RE: [PATCH v2 2/6] eal: adapt EAL to present rte optional atomics API

["Morten Brørup" <mb@smartsharesystems.com>]

> From: Tyler Retzlaff [mailto:roretzla@linux.microsoft.com]
> Sent: Friday, 11 August 2023 19.32
> 
> Adapt the EAL public headers to use rte optional atomics API instead of
> directly using and exposing toolchain specific atomic builtin intrinsics.
> 
> Signed-off-by: Tyler Retzlaff <roretzla@linux.microsoft.com>
> ---

[...]

> --- a/app/test/test_mcslock.c
> +++ b/app/test/test_mcslock.c
> @@ -36,9 +36,9 @@
>   *   lock multiple times.
>   */
> 
> -rte_mcslock_t *p_ml;
> -rte_mcslock_t *p_ml_try;
> -rte_mcslock_t *p_ml_perf;
> +rte_mcslock_t * __rte_atomic p_ml;
> +rte_mcslock_t * __rte_atomic p_ml_try;
> +rte_mcslock_t * __rte_atomic p_ml_perf;

Although this looks weird, it is pointers themselves, not the structures, that are used atomically. So it is correct.

> diff --git a/lib/eal/include/generic/rte_pause.h
> b/lib/eal/include/generic/rte_pause.h
> index bebfa95..c816e7d 100644
> --- a/lib/eal/include/generic/rte_pause.h
> +++ b/lib/eal/include/generic/rte_pause.h
> @@ -36,13 +36,13 @@
>   *  A 16-bit expected value to be in the memory location.
>   * @param memorder
>   *  Two different memory orders that can be specified:
> - *  __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to
> + *  rte_memory_order_acquire and rte_memory_order_relaxed. These map to
>   *  C++11 memory orders with the same names, see the C++11 standard or
>   *  the GCC wiki on atomic synchronization for detailed definition.

Delete the last part of the description, starting at "These map to...".

>   */
>  static __rte_always_inline void
>  rte_wait_until_equal_16(volatile uint16_t *addr, uint16_t expected,
> -		int memorder);
> +		rte_memory_order memorder);
> 
>  /**
>   * Wait for *addr to be updated with a 32-bit expected value, with a relaxed
> @@ -54,13 +54,13 @@
>   *  A 32-bit expected value to be in the memory location.
>   * @param memorder
>   *  Two different memory orders that can be specified:
> - *  __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to
> + *  rte_memory_order_acquire and rte_memory_order_relaxed. These map to
>   *  C++11 memory orders with the same names, see the C++11 standard or
>   *  the GCC wiki on atomic synchronization for detailed definition.

Delete the last part of the description, starting at "These map to...".

>   */
>  static __rte_always_inline void
>  rte_wait_until_equal_32(volatile uint32_t *addr, uint32_t expected,
> -		int memorder);
> +		rte_memory_order memorder);
> 
>  /**
>   * Wait for *addr to be updated with a 64-bit expected value, with a relaxed
> @@ -72,42 +72,42 @@
>   *  A 64-bit expected value to be in the memory location.
>   * @param memorder
>   *  Two different memory orders that can be specified:
> - *  __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to
> + *  rte_memory_order_acquire and rte_memory_order_relaxed. These map to
>   *  C++11 memory orders with the same names, see the C++11 standard or
>   *  the GCC wiki on atomic synchronization for detailed definition.

Delete the last part of the description, starting at "These map to...".

>   */
>  static __rte_always_inline void
>  rte_wait_until_equal_64(volatile uint64_t *addr, uint64_t expected,
> -		int memorder);
> +		rte_memory_order memorder);

[...]

> @@ -125,16 +125,16 @@
>   *  An expected value to be in the memory location.
>   * @param memorder
>   *  Two different memory orders that can be specified:
> - *  __ATOMIC_ACQUIRE and __ATOMIC_RELAXED. These map to
> + *  rte_memory_order_acquire and rte_memory_order_relaxed. These map to
>   *  C++11 memory orders with the same names, see the C++11 standard or
>   *  the GCC wiki on atomic synchronization for detailed definition.

Delete the last part of the description, starting at "These map to...".

There might be more similar comments that need removal; I haven't tried searching.

>   */
>  #define RTE_WAIT_UNTIL_MASKED(addr, mask, cond, expected, memorder) do { \

[...]

> --- a/lib/eal/include/generic/rte_spinlock.h
> +++ b/lib/eal/include/generic/rte_spinlock.h
> @@ -29,7 +29,7 @@
>   * The rte_spinlock_t type.
>   */
>  typedef struct __rte_lockable {
> -	volatile int locked; /**< lock status 0 = unlocked, 1 = locked */
> +	volatile int __rte_atomic locked; /**< lock status 0 = unlocked, 1 =
> locked */

I think __rte_atomic should be before the type:
	volatile __rte_atomic int locked; /**< lock status [...]
Alternatively (just mentioning it, I know we don't use this form):
	volatile __rte_atomic(int) locked; /**< lock status [...]

Thinking of where you would put "const" might help.

Maybe your order is also correct, so it is a matter of preference.

The DPDK coding style guidelines doesn't mention where to place "const", but looking at the code, it seems to use "const unsigned int" and "const char *".

>  } rte_spinlock_t;
> 
>  /**

[...]

> --- a/lib/eal/include/rte_mcslock.h
> +++ b/lib/eal/include/rte_mcslock.h
> @@ -33,8 +33,8 @@
>   * The rte_mcslock_t type.
>   */
>  typedef struct rte_mcslock {
> -	struct rte_mcslock *next;
> -	int locked; /* 1 if the queue locked, 0 otherwise */
> +	struct rte_mcslock * __rte_atomic next;

Correct, the pointer is atomic, not the struct.

