On Fri, Apr 19, 2024 at 08:54:40AM -0700, Song Liu wrote: > On Thu, Apr 18, 2024 at 11:56 PM Mike Rapoport <rppt@xxxxxxxxxx> wrote: > > > > On Thu, Apr 18, 2024 at 02:01:22PM -0700, Song Liu wrote: > > > On Thu, Apr 18, 2024 at 10:54 AM Mike Rapoport <rppt@xxxxxxxxxx> wrote: > > > > > > > > On Thu, Apr 18, 2024 at 09:13:27AM -0700, Song Liu wrote: > > > > > On Thu, Apr 18, 2024 at 8:37 AM Mike Rapoport <rppt@xxxxxxxxxx> wrote: > > > > > > > > > > > > > > > > I'm looking at execmem_types more as definition of the consumers, maybe I > > > > > > > > should have named the enum execmem_consumer at the first place. > > > > > > > > > > > > > > I think looking at execmem_type from consumers' point of view adds > > > > > > > unnecessary complexity. IIUC, for most (if not all) archs, ftrace, kprobe, > > > > > > > and bpf (and maybe also module text) all have the same requirements. > > > > > > > Did I miss something? > > > > > > > > > > > > It's enough to have one architecture with different constrains for kprobes > > > > > > and bpf to warrant a type for each. > > > > > > > > > > AFAICT, some of these constraints can be changed without too much work. > > > > > > > > But why? > > > > I honestly don't understand what are you trying to optimize here. A few > > > > lines of initialization in execmem_info? > > > > > > IIUC, having separate EXECMEM_BPF and EXECMEM_KPROBE makes it > > > harder for bpf and kprobe to share the same ROX page. In many use cases, > > > a 2MiB page (assuming x86_64) is enough for all BPF, kprobe, ftrace, and > > > module text. It is not efficient if we have to allocate separate pages for each > > > of these use cases. If this is not a problem, the current approach works. > > > > The caching of large ROX pages does not need to be per type. > > > > In the POC I've posted for caching of large ROX pages on x86 [1], the cache is > > global and to make kprobes and bpf use it it's enough to set a flag in > > execmem_info. > > > > [1] https://lore.kernel.org/all/20240411160526.2093408-1-rppt@xxxxxxxxxx > > For the ROX to work, we need different users (module text, kprobe, etc.) to have > the same execmem_range. From [1]: > > static void *execmem_cache_alloc(struct execmem_range *range, size_t size) > { > ... > p = __execmem_cache_alloc(size); > if (p) > return p; > err = execmem_cache_populate(range, size); > ... > } > > We are calling __execmem_cache_alloc() without range. For this to work, > we can only call execmem_cache_alloc() with one execmem_range. Actually, on x86 this will "just work" because everything shares the same address space :) The 2M pages in the cache will be in the modules space, so __execmem_cache_alloc() will always return memory from that address space. For other architectures this indeed needs to be fixed with passing the range to __execmem_cache_alloc() and limiting search in the cache for that range. > Did I miss something? > > Thanks, > Song -- Sincerely yours, Mike.
