Sorry to self-reply. On 2022/10/12 12:06, fengchengwen wrote: > Hi Dmitry, > > On 2022/10/11 23:58, Dmitry Kozlyuk wrote: >> 2022-10-11 12:17 (UTC+0000), Chengwen Feng: >> [...] >>> diff --git a/doc/guides/prog_guide/memarea_lib.rst >>> b/doc/guides/prog_guide/memarea_lib.rst >>> new file mode 100644 >>> index 0000000000..85ad57145f >>> --- /dev/null >>> +++ b/doc/guides/prog_guide/memarea_lib.rst >>> @@ -0,0 +1,39 @@ >>> +.. SPDX-License-Identifier: BSD-3-Clause >>> + Copyright(c) 2022 HiSilicon Limited >>> + >>> +Memarea Library >>> +=============== >>> + >>> +Introduction >>> +------------ >>> + >>> +The memarea library provides an allocator of variable-size objects, it is >>> +oriented towards the application layer, which could provides 'region-based >>> +memory management' function [1]. >>> + >>> +The main features are as follows: >>> + >>> +* The allocated object aligned at ``RTE_CACHE_LINE_SIZE`` default. >> >> Isn't this an implementation detail? >> Stating it in the API description limits optimization opportunities. >> Cache line alignment is good in many cases, >> but it can also be a waste of space, >> e.g. for a thread-unsafe region for small objects. >> Can this limitation only (temporarily?) apply to user memory? >> Or can the minimal alignment be a property of memarea? > > You are right, to be more general, I'll add the align parameter to > rte_memarea_alloc instead of stating it here.
The memarea is targeted to not so general usage, so v9 doesn't add align parameter to rte_memarea_alloc, at least from the internal feedback. We could add rte_memarea_align_alloc() if the align requirement does exist. > >> >>> + >>> +* The memory region can be initialized from the following memory sources: >>> + a) HEAP: e.g. invoke ``rte_malloc_socket``. b) LIBC: e.g. invoke >>> + posix_memalign to obtain. c) User memory: it can be from e.g. >>> rte_extmem_xxx >>> + as long as it is available. d) Another memarea: it can be allocated from >>> + another memarea. >> >> I think mentioning rte_extmem_xxx() is bogus >> because user memory does not need to be registered with DPDK >> (I understand it's an example, but still an unrelated reference). >> Please format as a list. > > okay > >> >>> + >>> +* It provides refcnt feature which could be useful in multi-reader >>> scenario. >>> + >>> +* It supports MT-safe as long as it's specified at creation time. >>> + >>> +Library API Overview >>> +-------------------- >>> + >>> +The ``rte_memarea_create()`` function is used to create a memarea, the >>> function >>> +returns the pointer to the created memarea or ``NULL`` if the creation >>> failed. >>> + >>> +The ``rte_memarea_destroy()`` function is used to destroy a memarea. >>> + >>> +Reference >>> +--------- >>> + >>> +[1] https://en.wikipedia.org/wiki/Region-based_memory_management >>> diff --git a/doc/guides/rel_notes/release_22_11.rst >>> b/doc/guides/rel_notes/release_22_11.rst >>> index 2da8bc9661..f5a67cec7b 100644 >>> --- a/doc/guides/rel_notes/release_22_11.rst >>> +++ b/doc/guides/rel_notes/release_22_11.rst >>> @@ -63,6 +63,12 @@ New Features >>> In theory this implementation should work with any target based on >>> ``LoongArch`` ISA. >>> >>> +* **Added memarea library.** >>> + >>> + The memarea library is an allocator of variable-size objects, it is >>> oriented >>> + towards the application layer, which could provides 'region-based memory >>> + management' function. >> >> "which could provides" -> "providing" > > okay > >> >>> + >>> * **Added support for multiple mbuf pools per ethdev Rx queue.** >>> >>> The capability allows application to provide many mempools >> [...] >> . >> > > . >
