linux-kernel - [PATCH 00/11] x86/mm: Implement lockless pgd_alloc()/pgd

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Message-Id: <1442903021-3893-1-git-send-email-mingo@kernel.org>
Date:	Tue, 22 Sep 2015 08:23:30 +0200
From:	Ingo Molnar <mingo@...nel.org>
To:	linux-kernel@...r.kernel.org, linux-mm@...ck.org
Cc:	Andy Lutomirski <luto@...capital.net>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Denys Vlasenko <dvlasenk@...hat.com>,
	Brian Gerst <brgerst@...il.com>,
	Peter Zijlstra <peterz@...radead.org>,
	Borislav Petkov <bp@...en8.de>,
	"H. Peter Anvin" <hpa@...or.com>,
	Linus Torvalds <torvalds@...ux-foundation.org>,
	Oleg Nesterov <oleg@...hat.com>,
	Waiman Long <waiman.long@...com>,
	Thomas Gleixner <tglx@...utronix.de>
Subject: [PATCH 00/11] x86/mm: Implement lockless pgd_alloc()/pgd_free()

So this is the somewhat belated latest iteration of the series.
I (think I) fixed all correctness bugs in the code pointed out by Oleg.

The task list walk is still 'dumb', using for_each_process(), as none of
the call sites are performance critical.

Oleg, can you see any problems with this code?

Background:

Waiman Long reported 'pgd_lock' contention on high CPU count systems and proposed
moving pgd_lock on a separate cacheline to eliminate false sharing and to reduce
some of the lock bouncing overhead.

I think we can do much better: this series eliminates the pgd_list and makes
pgd_alloc()/pgd_free() lockless.

Now the lockless initialization of the PGD has a few preconditions, which the
initial part of the series implements:

 - no PGD clearing is allowed, only additions. This makes sense as a single PGD
   entry covers 512 GB of RAM so the 4K overhead per 0.5TB of RAM mapped is
   miniscule.

The patches after that convert existing pgd_list users to walk the task list.

PGD locking is kept intact: coherency guarantees between the CPA, vmalloc,
hotplug, etc. code are unchanged.

The final patches eliminate the pgd_list and thus make pgd_alloc()/pgd_free()
lockless.

The patches have been boot tested on 64-bit and 32-bit x86 systems.

Architectures not making use of the new facility are unaffected.

Thanks,

	Ingo

===
Ingo Molnar (11):
  x86/mm/pat: Don't free PGD entries on memory unmap
  x86/mm/hotplug: Remove pgd_list use from the memory hotplug code
  x86/mm/hotplug: Don't remove PGD entries in remove_pagetable()
  x86/mm/hotplug: Simplify sync_global_pgds()
  mm: Introduce arch_pgd_init_late()
  x86/virt/guest/xen: Remove use of pgd_list from the Xen guest code
  x86/mm: Remove pgd_list use from vmalloc_sync_all()
  x86/mm/pat/32: Remove pgd_list use from the PAT code
  x86/mm: Make pgd_alloc()/pgd_free() lockless
  x86/mm: Remove pgd_list leftovers
  x86/mm: Simplify pgd_alloc()

 arch/Kconfig                      |   9 +++
 arch/x86/Kconfig                  |   1 +
 arch/x86/include/asm/pgtable.h    |   3 -
 arch/x86/include/asm/pgtable_64.h |   3 +-
 arch/x86/mm/fault.c               |  32 +++++++---
 arch/x86/mm/init_64.c             |  92 ++++++++++++--------------
 arch/x86/mm/pageattr.c            |  40 ++++++------
 arch/x86/mm/pgtable.c             | 131 +++++++++++++++++++-------------------
 arch/x86/xen/mmu.c                |  45 +++++++++++--
 fs/exec.c                         |   3 +
 include/linux/mm.h                |   6 ++
 kernel/fork.c                     |  16 +++++
 12 files changed, 227 insertions(+), 154 deletions(-)

--
2.1.4

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/