On Thu, 2 Nov 2023 16:58:52 +0000 Bruce Richardson <bruce.richard...@intel.com> wrote:
> On Thu, Nov 02, 2023 at 05:28:42PM +0100, Thomas Monjalon wrote: > > 02/11/2023 15:56, Bruce Richardson: > > > Hi all, > > > > > > looking to start a discussion and get some input here. > > > > > > There are a number of our examples in DPDK which still track core usage > > > via > > > a 64-bit bitmask, and, as such, cannot run on cores between 64 and > > > RTE_MAX_LCORE. Two examples I have recently come across with this issue > > > are > > > "eventdev_pipeline" and "qos_sched", but I am sure there are others. The > > > former is a good example (or bad example depending on your viewpoint) of > > > this as it takes multiple coremask parameters - for RX cores, for TX > > > cores, > > > for worker cores and optionally for scheduler cores. > > > > > > Now, the simple solution to this is to just expand the 64-bit bitmask to > > > 128 bit or more, but I think that is just making things harder for the > > > user, since dealing with long bitmasks is very awkward and unwieldy. > > > Better > > > instead to convert all examples using coremasks to using core lists > > > instead. > > > > > > First step should be to take our EAL corelist processing code and refactor > > > it into a function that can be made public, so that it can be used by all > > > apps for parsing core lists. Simple enough! > > > > OK to add some command line parsing helpers. > > It should probably be the start of a new library for command line. > > > > Funnily enough, separate to this I had already been working on an > "rte_args" library to have some functions for working on argc/argv > parameters. I'm planning on pushing out an RFC for 24.03 fairly shortly. > > However, pulling in functions for arg parsing is a different set of > functionality to what I had in mind, so we may yet get two libraries out of > this. [Merging may be tricky due to issues around circular dependencies > with EAL. My arg management library is designed to "sit on top of" EAL, > while any lib offering e.g. coremask, corelist parsing functions would need > to "sit beneath" EAL, so EAL can re-use it's functions]. > > Let's postpone the details of both these to when we get some RFCs out > though. > > > > > The next part I'm looking for input on is - how do we switch the apps from > > > coremasks to core lists? Some options: > > > > > > 1. Add in new commandline parameters for each app to work with core lists. > > > This is what we did in the past with EAL, by adding -l as a replacement > > > for -c. The advantage is that we maintain backward compatibility, but > > > the > > > downside is that it becomes hard to find new suitable letter options for > > > the core lists. Taking eventdev_pipeline again, we would need "new" > > > options for "-r", "-t", "-w" and "-s" parameters. Using the capitalized > > > versions of these would be a simple alternative, but "-W" is already > > > used > > > as an app parameter so we can't do that. > > > > > > 2. Just break backward compatibility and switch the apps to taking > > > core lists instead of masks. Advantage is that it gives us the cleanest > > > solution, but the downside is that and testing done using these > > > examples, > > > or any users who may have run them in the past, get different > > > behaviour. > > > > We don't need to offer backward compatibility in examples. > > So this option is acceptable. > > > > Glad to hear it. Makes life simpler. > > > > 3. An interesting further alternative is to allow apps to take both > > > coremasks and corelists and use heuristics to determine which is which. > > > For example, anything starting with "0x" is a mask, anything containing > > > "-" or "," is a list. There would be ambiguous values such as e.g. 2, > > > which could be either, but we can always find ways to disambiguate > > > these, e.g. allow trailing commas in lists, so that "0x2" is the > > > coremask, and "2," is the corelist. [Could be other alternatives]. This > > > largely keeps backward compatibility and also allows use of corelists. > > > > The option 3 can be interesting as well. > > > Yep. If we start offering a library of arg-parsing functions, one of those > could be a function using heuristics to identify core-mask, core-list or > ambiguous values. Then each app can decide what to do in the latter case. > Since we don't care about backward compatibility, the examples can just parse > ambiguous values as core-list. User could then still use coremasks by > prefixing them with "0x". Noticed that lib/graph is using 64 bit coremask internally. Wonder if others have the same issue. Would be good if DPDK had a library to handle cpusets better, something like the Linux kernel cpuset which uses comma separated list of cpu masks.
