• Mel Gorman's avatar
    mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages · 072bb0aa
    Mel Gorman authored
    When a user or administrator requires swap for their application, they
    create a swap partition and file, format it with mkswap and activate it
    with swapon.  Swap over the network is considered as an option in diskless
    systems.  The two likely scenarios are when blade servers are used as part
    of a cluster where the form factor or maintenance costs do not allow the
    use of disks and thin clients.
    
    The Linux Terminal Server Project recommends the use of the Network Block
    Device (NBD) for swap according to the manual at
    https://sourceforge.net/projects/ltsp/files/Docs-Admin-Guide/LTSPManual.pdf/download
    There is also documentation and tutorials on how to setup swap over NBD at
    places like https://help.ubuntu.com/community/UbuntuLTSP/EnableNBDSWAP The
    nbd-client also documents the use of NBD as swap.  Despite this, the fact
    is that a machine using NBD for swap can deadlock within minutes if swap
    is used intensively.  This patch series addresses the problem.
    
    The core issue is that network block devices do not use mempools like
    normal block devices do.  As the host cannot control where they receive
    packets from, they cannot reliably work out in advance how much memory
    they might need.  Some years ago, Peter Zijlstra developed a series of
    patches that supported swap over an NFS that at least one distribution is
    carrying within their kernels.  This patch series borrows very heavily
    from Peter's work to support swapping over NBD as a pre-requisite to
    supporting swap-over-NFS.  The bulk of the complexity is concerned with
    preserving memory that is allocated from the PFMEMALLOC reserves for use
    by the network layer which is needed for both NBD and NFS.
    
    Patch 1 adds knowledge of the PFMEMALLOC reserves to SLAB and SLUB to
    	preserve access to pages allocated under low memory situations
    	to callers that are freeing memory.
    
    Patch 2 optimises the SLUB fast path to avoid pfmemalloc checks
    
    Patch 3 introduces __GFP_MEMALLOC to allow access to the PFMEMALLOC
    	reserves without setting PFMEMALLOC.
    
    Patch 4 opens the possibility for softirqs to use PFMEMALLOC reserves
    	for later use by network packet processing.
    
    Patch 5 only sets page->pfmemalloc when ALLOC_NO_WATERMARKS was required
    
    Patch 6 ignores memory policies when ALLOC_NO_WATERMARKS is set.
    
    Patches 7-12 allows network processing to use PFMEMALLOC reserves when
    	the socket has been marked as being used by the VM to clean pages. If
    	packets are received and stored in pages that were allocated under
    	low-memory situations and are unrelated to the VM, the packets
    	are dropped.
    
    	Patch 11 reintroduces __skb_alloc_page which the networking
    	folk may object to but is needed in some cases to propogate
    	pfmemalloc from a newly allocated page to an skb. If there is a
    	strong objection, this patch can be dropped with the impact being
    	that swap-over-network will be slower in some cases but it should
    	not fail.
    
    Patch 13 is a micro-optimisation to avoid a function call in the
    	common case.
    
    Patch 14 tags NBD sockets as being SOCK_MEMALLOC so they can use
    	PFMEMALLOC if necessary.
    
    Patch 15 notes that it is still possible for the PFMEMALLOC reserve
    	to be depleted. To prevent this, direct reclaimers get throttled on
    	a waitqueue if 50% of the PFMEMALLOC reserves are depleted.  It is
    	expected that kswapd and the direct reclaimers already running
    	will clean enough pages for the low watermark to be reached and
    	the throttled processes are woken up.
    
    Patch 16 adds a statistic to track how often processes get throttled
    
    Some basic performance testing was run using kernel builds, netperf on
    loopback for UDP and TCP, hackbench (pipes and sockets), iozone and
    sysbench.  Each of them were expected to use the sl*b allocators
    reasonably heavily but there did not appear to be significant performance
    variances.
    
    For testing swap-over-NBD, a machine was booted with 2G of RAM with a
    swapfile backed by NBD.  8*NUM_CPU processes were started that create
    anonymous memory mappings and read them linearly in a loop.  The total
    size of the mappings were 4*PHYSICAL_MEMORY to use swap heavily under
    memory pressure.
    
    Without the patches and using SLUB, the machine locks up within minutes
    and runs to completion with them applied.  With SLAB, the story is
    different as an unpatched kernel run to completion.  However, the patched
    kernel completed the test 45% faster.
    
    MICRO
                                             3.5.0-rc2 3.5.0-rc2
    					 vanilla     swapnbd
    Unrecognised test vmscan-anon-mmap-write
    MMTests Statistics: duration
    Sys Time Running Test (seconds)             197.80    173.07
    User+Sys Time Running Test (seconds)        206.96    182.03
    Total Elapsed Time (seconds)               3240.70   1762.09
    
    This patch: mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages
    
    Allocations of pages below the min watermark run a risk of the machine
    hanging due to a lack of memory.  To prevent this, only callers who have
    PF_MEMALLOC or TIF_MEMDIE set and are not processing an interrupt are
    allowed to allocate with ALLOC_NO_WATERMARKS.  Once they are allocated to
    a slab though, nothing prevents other callers consuming free objects
    within those slabs.  This patch limits access to slab pages that were
    alloced from the PFMEMALLOC reserves.
    
    When this patch is applied, pages allocated from below the low watermark
    are returned with page->pfmemalloc set and it is up to the caller to
    determine how the page should be protected.  SLAB restricts access to any
    page with page->pfmemalloc set to callers which are known to able to
    access the PFMEMALLOC reserve.  If one is not available, an attempt is
    made to allocate a new page rather than use a reserve.  SLUB is a bit more
    relaxed in that it only records if the current per-CPU page was allocated
    from PFMEMALLOC reserve and uses another partial slab if the caller does
    not have the necessary GFP or process flags.  This was found to be
    sufficient in tests to avoid hangs due to SLUB generally maintaining
    smaller lists than SLAB.
    
    In low-memory conditions it does mean that !PFMEMALLOC allocators can fail
    a slab allocation even though free objects are available because they are
    being preserved for callers that are freeing pages.
    
    [a.p.zijlstra@chello.nl: Original implementation]
    [sebastian@breakpoint.cc: Correct order of page flag clearing]
    Signed-off-by: default avatarMel Gorman <mgorman@suse.de>
    Cc: David Miller <davem@davemloft.net>
    Cc: Neil Brown <neilb@suse.de>
    Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
    Cc: Mike Christie <michaelc@cs.wisc.edu>
    Cc: Eric B Munson <emunson@mgebm.net>
    Cc: Eric Dumazet <eric.dumazet@gmail.com>
    Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc>
    Cc: Mel Gorman <mgorman@suse.de>
    Cc: Christoph Lameter <cl@linux.com>
    Signed-off-by: default avatarAndrew Morton <akpm@linux-foundation.org>
    Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
    072bb0aa
slub.c 130 KB