On Wed, Oct 17, 2018 at 08:19:41PM +0530, Sai Prakash Ranjan wrote:
> On 10/17/2018 5:08 PM, Sai Prakash Ranjan wrote:
> > >
> > > What do you think about the (untested) patch below? It seems to me
> > > that it
> > > should solve the issue of missing early crash dumps, but I have not
> > > tested it
> > > yet. Sai, would you mind trying it out and let me know if you can see the
> > > early crash dumps properly now?
> > >
> > > ----8<---
> > > From: "Joel Fernandes (Google)" <joel@xxxxxxxxxxxxxxxxx>
> > > Subject: [RFC] pstore: allocate compression during late_initcall
> > >
> > > ramoop's pstore registration (using pstore_register) has to run during
> > > late_initcall because crypto backend may not be ready during
> > > postcore_initcall. This causes missing of dmesg crash dumps which could
> > > have been caught by pstore.
> > >
> > > Instead, lets allow ramoops pstore registration earlier, and once crypto
> > > is ready we can initialize the compression.
> > >
> > > Reported-by: Sai Prakash Ranjan <saiprakash.ranjan@xxxxxxxxxxxxxx>
> > > Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > > ---
> > > fs/pstore/platform.c | 13 +++++++++++++
> > > fs/pstore/ram.c | 2 +-
> > > 2 files changed, 14 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/fs/pstore/platform.c b/fs/pstore/platform.c
> > > index 15e99d5a681d..f09066db2d4d 100644
> > > --- a/fs/pstore/platform.c
> > > +++ b/fs/pstore/platform.c
> > > @@ -780,6 +780,19 @@ void __init pstore_choose_compression(void)
> > > }
> > > }
> > > +static int __init pstore_compression_late_init(void)
> > > +{
> > > + /*
> > > + * Check if any pstore backends registered earlier but did not
> > > allocate
> > > + * for compression because crypto was not ready, if so then
> > > initialize
> > > + * compression.
> > > + */
> > > + if (psinfo && !tfm)
> > > + allocate_buf_for_compression();
> > > + return 0;
> > > +}
> > > +late_initcall(pstore_compression_late_init);
> > > +
> > > module_param(compress, charp, 0444);
> > > MODULE_PARM_DESC(compress, "Pstore compression to use");
> > > diff --git a/fs/pstore/ram.c b/fs/pstore/ram.c
> > > index bbd1e357c23d..98e48d1a9776 100644
> > > --- a/fs/pstore/ram.c
> > > +++ b/fs/pstore/ram.c
> > > @@ -940,7 +940,7 @@ static int __init ramoops_init(void)
> > > ramoops_register_dummy();
> > > return platform_driver_register(&ramoops_driver);
> > > }
> > > -late_initcall(ramoops_init);
> > > +postcore_initcall(ramoops_init);
> > > static void __exit ramoops_exit(void)
> > > {
> > >
> >
> > Yes I could see the early crash dump. Also I tested with different
> > compression (LZO) instead of deflate just to be sure and it works fine,
> > thanks :)
> >
> > Tested-by: Sai Prakash Ranjan <saiprakash.ranjan@xxxxxxxxxxxxxx>
> >
Thanks.
> I just noticed that allocate_buf_for_compression() is also called from
> pstore_register(). Shouldn't that call be removed now that ramoops_init is
> moved to postcore_initcall and allocate_buf_for_compression() will just
> return doing nothing when called from pstore_register()?
Yes, that is the point. If crypto is not ready then my thought is
allocate_buf_for_compression() called from pstore_register() should do
nothing and pstore will work uncompressed and the kdump infrastructure will
only cause uncompressed writes. But say if the kdump triggered *after* crypto
was ready, then the dump write out would be compressed since pstore is ready
for compression.
The idea is if a future pstore backend calls pstore_register late, then it
may as well do the allocate_buf_for_compression as well at that time when it
runs. In that cause pstore_compression_late_init would do nothing.
So this approach is both dynamic and future proof.
- Joel
