[dpdk-dev] Windows Support Plan
Dmitry Kozliuk
dmitry.kozliuk at gmail.com
Sun Feb 2 21:37:36 CET 2020
Hi everyone!
Where do I find a high-level plan of comprehensive Windows support: design
decisions, implementation order, etc?
Information on the subject is very scarce, one may think it is abandoned.
Googling for "site:dpdk.org/ml/archives/dev/ windows" yields only two pages
of disjoint messages. I learned about "netuio" days ago from a tiny remark in
a "Minutes of Technical Board Meetings" email, and even then it took
enumerating "dpdk-next-windows" branches to find the source.
The matter is, as a New Year's holiday project of mine I implemented Windows
support from scratch to the point it runs in QEMU with virtio-pci [0]. It is
not of production quality, cuts some corners and lacks major features (see
bottom). My primary goal was fun^W making it work. Comparing it to
"windpdk-v18.08" branch of "dpdk-next-windows", I can see that 1) our
implementations take rather different approaches in some cases, and 2) both
have severe issues and would benefit from amalgamation. I'd like to
contribute to Windows support with this code, but to do so, coordination is
required, because changes are significant.
Primary topics to discuss:
1. Memory management (@Anatoly)
1.1. MM changed radically since v18.08 and dpdk-next-windows does not
implement it properly anyway, it allocates segment lists in a PCI bus
driver. My implementation closely follows the Linux one using
VirtualAlloc2() with XXX_PLACEHOLDER flags to reserve and commit
memory, but does not map hugepages to files. Is there
a consensus on MM approach in Windows?
Anyway, I think EAL private MM API would have to be changed,
because memory reservation, allocation, and mapping are
completely different operations. Hiding this with an mmap() shim
doesn't look right, because mmap()'s behavior differs even among
Unix platforms.
1.2. In Windows, there is no /dev/mem to implement rte_virt2iova(),
so a simple kernel driver is required for mapping. Moreover,
Windows kernel abstracts IOMMU, so those physical addresses may
be unsuitable for DMA at all (see below).
2. Kernel drivers (@Harini, @Stephen)
2.1. The most serious issue is that Windows formally prohibits using
arbitrary physical addresses with DMA in favor of allocating
special buffers (presumably because IOMMU may be engaged, and
there is no way to check). We can either live with it
(technically, everything works with PA mode), or we could revive
DMA allocation API from ethdev to ask the driver for a proper
DMA buffer.
2.2. Neither netuio, nor my driver (userpci) support interrupts.
I see not inherent difficulty here, but interface should be
designed carefully.
2.3. Windows allows mapping I/O ports into user-space, but there is
no API to change IOPL, which makes mapping useless and requires
a syscall for every I/O port access. This demolishes
virtio-legacy performance. Perhaps Microsoft could give some
advice here. OTOH, PIO is all legacy, so might be much effort is
not justified.
2.4. I believe GUIDs approach for identifying compatible devices
should be strictly preferred, and not DosDevices symlinks. Think
of Mellanox OFED on Linux, which uses a different driver, but
could provide a compatible interface. Another reason is that
a single driver can implement multiple kernel interfaces with
appropriate GUIDs.
2.5. DPDK Windows driver guidelines, driver review, and certification.
The quality of both netuio and userpci is below standards now
(e. g. netuio does not mind its context when mapping memory,
and userpci lacks synchronization), code style is a mix of
Windows and DPDK, logging may be insufficient.
3. POSIX shim vs EAL wrappers (@Thomas, @Pallavi, @Ranjit)
What is the policy: to implement a POSIX shim in EAL (as the latest
patches from Pallavi Kadam do), or to add dependencies (as [1] suggests)?
IMO creating a shim is wrong. First, some POSIX concepts do not
easily map to Windows, like poll() interface and I/O model in
general. Second, there are numerous getopt, pthread, etc.
implementations for Windows, no point wasting resources and repeat
them, adding bugs. I can think of two exceptions:
* <sys/queue.h>, which is header-only.
* Berkeley sockets. Adding <winsock2.h> to public headers creates
more trouble that its worth: definitions for a few structures and
constants. May be there should be some <rte_socket.h> to abstract
platform differences.
Some highlights on my implementation:
* Major features NOT supported:
* multi-process (due to limited time)
* interrupts (limited time + explained above)
* eventdev (requires access to physical memory)
* hot-plug (due to limited time and Windows knowledge)
* bbdev (see comments in config/common_windows)
* FreeBSD (trivial, I just don't use it)
* DPDK is built using MinGW-w64 with GNUmake or Meson.
Drivers are built using DDK (msbuild or Visual Studio).
Actually, I cross-compile DPDK and build drivers natively.
* Only tested on Windows 10 in QEMU with virtio-legacy.
* No docs, but there's nothing unusual for those familiar with Windows.
Bind virt2phys driver to Root\virt2phys, bind userpci driver to device(s).
* Commit history is squashed, because it was a mess from experiments.
There also may be some leftover changes, but those commits are not proper
patches anyway.
References:
[0]: https://github.com/PlushBeaver/dpdk/commits/windows
[1]: http://mails.dpdk.org/archives/dev/2015-February/014245.html
--
Dmitry Kozlyuk
More information about the dev
mailing list