Message-ID: <e678affb-5eee-a055-7af1-1d29a965663b@google.com>
Date:   Tue, 4 Jul 2023 10:03:57 -0700 (PDT)
From:   Hugh Dickins <hughd@...gle.com>
To:     Gerald Schaefer <gerald.schaefer@...ux.ibm.com>
cc:     Hugh Dickins <hughd@...gle.com>, Jason Gunthorpe <jgg@...pe.ca>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Vasily Gorbik <gor@...ux.ibm.com>,
        Mike Kravetz <mike.kravetz@...cle.com>,
        Mike Rapoport <rppt@...nel.org>,
        "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>,
        Matthew Wilcox <willy@...radead.org>,
        David Hildenbrand <david@...hat.com>,
        Suren Baghdasaryan <surenb@...gle.com>,
        Qi Zheng <zhengqi.arch@...edance.com>,
        Yang Shi <shy828301@...il.com>,
        Mel Gorman <mgorman@...hsingularity.net>,
        Peter Xu <peterx@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Will Deacon <will@...nel.org>, Yu Zhao <yuzhao@...gle.com>,
        Alistair Popple <apopple@...dia.com>,
        Ralph Campbell <rcampbell@...dia.com>,
        Ira Weiny <ira.weiny@...el.com>,
        Steven Price <steven.price@....com>,
        SeongJae Park <sj@...nel.org>,
        Lorenzo Stoakes <lstoakes@...il.com>,
        Huang Ying <ying.huang@...el.com>,
        Naoya Horiguchi <naoya.horiguchi@....com>,
        Christophe Leroy <christophe.leroy@...roup.eu>,
        Zack Rusin <zackr@...are.com>,
        Axel Rasmussen <axelrasmussen@...gle.com>,
        Anshuman Khandual <anshuman.khandual@....com>,
        Pasha Tatashin <pasha.tatashin@...een.com>,
        Miaohe Lin <linmiaohe@...wei.com>,
        Minchan Kim <minchan@...nel.org>,
        Christoph Hellwig <hch@...radead.org>,
        Song Liu <song@...nel.org>,
        Thomas Hellstrom <thomas.hellstrom@...ux.intel.com>,
        Russell King <linux@...linux.org.uk>,
        "David S. Miller" <davem@...emloft.net>,
        Michael Ellerman <mpe@...erman.id.au>,
        "Aneesh Kumar K.V" <aneesh.kumar@...ux.ibm.com>,
        Heiko Carstens <hca@...ux.ibm.com>,
        Christian Borntraeger <borntraeger@...ux.ibm.com>,
        Claudio Imbrenda <imbrenda@...ux.ibm.com>,
        Alexander Gordeev <agordeev@...ux.ibm.com>,
        Jann Horn <jannh@...gle.com>,
        Vishal Moola <vishal.moola@...il.com>,
        Vlastimil Babka <vbabka@...e.cz>,
        linux-arm-kernel@...ts.infradead.org, sparclinux@...r.kernel.org,
        linuxppc-dev@...ts.ozlabs.org, linux-s390@...r.kernel.org,
        linux-kernel@...r.kernel.org, linux-mm@...ck.org
Subject: Re: [PATCH v2 07/12] s390: add pte_free_defer() for pgtables sharing
 page

On Tue, 4 Jul 2023, Gerald Schaefer wrote:
> On Sat, 1 Jul 2023 21:32:38 -0700 (PDT)
> Hugh Dickins <hughd@...gle.com> wrote:
> > On Thu, 29 Jun 2023, Hugh Dickins wrote:
> > > 
> > > I've grown to dislike the (ab)use of pt_frag_refcount even more, to the
> > > extent that I've not even tried to verify it; but I think I do get the
> > > point now, that we need further info than just PPHHAA to know whether
> > > the page is on the list or not.  But I think that if we move where the
> > > call_rcu() is done, then the page can stay on or off the list by same
> > > rules as before (but need to check HH bits along with PP when deciding
> > > whether to allocate, and whether to list_add_tail() when freeing).  
> > 
> > No, not quite the same rules as before: I came to realize that using
> > list_add_tail() for the HH pages would be liable to put a page on the
> > list which forever blocked reuse of PP list_add_tail() pages after it
> > (could be solved by a list_move() somewhere, but we have agreed to
> > prefer simplicity).
> > 
> > I've dropped the HH bits, I'm using PageActive like we did on powerpc,
> > I've dropped most of the pte_free_*() helpers, and list_del_init() is
> > an easier way of dealing with those "is it on the list" questions.
> > I expect that we shall be close to reaching agreement on...
> 
> This looks really nice, almost too good and easy to be true. I did not
> find any obvious flaw, just some comments below. It also survived LTP
> without any visible havoc, so I guess this approach is the best so far.

Phew! I'm of course glad to hear this: thanks for your efforts on it.

