Kernel Synchronization

Report
Kernel Synchronization
Examples From the Linux Kernel
Michael E. Locasto
kernel control flow is a complicated, asynchronous interleaving
BIG PICTURE: HOW CAN THE KERNEL
CORRECTLY SERVICE REQUESTS?
Main Ideas / Concepts
Atomic operations in x86
Kernel locking / synchronization primitives
Kernel preemption
Read-Copy-Update
The “big kernel lock”
Kernel Preemption
Kernel preemption is a concept in which the
kernel can preempt other running kernel
control paths (be they on behalf of a user or
another kernel thread)
Acquiring a spinlock automatically disables
kernel preemption (as we will see in the code)
Synchronization Primitives
Atomic operations
Disable interrupts (cli/sti modify IF of eflags)
Lock memory bus (x86 lock prefix)
Spin locks
Semaphores
Sequence Locks
Read-copy-update (RCU) (lock free)
Barriers
Barriers are serializing
operations; they
“gather” and make
operations sequential.
Memory barrier:
x86 in/out on I/O ports
x86 lock prefix
x86 writes to CReg,
SReg/eflags, DReg
x86 instr
meaning
lfence
read barrier
sfence
write barrier
mfence
r/w barrier
Barrier Implementation
Motivating Example: Using
Semaphores in the Kernel
what are: down_read, up_read, and mmap_sem
Let’s start with the data structure and see where that leads…
START WITH THE DATA STRUCTURE:
MM->MMAP_SEM
current->mm->mmap_sem
struct mm_struct: include/linux/mm_types.h
PRIMITIVE ONE: ATOMIC TYPE AND
OPERATIONS
On x86, these operations are atomic
simple asm instructions that involve 0 or 1
aligned memory access
read-modify-update in 1 clock cycle (e.g., inc,
dec)
anything prefixed by the IA-32 ‘lock’ prefix
atomic_t: include/linux/types.h
Example: Reference Counters
Refcounts: atomic_t; associated with resources,
but keeps count of kernel control paths
accessing the resource
PRIMITIVE TWO: SPINLOCKS
/include/linux/spinlock_types.h
typedef struct spinlock{
struct raw_spinlock rlock;
} spinlock_t;
typedef struct raw_spinlock{
arch_spinlock_t raw_lock;
} raw_spinlock_t;
arch/x86/include/asm/spinlock_types.
h#L10
slock=1 (unlocked), slock=0 (locked)
spinlock API (partial)
/include/linux/spinlock.h
/kernel/spinlock.c
include/linux/spinlock_api_smp.h
Linux Tracks Lock Dependencies @
Runtime
Here we mainly consider Read/Write Semaphores
PRIMITIVE THREE: SEMAPHORES
Important Caveats about Kernel
Semaphores
Semaphores are *not* like spinlocks in the
sense that the invoking process is put to sleep
rather than busy waits.
As a result, kernel semaphores should only be
used by functions that can safely sleep (i.e.,
not interrupt handlers)
might_sleep() leads (eventually) to:
rwsem_wake
__rwsem_do_wake
On our way out, allow a writer at the
front of the waiting queue to proceed.
Then allow unlimited numbers of
readers to access the critical region.
Advanced Techniques
Sequence Locks
A solution to the multiple
readers-writer problem in
that a writer is permitted to
advance even if readers are
in the critical section.
Readers must check both an
entry and exit flag to see if
data has been modified
underneath them.
Read-Copy-Update (RCU)
Designed to protect data
structures accessed by
multiple CPUs; allows many
readers and writers.
Basic idea is simple (and in the
name). Readers access data
structure via a pointer; writers
initially act as readers & create
a copy to modify. “Writing” is
just a matter of updating the
pointer.
RCU
Only for kernel control paths; disables
preemption.
Used to protect data structures
accessed through a pointer
by adding a layer of indirection, we can
reduce wholesale writes/updates to a
single atomic write/update
Heavy restrictions: RCU tasks cannot
sleep
readers do little work
writers act as readers, make a copy,
then update copy. Finally, they rewrite
the pointer.
cleanup is correspondingly
complicated.
http://lxr.linux.no/#linux+v2.6.35.14/kernel/timer.c#L1354
RCU EXAMPLE: GETPPID(2)
Does synchronization impose a significant cost?
(test at user level)
EXERCISE: TIME PERFORMANCE
COST OF SYNCHRONIZATION
CODE: AUTOMATICALLY DRAWING
RESOURCE GRAPHS

similar documents