Skip to content

Latest commit

 

History

History
56 lines (36 loc) · 2.13 KB

AHCILocking.md

File metadata and controls

56 lines (36 loc) · 2.13 KB

AHCI Locking

Introduction to hard locks, soft locks and what they do

Soft lock - Lock

A soft lock is basically a regular lock in the kernel. We use it with a Locker class, to create a scoped locking of that lock:

Locker locker(m_lock);

...
...

return true;

This lock doesn't disable interrupts at all, and if it is already in use, the scheduler will simply yield away from that section until it tries to lock it again.

Hard lock - Spinlock

A hard lock is essentially a lock that is used in critical sections in the kernel. We use it with a ScopedSpinLock class, to create a scoped locking of that lock:

ScopedSpinLock lock(m_lock);

...
...

return true;

Why do we need soft and hard locking in the AHCI code?

First of all, the proper way of taking a SpinLock and Lock is to:

Locker locker(m_soft_lock);
ScopedSpinLock lock(m_spinlock);

...
...

return true;

This sequence is relevant for any pattern of taking a soft and hard lock together in the kernel. The reason for this order is that SpinLock will disable interrupts, while Lock will still allow the system to yield execution to another thread if we can't lock the soft lock, because interrupts are not disabled. Taking a SpinLock and then a Lock is considered a bug, because we already disabled interrupts so yielding from this section is not possible anymore.

We need both types of locking to implement hardware access safely. When we use the SpinLock object, we ensure that only one CPU can run the scoped code section without any interruptions at all. This is important, because interrupts can be fatal in essentially what is a critical section.

We use the Lock object for basically anything else, most of the time together with SpinLock as described earlier. This object becomes important when we schedule IO work to happen in the IO WorkQueue. When we run in WorkQueue, it is guaranteed that we will have interrupts enabled - therefore we will not use the SpinLock to allow the kernel to handle page fault interrupts, but we still want to ensure no other concurrent operation can happen, so we still hold the Lock.