On Tue, Aug 04, 2020 at 07:34:18PM +0200, Vlastimil Babka wrote: > On 8/4/20 7:12 PM, Matthew Wilcox wrote: > > On Tue, Aug 04, 2020 at 07:02:14PM +0200, Vlastimil Babka wrote: > >> > 2) There was a proposal from Matthew Wilcox: https://lkml.org/lkml/2020/7/31/1015 > >> > > >> > <snip> > >> > On non-RT, we could make that lock a raw spinlock. On RT, we could > >> > decline to take the lock. We'd need to abstract the spin_lock() away > >> > behind zone_lock(zone), but that should be OK. > >> > <snip> > >> > > >> > It would be great to use any existing flag, say GFP_NOWAIT. Suppose we > >> > decline to take the lock across the page allocator for RT. But there is > >> > at least one path that does it outside of the page allocator. GFP_NOWAIT > >> > can wakeup the kswapd, whereas a "wake-up path" uses sleepable lock: > >> > > >> > wakeup_kswapd() -> wake_up_interruptible(&pgdat->kswapd_wait). > >> > > >> > Probably it can be fixed by the excluding of waking of the kswapd process > >> > defining something like below: > >> > >> Is something missing here? > >> > >> > what is equal to zero and i am not sure if __get_free_page(0) handles > >> > all that correctly, though it allocates and seems working on my test > >> > machine! Please note it is related to "if we can reuse existing flags". > >> > > >> > In the meantime, please see below for a patch that adds a __GFP_FAST_TRY, > >> > which can at least serve as a baseline against which other proposals can > >> > be compared. The patch is based on the 5.8.0-rc3. > >> > > >> > Please RFC. > >> > >> At first glance __GFP_FAST_TRY (more descriptive name? __GFP_NO_LOCKS?) seems > >> better than doing weird things with GFP_NOWAIT, but depends on the real benefits > >> (hence my first questions). > > > > I think what Vlad is trying to say is that even GFP_NOWAIT will wake > > kswapd, which involves taking a spinlock. If you specify 0 in your GFP > > flags, then we won't wake kswapd. So a simple: > > > > #define GFP_NOLOCKS 0 > > > > should do the trick (modulo various casting, blah blah blah) > > Ah, you're right, waking up kswapd is is only done with __GFP_KSWAPD_RECLAIM and > GFP_NOWAIT equals to that. So that's easy to avoid for the rcu allocation. > > But still IIUC option 2) would mean that even with "#define GFP_NOLOCKS 0" would > mean we need to abstract away the zone lock, and behave differently depending on > the kernel being RT, and inadvertedly changing other users that happen to > specify gfp where "gfp & GFP_RECLAIM_MASK == 0" (or however we would exactly > check if we can take the lock on RT kernel). That sounds too complicated to me. > I think a different behaviour, i mean RT/non-rt, is not a way forward, because the things will be over complicated. Please note, the proposed variant is common. It provides a fast access to pcp-cache, what can be done lock-less. If we could extend the "fast path" even do the lock-less prefetch(make fast path fully lock-less) from the body would be fantastic, but that is a bit out of the question. For example implement removing/inserting pages from "zone->free_area" as lock-less: llist_add()/llist_del(). But that is theory and on the high level. During investigation the things might become complicated. -- Vlad Rezki
