lkml.org 
[lkml]   [2019]   [Jan]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCHv5] x86/kdump: bugfix, make the behavior of crashkernel=X consistent with kaslr
On Tue, Jan 8, 2019 at 11:49 PM Mike Rapoport <rppt@linux.ibm.com> wrote:
>
> On Tue, Jan 08, 2019 at 05:01:38PM +0800, Baoquan He wrote:
> > Hi Mike,
> >
> > On 01/08/19 at 10:05am, Mike Rapoport wrote:
> > > I'm not thrilled by duplicating this code (yet again).
> > > I liked the v3 of this patch [1] more, assuming we allow bottom-up mode to
> > > allocate [0, kernel_start) unconditionally.
> > > I'd just replace you first patch in v3 [2] with something like:
> >
> > In initmem_init(), we will restore the top-down allocation style anyway.
> > While reserve_crashkernel() is called after initmem_init(), it's not
> > appropriate to adjust memblock_find_in_range_node(), and we really want
> > to find region bottom up for crashkernel reservation, no matter where
> > kernel is loaded, better call __memblock_find_range_bottom_up().
> >
> > Create a wrapper to do the necessary handling, then call
> > __memblock_find_range_bottom_up() directly, looks better.
>
> What bothers me is 'the necessary handling' which is already done in
> several places in memblock in a similar, but yet slightly different way.
>
> memblock_find_in_range() and memblock_phys_alloc_nid() retry with different
> MEMBLOCK_MIRROR, but memblock_phys_alloc_try_nid() does that only when
> allocating from the specified node and does not retry when it falls back to
> any node. And memblock_alloc_internal() has yet another set of fallbacks.
>
> So what should be the necessary handling in the wrapper for
> __memblock_find_range_bottom_up() ?
>
Well, it is a hard choice.
> BTW, even without any memblock modifications, retrying allocation in
> reserve_crashkerenel() for different ranges, like the proposal at [1] would
> also work, wouldn't it?
>
Yes, it can work. Then is it worth to expose the bottom-up allocation
style beside for hotmovable purpose?

Thanks,
Pingfan
> [1] http://lists.infradead.org/pipermail/kexec/2017-October/019571.html
>
> > Thanks
> > Baoquan
> >
> > >
> > > diff --git a/mm/memblock.c b/mm/memblock.c
> > > index 7df468c..d1b30b9 100644
> > > --- a/mm/memblock.c
> > > +++ b/mm/memblock.c
> > > @@ -274,24 +274,14 @@ phys_addr_t __init_memblock memblock_find_in_range_node(phys_addr_t size,
> > > * try bottom-up allocation only when bottom-up mode
> > > * is set and @end is above the kernel image.
> > > */
> > > - if (memblock_bottom_up() && end > kernel_end) {
> > > - phys_addr_t bottom_up_start;
> > > -
> > > - /* make sure we will allocate above the kernel */
> > > - bottom_up_start = max(start, kernel_end);
> > > -
> > > + if (memblock_bottom_up()) {
> > > /* ok, try bottom-up allocation first */
> > > - ret = __memblock_find_range_bottom_up(bottom_up_start, end,
> > > + ret = __memblock_find_range_bottom_up(start, end,
> > > size, align, nid, flags);
> > > if (ret)
> > > return ret;
> > >
> > > /*
> > > - * we always limit bottom-up allocation above the kernel,
> > > - * but top-down allocation doesn't have the limit, so
> > > - * retrying top-down allocation may succeed when bottom-up
> > > - * allocation failed.
> > > - *
> > > * bottom-up allocation is expected to be fail very rarely,
> > > * so we use WARN_ONCE() here to see the stack trace if
> > > * fail happens.
> > >
> > > [1] https://lore.kernel.org/lkml/1545966002-3075-3-git-send-email-kernelfans@gmail.com/
> > > [2] https://lore.kernel.org/lkml/1545966002-3075-2-git-send-email-kernelfans@gmail.com/
> > >
> > > > +
> > > > + return ret;
> > > > +}
> > > > +
> > > > /**
> > > > * __memblock_find_range_top_down - find free area utility, in top-down
> > > > * @start: start of candidate range
> > > > --
> > > > 2.7.4
> > > >
> > >
> > > --
> > > Sincerely yours,
> > > Mike.
> > >
> >
>
> --
> Sincerely yours,
> Mike.
>

\
 
 \ /
  Last update: 2019-01-09 14:03    [W:0.097 / U:0.024 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site