Re: [PATCH v13 1/7] eal: add static per-lcore memory allocation facility

["Mattias Rönnblom" <hofors@lysator.liu.se>]

On 2024-10-16 00:33, Stephen Hemminger wrote:
> On Tue, 15 Oct 2024 11:33:38 +0200
> Mattias Rönnblom <mattias.ronnblom@ericsson.com> wrote:
> 
>> + * Lcore variables
>> + *
>> + * This API provides a mechanism to create and access per-lcore id
>> + * variables in a space- and cycle-efficient manner.
>> + *
>> + * A per-lcore id variable (or lcore variable for short) holds a
>> + * unique value for each EAL thread and registered non-EAL
>> + * thread. There is one instance for each current and future lcore
>> + * id-equipped thread, with a total of @c RTE_MAX_LCORE instances. The
>> + * value of the lcore variable for one lcore id is independent from
>> + * the values assigned to other lcore ids within the same variable.
>> + *
>> + * In order to access the values of an lcore variable, a handle is
>> + * used. The type of the handle is a pointer to the value's type
>> + * (e.g., for an @c uint32_t lcore variable, the handle is a
>> + * <code>uint32_t *</code>). The handle type is used to inform the
>> + * access macros of the type of the values. A handle may be passed
>> + * between modules and threads just like any pointer, but its value
>> + * must be treated as an opaque identifier. An allocated handle never
>> + * has the value NULL.
>> + *
>> + * @b Creation
>> + *
>> + * An lcore variable is created in two steps:
>> + *  1. Define an lcore variable handle by using @ref RTE_LCORE_VAR_HANDLE.
>> + *  2. Allocate lcore variable storage and initialize the handle with
>> + *     a unique identifier by @ref RTE_LCORE_VAR_ALLOC or
>> + *     @ref RTE_LCORE_VAR_INIT. Allocation generally occurs at the time
>> + *     of module initialization, but may be done at any time.
>> + *
>> + * The lifetime of an lcore variable is not tied to the thread that
>> + * created it. Its per lcore id values (up to @c RTE_MAX_LCORE) are
>> + * available from the moment the lcore variable is created and
>> + * continue to exist throughout the entire lifetime of the EAL,
>> + * whether or not the lcore id is currently in use.
>> + *
>> + * Lcore variables cannot and need not be freed.
>> + *
>> + * @b Access
>> + *
>> + * The value of any lcore variable for any lcore id may be accessed
>> + * from any thread (including unregistered threads), but it should
>> + * only be *frequently* read from or written to by the owner.
>> + *
>> + * Values of the same lcore variable, associated with different lcore
>> + * ids may be frequently read or written by their respective owners
>> + * without risking false sharing.
>> + *
>> + * An appropriate synchronization mechanism (e.g., atomic loads and
>> + * stores) should be employed to prevent data races between the owning
>> + * thread and any other thread accessing the same value instance.
>> + *
>> + * The value of the lcore variable for a particular lcore id is
>> + * accessed using @ref RTE_LCORE_VAR_LCORE.
>> + *
>> + * A common pattern is for an EAL thread or a registered non-EAL
>> + * thread to access its own lcore variable value. For this purpose, a
>> + * shorthand exists as @ref RTE_LCORE_VAR.
>> + *
>> + * Although the handle (as defined by @ref RTE_LCORE_VAR_HANDLE) is a
>> + * pointer with the same type as the value, it may not be directly
>> + * dereferenced and must be treated as an opaque identifier.
>> + *
>> + * Lcore variable handles and value pointers may be freely passed
>> + * between different threads.
>> + *
>> + * @b Storage
>> + *
>> + * An lcore variable's values may be of a primitive type like @c int,
>> + * but would more typically be a @c struct.
>> + *
>> + * The lcore variable handle introduces a per-variable (not
>> + * per-value/per-lcore id) overhead of @c sizeof(void *) bytes, so
>> + * there are some memory footprint gains to be made by organizing all
>> + * per-lcore id data for a particular module as one lcore variable
>> + * (e.g., as a struct).
>> + *
>> + * An application may define an lcore variable handle without ever
>> + * allocating it.
>> + *
>> + * The size of an lcore variable's value must be less than the DPDK
>> + * build-time constant @c RTE_MAX_LCORE_VAR.
>> + *
>> + * Lcore variables are stored in a series of lcore buffers, which are
>> + * allocated from the libc heap. Heap allocation failures are treated
>> + * as fatal.
>> + *
>> + * Lcore variables should generally *not* be @ref __rte_cache_aligned
>> + * and need *not* include a @ref RTE_CACHE_GUARD field, since the use
>> + * of these constructs are designed to avoid false sharing. In the
>> + * case of an lcore variable instance, the thread most recently
>> + * accessing nearby data structures should almost-always be the lcore
>> + * variable's owner. Adding padding will increase the effective memory
>> + * working set size, potentially reducing performance.
>> + *
>> + * Lcore variable values are initialized to zero by default.
>> + *
>> + * Lcore variables are not stored in huge page memory.
>> + *
>> + * @b Example
>> + *
>> + * Below is an example of the use of an lcore variable:
>> + *
>> + * @code{.c}
