Re: [PATCH v2 2/2] vmalloc: Remove work as from vfree path

[Edgecombe, Rick P]

On Tue, 2019-05-21 at 10:00 -0700, Andy Lutomirski wrote:
> On Tue, May 21, 2019 at 9:51 AM Edgecombe, Rick P
> <rick.p.edgecombe@xxxxxxxxx> wrote:
> > On Tue, 2019-05-21 at 09:17 -0700, Andy Lutomirski wrote:
> > > On Mon, May 20, 2019 at 4:39 PM Rick Edgecombe
> > > <rick.p.edgecombe@xxxxxxxxx> wrote:
> > > > From: Rick Edgecombe <redgecombe.lkml@xxxxxxxxx>
> > > > 
> > > > Calling vm_unmap_alias() in vm_remove_mappings() could
> > > > potentially
> > > > be a
> > > > lot of work to do on a free operation. Simply flushing the TLB
> > > > instead of
> > > > the whole vm_unmap_alias() operation makes the frees faster and
> > > > pushes
> > > > the heavy work to happen on allocation where it would be more
> > > > expected.
> > > > In addition to the extra work, vm_unmap_alias() takes some
> > > > locks
> > > > including
> > > > a long hold of vmap_purge_lock, which will make all other
> > > > VM_FLUSH_RESET_PERMS vfrees wait while the purge operation
> > > > happens.
> > > > 
> > > > Lastly, page_address() can involve locking and lookups on some
> > > > configurations, so skip calling this by exiting out early when
> > > > !CONFIG_ARCH_HAS_SET_DIRECT_MAP.
> > > 
> > > Hmm.  I would have expected that the major cost of
> > > vm_unmap_aliases()
> > > would be the flush, and at least informing the code that the
> > > flush
> > > happened seems valuable.  So would guess that this patch is
> > > actually
> > > a
> > > loss in throughput.
> > > 
> > You are probably right about the flush taking the longest. The
> > original
> > idea of using it was exactly to improve throughput by saving a
> > flush.
> > However with vm_unmap_aliases() the flush will be over a larger
> > range
> > than before for most arch's since it will likley span from the
> > module
> > space to vmalloc. From poking around the sparc tlb flush history, I
> > guess the lazy purges used to be (still are?) a problem for them
> > because it would try to flush each page individually for some CPUs.
> > Not
> > sure about all of the other architectures, but for any
> > implementation
> > like that, using vm_unmap_alias() would turn an occasional long
> > operation into a more frequent one.
> > 
> > On x86, it shouldn't be a problem to use it. We already used to
> > call
> > this function several times around a exec permission vfree.
> > 
> > I guess its a tradeoff that depends on how fast large range TLB
> > flushes
> > usually are compared to small ones. I am ok dropping it, if it
> > doesn't
> > seem worth it.
> 
> On x86, a full flush is probably not much slower than just flushing a
> page or two -- the main cost is in the TLB refill.  I don't know
> about
> other architectures.  I would drop this patch unless you have numbers
> suggesting that it's a win.

Ok. This patch also inadvertently improved some correctness in calls to
flush_tlb_kernel_range() for a rare situation. I'll work that into a
different patch.

Thanks,

Rick