Context Switching

Context switch sequence: Process A running on CPU 0 | Timer interrupt fires (every 4ms, configurable via CONFIG_HZ) | Scheduler decides to run Process B | switch_to(A, B): 1. Save A's registers to A's kernel stack 2. Save A's CR3 (page table pointer) 3. Save A's FPU state (if used FP instructions) 4. Load B's kernel stack pointer 5. Load B's CR3 (new page tables!) - CR3 change flushes the TLB (expensive!) 6. Restore B's FPU state 7. Restore B's registers from B's kernel stack 8. Jump to B's saved RIP (where B was interrupted) | Process B continues running

# PCID (Process-Context Identifiers) — reduce TLB flush cost # Modern CPUs (Sandy Bridge+): TLB entries tagged with PCID # Different processes can coexist in TLB without interfering # CR3 switch doesn't flush all TLB entries, only process's own # Linux uses PCID since kernel 4.14 (if CPU supports it) # Check if PCID is in use: dmesg | grep -i pcid # PCID disabled because TLB_HARDIRQ_CPU is set # or: PCID enabled and supported # Voluntary vs involuntary context switches: cat /proc/1234/status | grep ctxt # voluntary_ctxt_switches: 9876 (process called sleep/wait/IO) # nonvoluntary_ctxt_switches: 234 (preempted by scheduler) # High involuntary = CPU contention (process wants to run but preempted)

# Thread context switch (within same process): # - Save/restore registers: YES # - Change page tables (CR3): NO (same address space) # - TLB flush: NO (same PCID/address space) # - Switch kernel stack: YES (each thread has its own) # Cost: ~1-3 microseconds # Process context switch: # - Save/restore registers: YES # - Change page tables (CR3): YES # - TLB flush: YES (or PCID tag switch) # - Switch kernel stack: YES # Cost: ~3-10 microseconds (cold TLB effects linger much longer) # Why goroutines/green threads are faster: # Go runtime scheduler manages 1000s of goroutines on N OS threads # Goroutine switch = save/restore 7 registers (no syscall, no TLB flush) # OS thread switch needs kernel involvement # Goroutine switches take ~100ns vs 3000ns for OS thread switch

# System-wide context switch rate: vmstat 1 # r b swpd free buff cache si so bi bo in cs us sy id wa # 2 0 0 2048M 512M 8192M 0 0 0 0 512 1234 5 2 93 0 # ^^ cs = context switches/sec # Per-process: pidstat -w 1 # PID cswch/s nvcswch/s Command # 1234 45.0 2.3 nginx # Total context switches since boot: grep ctxt /proc/stat # ctxt 123456789 # Monitor in real time: watch -n1 "grep ctxt /proc/stat" # High context switch rate causes: # 1. Many threads competing for CPU # 2. Frequent I/O (blocking, then unblocking) # 3. Over-threading (more threads than CPUs, each getting tiny slices)