> +	int __rte_atomic locked; /* 1 if the queue locked, 0 otherwise */

Again, I think __rte_atomic should be before the type:
	__rte_atomic int locked; /* 1 if the queue locked, 0 otherwise */

>  } rte_mcslock_t;
> 

[...]

> @@ -101,34 +101,34 @@
>   *   A pointer to the node of MCS lock passed in rte_mcslock_lock.
>   */
>  static inline void
> -rte_mcslock_unlock(rte_mcslock_t **msl, rte_mcslock_t *me)
> +rte_mcslock_unlock(rte_mcslock_t * __rte_atomic *msl, rte_mcslock_t *
> __rte_atomic me)
>  {
>  	/* Check if there are more nodes in the queue. */
> -	if (likely(__atomic_load_n(&me->next, __ATOMIC_RELAXED) == NULL)) {
> +	if (likely(rte_atomic_load_explicit(&me->next, rte_memory_order_relaxed)
> == NULL)) {
>  		/* No, last member in the queue. */
> -		rte_mcslock_t *save_me = __atomic_load_n(&me, __ATOMIC_RELAXED);
> +		rte_mcslock_t *save_me = rte_atomic_load_explicit(&me,
> rte_memory_order_relaxed);
> 
>  		/* Release the lock by setting it to NULL */
> -		if (likely(__atomic_compare_exchange_n(msl, &save_me, NULL, 0,
> -				__ATOMIC_RELEASE, __ATOMIC_RELAXED)))
> +		if (likely(rte_atomic_compare_exchange_strong_explicit(msl,
> &save_me, NULL,
> +				rte_memory_order_release,
> rte_memory_order_relaxed)))
>  			return;
> 
>  		/* Speculative execution would be allowed to read in the
>  		 * while-loop first. This has the potential to cause a
>  		 * deadlock. Need a load barrier.
>  		 */
> -		__atomic_thread_fence(__ATOMIC_ACQUIRE);
> +		__rte_atomic_thread_fence(rte_memory_order_acquire);
>  		/* More nodes added to the queue by other CPUs.
>  		 * Wait until the next pointer is set.
>  		 */
> -		uintptr_t *next;
> -		next = (uintptr_t *)&me->next;
> +		uintptr_t __rte_atomic *next;
> +		next = (uintptr_t __rte_atomic *)&me->next;

This way around, I think:
		__rte_atomic uintptr_t *next;
		next = (__rte_atomic uintptr_t *)&me->next;

[...]

> --- a/lib/eal/include/rte_pflock.h
> +++ b/lib/eal/include/rte_pflock.h
> @@ -41,8 +41,8 @@
>   */
>  struct rte_pflock {
>  	struct {
> -		uint16_t in;
> -		uint16_t out;
> +		uint16_t __rte_atomic in;
> +		uint16_t __rte_atomic out;

Again, I think __rte_atomic should be before the type:
		__rte_atomic uint16_t in;
		__rte_atomic uint16_t out;

>  	} rd, wr;
>  };

[...]

> --- a/lib/eal/include/rte_seqcount.h
> +++ b/lib/eal/include/rte_seqcount.h
> @@ -32,7 +32,7 @@
>   * The RTE seqcount type.
>   */
>  typedef struct {
> -	uint32_t sn; /**< A sequence number for the protected data. */
> +	uint32_t __rte_atomic sn; /**< A sequence number for the protected data.
> */

Again, I think __rte_atomic should be before the type:
	__rte_atomic uint32_t sn; /**< A sequence [...]

>  } rte_seqcount_t;

[...]

> --- a/lib/eal/include/rte_ticketlock.h
> +++ b/lib/eal/include/rte_ticketlock.h
> @@ -30,10 +30,10 @@
>   * The rte_ticketlock_t type.
>   */
>  typedef union {
> -	uint32_t tickets;
> +	uint32_t __rte_atomic tickets;
>  	struct {
> -		uint16_t current;
> -		uint16_t next;
> +		uint16_t __rte_atomic current;
> +		uint16_t __rte_atomic next;

Again, I think __rte_atomic should be before the type:
		__rte_atomic uint16_t current;
		__rte_atomic uint16_t next;

>  	} s;
>  } rte_ticketlock_t;



> @@ -127,7 +129,7 @@
> 
>  typedef struct {
>  	rte_ticketlock_t tl; /**< the actual ticketlock */
> -	int user; /**< core id using lock, TICKET_LOCK_INVALID_ID for unused */
> +	int __rte_atomic user; /**< core id using lock, TICKET_LOCK_INVALID_ID
> for unused */

Again, I think __rte_atomic should be before the type:
	__rte_atomic int user; /**< core id [...]

>  	unsigned int count; /**< count of time this lock has been called */
>  } rte_ticketlock_recursive_t;

[...]

> --- a/lib/eal/include/rte_trace_point.h
> +++ b/lib/eal/include/rte_trace_point.h
> @@ -33,7 +33,7 @@
>  #include <rte_stdatomic.h>
> 
>  /** The tracepoint object. */
> -typedef uint64_t rte_trace_point_t;
> +typedef uint64_t __rte_atomic rte_trace_point_t;

Again, I think __rte_atomic should be before the type:
typedef __rte_atomic uint64_t rte_trace_point_t;

[...]

At the risk of having gone "speed blind" by all the search-replaces along the way...

Reviewed-by: Morten Brørup <mb@smartsharesystems.com>