The fault is a control transfer
A page fault is a processor exception indicating that the current translation or protection state cannot complete a memory access. Instead of silently returning a value, the processor transfers control to an operating-system handler with information about the attempted access. The handler evaluates the address against the process memory layout and the reason for the exception. The important question is whether the access is legitimate and repairable, not merely whether an exception occurred. A successful handler can arrange the mapping and allow execution to resume.
First touch of anonymous memory
A process can reserve a large anonymous virtual range without immediately receiving a distinct physical page for every part of it. On the first write to a particular page, a fault may cause the kernel to allocate a physical frame and install a writable mapping. No file content needs to be read for that operation. In an illustrative 4 KiB-page experiment, touching one byte in each page exposes first-touch costs that repeatedly updating a byte within one already resident page does not reveal.
Interpreting minor and major faults
Linux accounting distinguishes minor faults from major faults using whether resolving the event requires the relevant I/O work. Resolving copy-on-write or mapping content already present in memory can avoid a major storage read. A high fault count alone therefore does not identify a slow disk or prove that memory is exhausted. Record the process, measurement interval, region being accessed, and associated I/O activity. Repeating a workload after warming its pages can change the observed counts, which is part of the experiment rather than evidence that the counter is meaningless.
Not every fault can be repaired
An access outside a valid region, or one that violates its intended protection, cannot always be repaired by allocating another frame. The operating system may instead report an error to the process, for example through a signal. Disabling protection to reduce faults confuses a diagnostic observation with a correctness requirement. Investigate whether the event came from first allocation, file-backed demand paging, deliberate copy-on-write, or an invalid pointer. Those cases can share the exception mechanism while requiring completely different explanations and remedies.
Classify a fault by mapping and permission
Compare first write to an anonymous range, first mapping of a file page already cached, retrieval of a nonresident file page, and an unauthorized write to a read-only region. The first two can be resolved without fetching content from storage. The third may require I/O. The last is not made legitimate merely by adding a page. Copy-on-write is an important qualification: a page-table mapping can deliberately be read-only even though the operating system permits the process to obtain a private writable copy on fault. Mapping-level protection bits alone do not describe the full policy. Consult the valid memory region, intended permissions and reason for the fault before classifying its resolution.