Systems

A mindmap of OS-related notes.

• Interupts
  • Synchronous
    • Called Exceptions
      • Processor detected
        • Faults (before execution of instruction)
        • Traps (after execution of instruction)
        • aborts
      • Programmed
        • int d
  • Asynchronous
    • Interrupts from external events, like I/O devices
  • Can be controlled at:
    • Device level
    • APIC level (Advanced Programmable Interrupt Controller)
    • CPU level
      • cli (Clear interrupt flag)
      • sti (set interrupt flag)
  • Some can be masked, some not (NMI)
  • Interrupt Descriptor Table
  • When an interrupt happens
    • CPU checks privilege level
    • If needed, change to require privilege
    • Save stack info
    • Execute handler
  • Return with IRET
  • Top half, until IRET
    • IRQ disabled for this duration
    • Called the interrupt context
    • Can be preempted by one interrupt, not exception
    • Not allowed to context switch, sleep, schedule or access memory
  • Bottom half, deferred execution
    • Two categories
      • Interrupt context deferrables
      • Process context deferrables
    • Reason: avoid running too long in interupt handler with irq disabled
    • Three types
      • SoftIRQs
        • Interrupt context
        • Statically allocated
        • Same handler can run in parallel on many cores
      • Tasklet
        • Inerrup context
        • Dynamically allocated
        • Same handlers run serialized
      • Workqueues
        • Run in process context
• System calls
  • Flow
    • Set call number and parameters in registers
    • Issue a trap
    • Enter kernel context
    • Kernel saves registers in stack
    • Identify syscall from syscall table and run it
    • Restore registers
  • Max 6 parameters (set in registers)
    • Retrieve result
  • VDSO
    • Optimize system call implementation by vsyscalls that run on user space
    • Access VDSO page that is either static or modified by kernel
    • E.g. getpid or gettimeofday
• Booting
  • BIOS / POST
    • verify hardware integrity
      • cpu registers, bios code, dma, interrupt controllers
      • memory, bios init
      • identify and select devices for booting
    • Written in EEPROM
  • Bios reads MBR
    • MBR is 512 bytes, in the first sector of the selected boot disk
    • Contains program and partiotion table details
      • Last 2 bytes: 0xAA55, magic number
      • Contains first stage loader, loads second stage
  • Stage 1 boot loader
  • Stage 2 boot loader
    • Calls the kernel loader
  • Kernel stage
    • Kernel loaded in memory
    • Image file containing basic root FS with kernel modules is loaded in memory. This is initrd.
    • After kernel detectes hardware, the root file system on disk will take over the one from memory
  • INIT (systemd)
    • Read /etc/fsstab
  • UEFI is next level
    • Proves > 2TB partition size, has GUID Partition Table (GPT) instead of MBR
    • Comes with its own FS (/EFI)
      • Separate folder per system
      • Systems provide their own bootloader
• VFS
  • read(), write(), open(), close(), stat(), lseek()
  • inodes
  • dentry
  • file
  • mounting, superblock operations
• Limits
  • CGroups
    • /sys/fs/cgroups API
    • Applied to tasks
    • Has controllers
      • The core, pseudo-fs at /sys/fs/cgroup
      • Other controlles
        • Memory
        • IO
        • CPU
  • Namespaces
    • Used with unshare, nsenter
    • Mount
    • PID
    • Network
    • UTS (Hostname)
    • IPC
    • Cgroup
    • Implemented as flags in clone() syscall
    • Also unshare()
  • Capabilities
    • Set per thread, also to files
    • Finer grained control that setuid-as-root
    • CAP_IPC_LOCK
    • CAP_MKNOD
    • CAP_NET_RAW
    • CAP_SYS_ADMIN
    • shown in /proc/[pid]/status
  • setrlimits()
  • tc (qdisk - queue discipline)
    • Overlap with Cgroups
  • nice
  • seccomp - limit system calls
• Network
  • XDP and BPF
    • Process RX before any sk_buffs are allocated or queues entered
    • Not a Kernel bypass, integrated in the path
    • actions
      • Forward, drop, receive
    • PF_RING or AF_PACKET, in contrast, bypasses the kernel
  • Multiple receive queues in NICs
    • Receive side scaling
      • Remember to set IRQ affinity to all cpus
  • RPS, software implementation of RSS
  • RPS doesn’t take into account applicatino locality
    • RFS, use flow lookup table to steer packets to CPU where the flow is processed
  • XPS, Transmit packet steering