[RFC PATCH v3 0/7] device backed vmemmap crash dump support

[Zhijian Li (Fujitsu)]

ping


Any comment is welcome.


On 06/03/2024 18:28, Li Zhijian wrote:
> Hello folks,
> 
> Compared with the V2[1] I posted a long time ago, this time it is a
> completely new proposal design.
> 
> ### Background and motivate overview ###
> ---
> Crash dump is an important feature for troubleshooting the kernel. It is the
> final way to chase what happened at the kernel panic, slowdown, and so on. It
> is one of the most important tools for customer support.
> 
> Currently, there are 2 syscalls(kexec_file_load(2) and kexec_load(2)) to
> configure the dumpable regions. Generally, (A)iomem resources registered with
> flags (IORESOURCE_SYSTEM_RAM | IORESOUCE_BUSY) for kexec_file_load(2) or
> (B)iomem resources registered with "System RAM" name prefix for kexec_load(2)
> are dumpable.
> 
> The pmem use cases including fsdax and devdax, could map their vmemmap to
> their own devices. In this case, these part of vmemmap will not be dumped when
> crash happened since these regions are satisfied with neither the above (A)
> nor (B).
> 
> In fsdax, the vmemmap(struct page array) becomes very important, it is one of
> the key data to find status of reverse map. Lacking of the information may
> cause difficulty to analyze trouble around pmem (especially Filesystem-DAX).
> That means troubleshooters are unable to check more details about pmem from
> the dumpfile.
> 
> ### Proposal ###
> ---
> In this proposal, register the device backed vmemmap as a separate resource.
> This resource has its own new flag and name, and then teaches kexec_file_load(2)
> and kexec_load(2) to mark it as dumpable.
> 
> Proposed flag: IORESOURCE_DEVICE_BACKED_VMEMMAP
> Proposed name: "Device Backed Vmemmap"
> 
> NOTE: crash-utils also needs to adapt to this new name for kexec_load()
> 
> With current proposal, the /proc/iomem should show as following for device
> backed vmemmap
> # cat /proc/iomem
> ...
> fffc0000-ffffffff : Reserved
> 100000000-13fffffff : Persistent Memory
>    100000000-10fffffff : namespace0.0
>      100000000-1005fffff : Device Backed Vmemmap  # fsdax
> a80000000-b7fffffff : CXL Window 0
>    a80000000-affffffff : Persistent Memory
>      a80000000-affffffff : region1
>        a80000000-a811fffff : namespace1.0
>          a80000000-a811fffff : Device Backed Vmemmap # devdax
>        a81200000-abfffffff : dax1.0
> b80000000-c7fffffff : CXL Window 1
> c80000000-147fffffff : PCI Bus 0000:00
>    c80000000-c801fffff : PCI Bus 0000:01
> ...
> 
> ### Kdump service reloading ###
> ---
> Once the kdump service is loaded, if changes to CPUs or memory occur,
> either by hot un/plug or off/onlining, the crash elfcorehdr should also
> be updated. There are 2 approaches to make the reloading more efficient.
> 1) Use udev rules to watch CPU and memory events, then reload kdump
> 2) Enable kernel crash hotplug to automatically reload elfcorehdr (>= 6.5)
> 
> This reloading also needed when device backed vmemmap layouts change, Similar
> to what 1) does now, one could add the following as the first lines to the
> RHEL udev rule file /usr/lib/udev/rules.d/98-kexec.rules:
> 
> # namespace updated: watch daxX.Y(devdax) and pfnX.Y(fsdax) of nd
> SUBSYSTEM=="nd", KERNEL=="[dp][af][xn][0-9].*", ACTION=="bind", GOTO="kdump_reload"
> SUBSYSTEM=="nd", KERNEL=="[dp][af][xn][0-9].*", ACTION=="unbind", GOTO="kdump_reload"
> # devdax <-> system-ram updated: watch daxX.Y of dax
> SUBSYSTEM=="dax", KERNEL=="dax[0-9].*", ACTION=="bind", GOTO="kdump_reload"
> SUBSYSTEM=="dax", KERNEL=="dax[0-9].*", ACTION=="unbind", GOTO="kdump_reload"
> 
> Regarding 2), my idea is that it would need to call the memory_notify() in
> devm_memremap_pages_release() and devm_memremap_pages() to trigger the crash
> hotplug. This part is not yet mature, but it does not affect the whole feature
> because we can still use method 1) alternatively.
> 
> [1] https://lore.kernel.org/lkml/02066f0f-dbc0-0388-4233-8e24b6f8435b@fujitsu.com/T/
> --------------------------------------------
> changes from V2[1]
> - new proposal design
> 
> CC: Alison Schofield <alison.schofield at intel.com>
> CC: Andrew Morton <akpm at linux-foundation.org>
> CC: Baoquan He <bhe at redhat.com>
> CC: Borislav Petkov <bp at alien8.de>
> CC: Dan Williams <dan.j.williams at intel.com>
> CC: Dave Hansen <dave.hansen at linux.intel.com>
> CC: Dave Jiang <dave.jiang at intel.com>
> CC: Greg Kroah-Hartman <gregkh at linuxfoundation.org>
> CC: "H. Peter Anvin" <hpa at zytor.com>
> CC: Ingo Molnar <mingo at redhat.com>
> CC: Ira Weiny <ira.weiny at intel.com>
> CC: Thomas Gleixner <tglx at linutronix.de>
> CC: Vishal Verma <vishal.l.verma at intel.com>
> CC: linux-cxl at vger.kernel.org
> CC: linux-mm at kvack.org
> CC: nvdimm at lists.linux.dev
> CC: x86 at kernel.org
> 
> Li Zhijian (7):
>    mm: memremap: register/unregister altmap region to a separate resource
>    mm: memremap: add pgmap_parent_resource() helper
>    nvdimm: pmem: assign a parent resource for vmemmap region for the
>      fsdax
>    dax: pmem: assign a parent resource for vmemmap region for the devdax
>    resource: Introduce walk device_backed_vmemmap res() helper
>    x86/crash: make device backed vmemmap dumpable for kexec_file_load
>    nvdimm: set force_raw=1 in kdump kernel
> 
>   arch/x86/kernel/crash.c         |  5 +++++
>   drivers/dax/pmem.c              |  8 ++++++--
>   drivers/nvdimm/namespace_devs.c |  3 +++
>   drivers/nvdimm/pmem.c           |  9 ++++++---
>   include/linux/ioport.h          |  4 ++++
>   include/linux/memremap.h        |  4 ++++
>   kernel/resource.c               | 13 +++++++++++++
>   mm/memremap.c                   | 30 +++++++++++++++++++++++++++++-
>   8 files changed, 70 insertions(+), 6 deletions(-)
>