lkml.org 
[lkml]   [2015]   [Oct]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    SubjectRe: [Intel-wired-lan] [Patch V3 5/9] i40e: Use numa_mem_id() to better support memoryless node
    From
    Date
    On 2015/10/09 14:52, Jiang Liu wrote:
    > On 2015/10/9 4:20, Andrew Morton wrote:
    >> On Wed, 19 Aug 2015 17:18:15 -0700 (PDT) David Rientjes <rientjes@google.com> wrote:
    >>
    >>> On Wed, 19 Aug 2015, Patil, Kiran wrote:
    >>>
    >>>> Acked-by: Kiran Patil <kiran.patil@intel.com>
    >>>
    >>> Where's the call to preempt_disable() to prevent kernels with preemption
    >>> from making numa_node_id() invalid during this iteration?
    >>
    >> David asked this question twice, received no answer and now the patch
    >> is in the maintainer tree, destined for mainline.
    >>
    >> If I was asked this question I would respond
    >>
    >> The use of numa_mem_id() is racy and best-effort. If the unlikely
    >> race occurs, the memory allocation will occur on the wrong node, the
    >> overall result being very slightly suboptimal performance. The
    >> existing use of numa_node_id() suffers from the same issue.
    >>
    >> But I'm not the person proposing the patch. Please don't just ignore
    >> reviewer comments!
    > Hi Andrew,
    > Apologize for the slow response due to personal reasons!
    > And thanks for answering the question from David. To be honest,
    > I didn't know how to answer this question before. Actually this
    > question has puzzled me for a long time when dealing with memory
    > hot-removal. For normal cases, it only causes sub-optimal memory
    > allocation if schedule event happens between querying NUMA node id
    > and calling alloc_pages_node(). But what happens if system run into
    > following execution sequence?
    > 1) node = numa_mem_id();
    > 2) memory hot-removal event triggers
    > 2.1) remove affected memory
    > 2.2) reset pgdat to zero if node becomes empty after memory removal

    I'm sorry if I misunderstand something.
    After commit b0dc3a342af36f95a68fe229b8f0f73552c5ca08, there is no memset().

    > 3) alloc_pages_node(), which may access zero-ed pgdat structure.

    ?

    >
    > I haven't found a mechanism to protect system from above sequence yet,
    > so puzzled for a long time already:(. Does stop_machine() protect
    > system from such a execution sequence?

    To access pgdat, a pgdat's zone should be on per-pgdat-zonelist.
    Now, __build_all_zonelists() is called under stop_machine(). That's the reason
    why you're asking what stop_machine() does. And, as you know, stop_machine() is not
    protecting anything. The caller may fallback into removed zone.

    Then, let's think.

    At first, please note "pgdat" is not removed (and cannot be removed),
    accessing pgdat's memory will not cause segmentation fault.

    Just contents are problem. At removal, zone's page related information
    and pgdat's page related information is cleared.

    alloc_pages uses zonelist/zoneref/cache to walk each zones without accessing
    pgdat itself. I think accessing zonelist is safe because it's an array updated
    by stop_machine().

    So, the problem is alloc_pages() can work correctly even if zone contains no page.
    I think it should work.

    (Note: zones are included in pgdat. So, zeroing pgdat means zeroing zone and other
    structures. it will not work.)

    So, what problem you see now ?
    I'm sorry I can't chase old discusions.

    Thanks,
    -Kame

















    \
     
     \ /
      Last update: 2015-10-09 11:21    [W:3.157 / U:0.704 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site