A Locked Read is a special kind of Memory Read that is part of an atomic Read-Modify-Write operation — meaning it ensures that no one else can touch that memory location until the operation is complete.

Example use case:

• Semaphore Test-and-Set – Software wants to check if a semaphore (a memory flag) is 0 and then set it to 1 without anyone else writing to it in between.
• If this is not guaranteed, you could have a race condition where two CPUs or devices both think they successfully locked the semaphore, breaking synchronization.

How PCIe Handles This

Unlike ordinary memory reads, a locked read:

• Is initiated by the CPU only (Root Port generates MRdLk).
• Requires exclusive access to the path from the Root Port to the target memory.
• Prevents other devices from issuing transactions to that target memory while the lock is in effect.

Step-by-Step Flow

1. Root Port Sends a Locked Read Request (MRdLk)

   • The packet includes the memory address and a “lock” indicator.
   • It is routed through switches toward the target memory (or legacy endpoint).

2. Egress Port Locking

   • At each switch or routing device (called a service point), the egress port used by this request is “locked.”
   • This means no other packets are allowed to traverse that path toward the destination until the locked transaction is completed.

3. Target Completer Responds

   • The target gathers the data and returns Locked Completion TLPs back to the Requester.
   • These Completions are marked as “locked,” keeping the path locked until they all return.

4. If an Error Occurs

   • The Completer sends a locked completion without data and sets an error status.
   • The Requester (CPU) understands this means the lock failed.
   • Software decides what to do — retry, abort, or handle the error in some other way.

5. Unlocking

   • Once the locked Completion(s) are delivered, the path is unlocked so that normal traffic can resume.

Legacy Support

• Locked transactions were originally designed because PCI was meant to be a replacement for CPU front-side buses — so CPU-like atomic operations were supported.
• In modern PCIe:
  • New devices are not allowed to accept locked transactions unless they declare themselves as Legacy Endpoints.
  • This is partly to discourage reliance on this old mechanism and encourage newer atomic operations (PCIe added native atomic ops in later versions).

Key Takeaways

• MRdLk = Memory Read Locked.
• CPU-only feature (Root Port initiates it).
• Locks the data path (no other traffic can pass through that egress port) → ensures atomicity.
• Rarely used in modern systems, kept mainly for backward compatibility.
• PCIe introduced native atomic operations (like AtomicOp FetchAdd, Swap, CAS) which are better alternatives and don’t require locking whole paths.