cfg(feature = "indirect-desc")] is more better.

[qwandor]

The reason that indirect descriptor support is guarded by the alloc feature is that the way it's implemented here it requires heap allocation, i.e. a global allocator. In particular, it uses alloc::boxed::Box to allocate the array of indirect descriptors on the heap. If someone wants to use virtio-drivers in an environment where they don't have a heap allocator, then they will lose indirect descriptor support, along with some of the drivers that use heap allocation (currently console, GPU, input and net).

In an environment where heap allocation is available, I don't see any reason why someone would want to disable just indirect descriptor support at compile-time.

[equation314]

Okay, I mean we can enable it by default and make it dependent on alloc