fix: lfc read/write seconds calculation bug fix

[kevin336]

Problem

Bug 1: LFC Write metrics overflow (file_cache_write_wait_seconds_sum)

The INSTR_TIME_SUBTRACT arguments were reversed, causing a negative time value:

// Before (incorrect): io_start - io_end = negative value
INSTR_TIME_SUBTRACT(io_start, io_end);
time_spent_us = INSTR_TIME_GET_MICROSEC(io_start);

When this negative value was interpreted as uint64, it resulted in values close to 2^64 (~18 quintillion), making the metric unusable.

Observed symptom: file_cache_write_wait_seconds_sum showing values like 18446744073709.54 instead of realistic values.

Bug 2: LFC Read metrics always zero (file_cache_read_wait_seconds_sum)

The read path was missing time measurement code entirely. io_time_us was initialized to 0 but never updated after the preadv() call:

uint64 io_time_us = 0; // initialized
// ... preadv() called without timing ...
inc_page_cache_read_wait(io_time_us); // always 0

Observed symptom: file_cache_read_wait_seconds_sum always showing 0 regardless of actual I/O time.

Solution

Fix 1: Correct the argument order for write metrics

// After (correct): io_end - io_start = positive value
INSTR_TIME_SUBTRACT(io_end, io_start);
time_spent_us = INSTR_TIME_GET_MICROSEC(io_end);

This fix is applied in two locations:

• lfc_prefetch() (single block write)
• lfc_writev() (multi-block write)

After fixing the problem I can see below

Fix 2: Add time measurement for read metrics

Added proper instrumentation around preadv():

instr_time read_start, read_end;

INSTR_TIME_SET_CURRENT(read_start);
rc = preadv(lfc_desc, ...);
INSTR_TIME_SET_CURRENT(read_end);
INSTR_TIME_SUBTRACT(read_end, read_start);
io_time_us = INSTR_TIME_GET_MICROSEC(read_end);

After fixing the problem I can see below

Summary of changes

This PR fixes two bugs in the Local File Cache (LFC) I/O latency metrics that caused incorrect values to be reported.