From a library request to a system-call result

A system-call ABI specifies how an architecture represents the requested operation, arguments, results, and errors at the user-kernel boundary. Registers are common carriers, but their assignments and exceptional details differ across architectures and compatibility modes. A C library wrapper usually handles those details for its target. Its calling convention is not necessarily the same as the kernel's calling convention. Copying a register sequence from an unrelated machine can produce a different request or invalid arguments even when the source-level function name appears familiar.

Entry machinery establishes the kernel-side execution state before dispatch reaches an implementation. The operation number must be interpreted within the appropriate ABI, and unsupported numbers require a defined failure path. Policy mechanisms can also affect whether an operation is allowed or observed. Implementations must validate lengths, access rights, and object state; a user-supplied pointer is not a trusted kernel pointer. Appropriate access helpers handle faults, and a check does not freeze user memory that another thread can change. Copying or synchronization must match the interface's actual ownership contract. These are reasons to learn the boundary as a contract rather than memorize a particular internal function name. Kernel organization and mitigation settings evolve; a diagram of one implementation version is useful evidence for that version, not an invariant guaranteeing the same sequence or latency on every Linux installation.

For a read-like byte interface, a positive result reports the number of bytes actually obtained, which can be smaller than the requested count without indicating failure. A zero result for an ordinary nonzero-length read can indicate end of input, depending on the object and interface. A negative wrapper return accompanied by an error indication describes another outcome. Write code around the documented object-specific semantics instead of assuming that a successful request always fills the entire application buffer. A stream parser may need to retain incomplete application messages across several successful reads.

An interrupted operation requires equally careful treatment. An error such as EINTR can mean interruption before progress for a particular call, while an operation that already transferred bytes may report that positive partial count. Restart rules depend on the interface and signal configuration. Advance the buffer position only by confirmed progress, preserve an application deadline across retries, and stop according to the protocol's completion condition. A retry loop that blindly resets its timeout can exceed the caller's latency budget even if each individual attempt observes a locally reasonable limit.

Reducing calls can amortize fixed overhead, but larger batches may need more buffering and can delay the first record while the batch fills. A batching interface can also have different atomicity, ordering, completion, and failure behavior. Compare completed application work under equivalent conditions. For an inference service, the relevant outcome may be replies meeting a deadline, not merely fewer entries into the kernel. Measure both CPU service and elapsed latency, document the batch policy, and retain the actual partial-completion logic after introducing a supposedly faster data path.

The queue lab offers an overload thought experiment rather than a system-call implementation. Compare separate runs with arrival rate 2 and service rate 1 versus arrival rate 2 and service rate 3. Before each run, press Reset, re-set both desired rates, predict whether the fresh four-job backlog grows or drains, then press Play or Step. Changing a control alone does not begin a new run or preserve the previous backlog as a new starting point. A real batching improvement could raise effective service capacity, but the control does not compute that improvement from an ABI or a batch size. The separate residence-time readout uses a stable M/M/1 model with Poisson arrivals and exponential service, which batching can violate. Use the lab to form a capacity question, then determine the actual answer with a workload and a measured completion contract.