Hello, I am working on a bpf program of type perf_event (thus, it can't be sleepable). This program needs to be able to unwind the user-mode stack of the current task, which of course involves reading stack memory. In most cases, this works fine with bpf_probe_read_user. However, in some cases, it causes page faults, even when the relevant pages are resident in memory (and thus accessing them shouldn't cause a sleep). One example of when this can happen is when running on a NUMA system. If the process is migrated to a core on a different NUMA node, eventually Linux's numa balancing feature will kick in and the task's stack pages will be migrated to the proper node. When this happens, there will be a (minor) page fault the next time the process accesses the page so that the mapping in its virtual address space can be updated. But if a page is _never_ accessed again by the process, then bpf_probe_read_user for addresses in that page will always fail. Is there any workaround for cases like this, or is it simply impossible? One possible direction could be to allow bpf_probe_read_user to incur _minor_ page faults, but not _major_ ones, but I don't know whether that would actually be possible. -- Brennan (he/him) Staff Software Engineer | Polar Signals Inc. P: +16238249252 E: brennan.vincent@xxxxxxxxxxxxxxxx W: umanwizard.com TZ: America/New_York
