On Wed, Jun 03, 2020 at 09:37:18PM +1000, Michael Ellerman wrote: > Dan Carpenter <dan.carpenter@xxxxxxxxxx> writes: > > On Tue, Jun 02, 2020 at 09:23:57PM +1000, Michael Ellerman wrote: > >> Markus Elfring <Markus.Elfring@xxxxxx> writes: > >> >>>> Please just remove the message instead, it's a tiny allocation that's > >> >>>> unlikely to ever fail, and the caller will print an error anyway. > >> >>> > >> >>> How do you think about to take another look at a previous update suggestion > >> >>> like the following? > >> >>> > >> >>> powerpc/nvram: Delete three error messages for a failed memory allocation > >> >>> https://patchwork.ozlabs.org/project/linuxppc-dev/patch/00845261-8528-d011-d3b8-e9355a231d3a@xxxxxxxxxxxxxxxxxxxxx/ > >> >>> https://lore.kernel.org/linuxppc-dev/00845261-8528-d011-d3b8-e9355a231d3a@xxxxxxxxxxxxxxxxxxxxx/ > >> >>> https://lore.kernel.org/patchwork/patch/752720/ > >> >>> https://lkml.org/lkml/2017/1/19/537 > >> >> > >> >> That deleted the messages from nvram_scan_partitions(), but neither of > >> >> the callers of nvram_scan_paritions() check its return value or print > >> >> anything if it fails. So removing those messages would make those > >> >> failures silent which is not what we want. > >> > > >> > * How do you think about information like the following? > >> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/process/coding-style.rst?id=f359287765c04711ff54fbd11645271d8e5ff763#n883 > >> > “… > >> > These generic allocation functions all emit a stack dump on failure when used > >> > without __GFP_NOWARN so there is no use in emitting an additional failure > >> > message when NULL is returned. > >> > …” > >> > >> Are you sure that's actually true? > >> > >> A quick look around in slub.c leads me to: > >> > >> slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) > >> { > >> #ifdef CONFIG_SLUB_DEBUG > > > > You first have to enable EXPERT mode before you can disable SLUB_DEBUG. > > I see ~175 defconfigs with CONFIG_EXPERT=y, so that's not really a high > bar unfortunately. > > And there's 38 defconfigs with SLUB_DEBUG=n. > > So for kernels built with those defconfigs that documentation is plain > wrong and misleading. > > And then there's SLOB which doesn't dump stack anywhere AFAICS. > > In fact slab_out_of_memory() doesn't emit a stack dump either, it just > prints a bunch of slab related info! > > > So that hopefully means you *really* want to save memory. It doesn't > > make sense to add a bunch of memory wasting printks when the users want > > to go to extra lengths to conserve memory. > > I agree that in many cases those printks are just a waste of space in > the source and the binary and should be removed. > > But I dislike being told "these generic allocation functions all emit a > stack dump" only to find out that actually they don't, they print some > other debug info, and depending on config settings they actually don't > print _anything_. Wait... It *does* print a stack trace. We must but looking at the wrong function. Huh... The stack trace comes from warn_alloc(). What happen is this: mm/slub.c 2673 2674 freelist = new_slab_objects(s, gfpflags, node, &c); 2675 2676 if (unlikely(!freelist)) { 2677 slab_out_of_memory(s, gfpflags, node); 2678 return NULL; 2679 } 2680 The new_slab_objects() will call allocate_slab() which calls __alloc_pages_slowpath() which calls warn_alloc() on failure. There are some error paths from alloc_pages() which look like they could return without the stack dump, but those are impossible paths from kmalloc or error injection. regards, dan carpenter