...
> > --- a/arch/s390/mm/pgalloc.c
> > +++ b/arch/s390/mm/pgalloc.c
> > @@ -229,6 +229,15 @@ void page_table_free_pgste(struct page *page)
> >   * logic described above. Both AA bits are set to 1 to denote a 4KB-pgtable
> >   * while the PP bits are never used, nor such a page is added to or removed
> >   * from mm_context_t::pgtable_list.
> > + *
> > + * pte_free_defer() overrides those rules: it takes the page off pgtable_list,
> > + * and prevents both 2K fragments from being reused. pte_free_defer() has to
> > + * guarantee that its pgtable cannot be reused before the RCU grace period
> > + * has elapsed (which page_table_free_rcu() does not actually guarantee).
> 
> Hmm, I think page_table_free_rcu() has to guarantee the same, i.e. not
> allow reuse before grace period elapsed. And I hope that it does so, by
> setting the PP bits, which would be noticed in page_table_alloc(), in
> case the page would be seen there.
> 
> Unlike pte_free_defer(), page_table_free_rcu() would add pages back to the
> end of the list, and so they could be seen in page_table_alloc(), but they
> should not be reused before grace period elapsed and __tlb_remove_table()
> cleared the PP bits, as far as I understand.
> 
> So what exactly do you mean with "which page_table_free_rcu() does not actually
> guarantee"?

I'll answer without locating and re-reading what Jason explained earlier,
perhaps in a separate thread, about pseudo-RCU-ness in tlb_remove_table():
he may have explained it better.  And without working out again all the
MMU_GATHER #defines, and which of them do and do not apply to s390 here.

The detail that sticks in my mind is the fallback in tlb_remove_table()
in mm/mmu_gather.c: if its __get_free_page(GFP_NOWAIT) fails, it cannot
batch the tables for freeing by RCU, and resorts instead to an immediate 
TLB flush (I think: that again involves chasing definitions) followed by
tlb_remove_table_sync_one() - which just delivers an interrupt to each CPU,
and is commented: 
/*
 * This isn't an RCU grace period and hence the page-tables cannot be
 * assumed to be actually RCU-freed.
 *
 * It is however sufficient for software page-table walkers that rely on
 * IRQ disabling.
 */

Whether that's good for your PP pages or not, I've given no thought:
I've just taken it on trust that what s390 has working today is good.

If that __get_free_page(GFP_NOWAIT) fallback instead used call_rcu(),
then I would not have written "(which page_table_free_rcu() does not
actually guarantee)".  But it cannot use call_rcu() because it does
not have an rcu_head to work with - it's in some generic code, and
there is no MMU_GATHER_CAN_USE_PAGE_RCU_HEAD for architectures to set.

And Jason would have much preferred us to address the issue from that
angle; but not only would doing so destroy my sanity, I'd also destroy
20 architectures TLB-flushing, unbuilt and untested, in the attempt.

...
> > @@ -325,10 +346,17 @@ void page_table_free(struct mm_struct *mm, unsigned long *table)
> >  		 */
> >  		mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
> >  		mask >>= 24;
> > -		if (mask & 0x03U)
> > +		if ((mask & 0x03U) && !PageActive(page)) {
> > +			/*
> > +			 * Other half is allocated, and neither half has had
> > +			 * its free deferred: add page to head of list, to make
> > +			 * this freed half available for immediate reuse.
> > +			 */
> >  			list_add(&page->lru, &mm->context.pgtable_list);
> > -		else
> > -			list_del(&page->lru);
> > +		} else {
> > +			/* If page is on list, now remove it. */
> > +			list_del_init(&page->lru);
> > +		}
> 
> Ok, we might end up with some unnecessary list_del_init() here, e.g. if
> other half is still allocated, when called from pte_free_defer() on a
> fully allocated page, which was not on the list (and with PageActive, and
> (mask & 0x03U) true).
> Not sure if adding an additional mask check to the else path would be
> needed, but it seems that list_del_init() should also be able to handle
> this.

list_del_init() is very cheap in the unnecessary case: the cachelines
required are already there.  You don't want a flag to say whether to
call it or not, it is already the efficient approach.

(But you were right not to use it in your pt_frag_refcount version,
because there we were still trying to do the call_rcu() per fragment
rather than per page, so page->lru could have been on the RCU queue.)

> 
> Same thought applies to the similar logic in page_table_free_rcu()
> below.
> 
> >  		spin_unlock_bh(&mm->context.lock);
> >  		mask = atomic_xor_bits(&page->_refcount, 0x10U << (bit + 24));
> >  		mask >>= 24;
> > @@ -342,8 +370,10 @@ void page_table_free(struct mm_struct *mm, unsigned long *table)
> >  	}
> >  
> >  	page_table_release_check(page, table, half, mask);
> > -	pgtable_pte_page_dtor(page);
> > -	__free_page(page);
> > +	if (TestClearPageActive(page))
> > +		call_rcu(&page->rcu_head, pte_free_now);
> > +	else
> > +		pte_free_now(&page->rcu_head);
> 
> This ClearPageActive, and the similar thing in __tlb_remove_table() below,
> worries me a bit, because it is done outside the spin_lock. It "feels" like
> there could be some race with the PageActive checks inside the spin_lock,
> but when drawing some pictures, I could not find any such scenario yet.
> Also, our existing spin_lock is probably not supposed to protect against
> PageActive changes anyway, right?

Here (and similarly in __tlb_remove_table()) is where we are about to free
the page table page: both of the fragments have already been released,
there is nobody left who could be racing against us to set PageActive.

I chose PageActive for its name, not for any special behaviour of that
flag: nothing else could be setting or clearing it while we own the page.

Hugh

Powered by blists - more mailing lists