2021-03-01 21:43, Nick Connolly: > > Complete removal of non-standard dependencies in headers is within a grasp. > > Then we can remove shims and include whatever needed. > > Thoughts? > Sounds good. A couple of 'gotchas' that I've come across (but may not be > an issue for DPDK): > > * Memory allocation / free that spans a dll boundary (see earlier email).
Not sure which email you mean, could you give a link? > * posix_memalign is unfortunately specified as using free(), so we'll > either have to modify the calls to rte_posix_memalign / > rte_free_memalign (which means code analysis to find which free > calls need changing), or wrapper free() to distinguish the two types > of memory somehow. I 'solved' this for SPDK by using posix_memalign > for everything and I have a vague recollection that the mingw libc > does something similar. Is posix_memalign() used more extensively in SPDK? In DPDK, it's 2 PMDs: * mlx5 PMD uses _aligned_malloc/_aligned_free on Windows; * dpaax (PMD family) uses posix_memalign() to allocate pages. There are "malloc" and "alloc_size" attributes that can help code analysis. > * Sockets are unfortunately specified as using close(). This is > probably easy to address by rte_ wrapping all socket calls. Which public DPDK APIs operate on sockets? I don't like the idea of wrapping APIs like sockets or files. (Yes, we're discussing libpcap API wrappers in this thread, but we already agreed they are a mistake and they were internal in the first place.) I drafted what I was talking about: adding address types and removing shims: * librte_net/rte_ip.h then includes <netinet/ip.h> or <ws2tcpip.h> conditionally for AF_xxx, IPPROTO_xxx, and a few other constants. That's probably OK, there are similar places for Linux/FreeBSD differences, e.g. in <rte_endian.h>. * Some IPPROTO_xxx constants are missing on Windows, so rte_ip.h has to provide them. I hope Mirosoft will add them to system headers one day. * It affects cmdline (mostly), ethdev, security, crypto/dpaax.
