kexec, x86: Need a new e820 type support for kexec

[Baoquan He]

Hi Toshi,

Sorry for replying late.

On 08/06/15 at 07:13pm, Toshi Kani wrote:
> On Thu, 2015-08-06 at 16:12 +0800, Baoquan He wrote:
> > Hi Toshi,
> > 
> > Does this patch work for you?
> 
> Hi Baoquan,
> 
> I have tested the patch with both E820_PMEM and E820_PRAM setups, and
> confirmed it works fine for both cases. :-)  I did multiple kexec reboots
> followed by a kdump in my testing.  So, please feel free to add:
>  
> Tested-by: Toshi Kani <toshi.kani at hp.com>

Thanks for testing, I will repost with Tested-by info.

> 
> > There are things I am not sure. When jump to kexec/kdump kernel is this
> > PMEM still needed by system? 
> 
> Yes, after a kexec reboot, the kernel needs to be able to use NVDIMM as
> before.  While the kernel actually uses NFIT table, not e820, the range
> should be marked as PMEM for consistency.  The same goes to kdump kernel
> since NVDIMM may be used as a dump device in future.
> 
> > And what's the difference between PRAM and
> > PMEM? I saw in kernel commit ec776ef6 it introduced E820_PRAM for the
> > non-standard protected e820 type, then in kernel commit ad5fb870 it
> > introduced E820_PMEM for ACPI 6.0 persistent memory types. While it
> > doesn't add complete support for E820_PMEM like E820_PRAM if I
> > understand it correctly.
> 
> ACPI 6.0 spec defines E820_PMEM, which is used for NVDIMM devices from now
> on.  ACPI 6.0 also defines NFIT table for NVDIMM along with this type.
> 
> Before these are defined in ACPI, E820_PRAM type was "unofficially" used by
> some NVDIMM devices.  So, E820_PRAM was added for such legacy NVDIMMs. 
>  Since the E820_PRAM case is very simple (it does not have any other FW
> tables), it can be easily emulated with the "memmap=nn!ss" option.  So,
> people may use the memmap option to emulate this legacy NVDIMM.    

I was wrong. In fact in kexec-tools memory info can be passed to kdump
kernel by 2 ways. One is using memmap by specifying
--pass-memmap-cmdline. The other one is storing memory regions in
e820_map of real mode data structure by default. And the 1st way is
rarely used. So no need to worry about the "memmap=nn!ss" option.

Since kernel parse_memmap_one doesn't support E820_PMEM well, I would
like to ignore the PMEM adding in memmap way. So this patch is enough.

> 
> >  In this patch I simply pass E820_PMEM to kdump
> > kernel as E820_PRAM when it emerges since kernel can parse E820_PRAM
> > only in parse_memmap_one(), otherwise E820_PMEM has to be discarded or
> > need be passed as E820_RESERVED. What do you think about this, need
> > E820_PMEM be differentiated with E820_PRAM strictly? If yes, I think a
> > kernel patch need be posted to fix this. If not, this patch is enough
> > for supporting both of them in kexec.
> 
> E820_PMEM cannot be emulated by the "memmap=" option.  Do you have to use
> the "memmap=" options to pass the ranges for kdump kernel?  If so, I'd
> rather ignore E820_PMEM and let it be passed as E820_RESERVED.  The kdump
> kernel can still obtain the info from NFIT if necessary.
> 
> As for the code change...
> 
> > @@ -640,6 +644,8 @@ static void cmdline_add_memmap_internal(char *cmdline, 
> > unsigned long startk,
> >  		strcat (str_mmap, "K$");
> >  	else if (type == RANGE_ACPI || type == RANGE_ACPI_NVS)
> >  		strcat (str_mmap, "K#");
> > +	else if (type == RANGE_PMEM || type == RANGE_PRAM)
> > +		strcat (str_mmap, "K!");
> 
> It should only check with RANGE_PRAM, but I do not think this change matters
> much unless you also modify the caller cmdline_add_memmap(), which has the
> following check to skip other types.  I do not think we will use legacy
> NVDIMM device as a dump device, so you may ignore RANGE_PRAM and let it be
> passed as RESERVED as well (which is likely the case I tested with).
> 
>                 /* Only adding memory regions of RAM and ACPI */
>                 if (type != RANGE_RAM &&
>                     type != RANGE_ACPI &&
>                     type != RANGE_ACPI_NVS)
>                         continue;

Then if ignore PMEM adding into memmap, cmdline_add_memmap need not be
cared any more.

Thanks
Baoquan