>> + * struct foo_lcore_state {
>> + *         int a;
>> + *         long b;
>> + * };
>> + *
>> + * static RTE_LCORE_VAR_HANDLE(struct foo_lcore_state, lcore_states);
>> + *
>> + * long foo_get_a_plus_b(void)
>> + * {
>> + *         struct foo_lcore_state *state = RTE_LCORE_VAR(lcore_states);
>> + *
>> + *         return state->a + state->b;
>> + * }
>> + *
>> + * RTE_INIT(rte_foo_init)
>> + * {
>> + *         RTE_LCORE_VAR_ALLOC(lcore_states);
>> + *
>> + *         unsigned int lcore_id;
>> + *         struct foo_lcore_state *state;
>> + *         RTE_LCORE_VAR_FOREACH(lcore_id, state, lcore_states) {
>> + *                 (initialize 'state')
>> + *         }
>> + *
>> + *         (other initialization)
>> + * }
>> + * @endcode
>> + *
>> + *
>> + * @b Alternatives
>> + *
>> + * Lcore variables are designed to replace a pattern exemplified below:
>> + * @code{.c}
>> + * struct __rte_cache_aligned foo_lcore_state {
>> + *         int a;
>> + *         long b;
>> + *         RTE_CACHE_GUARD;
>> + * };
>> + *
>> + * static struct foo_lcore_state lcore_states[RTE_MAX_LCORE];
>> + * @endcode
>> + *
>> + * This scheme is simple and effective, but has one drawback: the data
>> + * is organized so that objects related to all lcores for a particular
>> + * module are kept close in memory. At a bare minimum, this requires
>> + * sizing data structures (e.g., using `__rte_cache_aligned`) to an
>> + * even number of cache lines to avoid false sharing. With CPU
>> + * hardware prefetching and memory loads resulting from speculative
>> + * execution (functions which seemingly are getting more eager faster
>> + * than they are getting more intelligent), one or more "guard" cache
>> + * lines may be required to separate one lcore's data from another's
>> + * and prevent false sharing.
>> + *
>> + * Lcore variables offer the advantage of working with, rather than
>> + * against, the CPU's assumptions. A next-line hardware prefetcher,
>> + * for example, may function as intended (i.e., to the benefit, not
>> + * detriment, of system performance).
>> + *
>> + * Another alternative to @ref rte_lcore_var.h is the @ref
>> + * rte_per_lcore.h API, which makes use of thread-local storage (TLS,
>> + * e.g., GCC __thread or C11 _Thread_local). The main differences
>> + * between by using the various forms of TLS (e.g., @ref
>> + * RTE_DEFINE_PER_LCORE or _Thread_local) and the use of lcore
>> + * variables are:
>> + *
>> + *   * The lifecycle of a thread-local variable instance is tied to
>> + *     that of the thread. The data cannot be accessed before the
>> + *     thread has been created, nor after it has exited. As a result,
>> + *     thread-local variables must be initialized in a "lazy" manner
>> + *     (e.g., at the point of thread creation). Lcore variables may be
>> + *     accessed immediately after having been allocated (which may occur
>> + *     before any thread beyond the main thread is running).
>> + *   * A thread-local variable is duplicated across all threads in the
>> + *     process, including unregistered non-EAL threads (i.e.,
>> + *     "regular" threads). For DPDK applications heavily relying on
>> + *     multi-threading (in conjunction to DPDK's "one thread per core"
>> + *     pattern), either by having many concurrent threads or
>> + *     creating/destroying threads at a high rate, an excessive use of
>> + *     thread-local variables may cause inefficiencies (e.g.,
>> + *     increased thread creation overhead due to thread-local storage
>> + *     initialization or increased total RAM footprint usage). Lcore
>> + *     variables *only* exist for threads with an lcore id.
>> + *   * If data in thread-local storage may be shared between threads
>> + *     (i.e., can a pointer to a thread-local variable be passed to
>> + *     and successfully dereferenced by non-owning thread) depends on
>> + *     the specifics of the TLS implementation. With GCC __thread and
>> + *     GCC _Thread_local, data sharing between threads is supported.
>> + *     In the C11 standard, accessing another thread's _Thread_local
>> + *     object is implementation-defined. Lcore variable instances may
>> + *     be accessed reliably by any thread.
>> + */
> 
> For me this comment too wordy for code and belongs in the documentation instead.
> Could also be reduced to more precise succinct language.
> 
> 

Provided this makes it into RC1, I can move most of this and some of the 
information in eal_common_lcore_var.c comments into "the documentation" 
as a RC2 patch.

If "the documentation" is a the EAL programmer's guide, a description of 
lcore variables (with pictures!) in sufficient detail (both API and 
implementation) would make up a large fraction of it. That would look 
silly and in the way of more important things. Lcore variables is just a 
tiny bit of infrastructure. Other, more central EAL features, like the 
RTE spinlock, they have no mention at all in the EAL docs.

Another option I suppose is to documentation it separately from the 
"main" EAL programmer's guide, but - correct me if I'm wrong here - 
there seem to be no precedent for doing this.