Re: [PATCH v9 07/11] arm64: kexec_file: add crash dump support

From: AKASHI Takahiro
Date: Fri May 18 2018 - 05:43:15 EST


On Tue, May 15, 2018 at 06:11:15PM +0100, James Morse wrote:
> Hi Akashi,
>
> On 25/04/18 07:26, AKASHI Takahiro wrote:
> > Enabling crash dump (kdump) includes
> > * prepare contents of ELF header of a core dump file, /proc/vmcore,
> > using crash_prepare_elf64_headers(), and
> > * add two device tree properties, "linux,usable-memory-range" and
> > "linux,elfcorehdr", which represent repsectively a memory range
>
> (Nit: respectively)

Will fix.

>
> > to be used by crash dump kernel and the header's location
>
> > arch/arm64/include/asm/kexec.h | 4 +
> > arch/arm64/kernel/kexec_image.c | 9 +-
> > arch/arm64/kernel/machine_kexec_file.c | 202 +++++++++++++++++++++++++
>
> In this patch, machine_kexec_file.c gains its own private fdt array encoder.

See below.

>
> > diff --git a/arch/arm64/kernel/machine_kexec_file.c b/arch/arm64/kernel/machine_kexec_file.c
> > index 37c0a9dc2e47..ec674f4d267c 100644
> > --- a/arch/arm64/kernel/machine_kexec_file.c
> > +++ b/arch/arm64/kernel/machine_kexec_file.c
> > @@ -76,6 +81,78 @@ int arch_kexec_walk_mem(struct kexec_buf *kbuf,
> > return ret;
> > }
> >
> > +static int __init arch_kexec_file_init(void)
> > +{
> > + /* Those values are used later on loading the kernel */
> > + __dt_root_addr_cells = dt_root_addr_cells;
> > + __dt_root_size_cells = dt_root_size_cells;
> > +
> > + return 0;
> > +}
> > +late_initcall(arch_kexec_file_init);
>
> If we need these is it worth taking them out of __initdata? I note they've been
> 'temporary' for quite a long time.

I think that I had some reason that I didn't do that, but don't remember now.
If there's no problem, I will take your suggestion.

>
> > +
> > +#define FDT_ALIGN(x, a) (((x) + (a) - 1) & ~((a) - 1))
> > +#define FDT_TAGALIGN(x) (FDT_ALIGN((x), FDT_TAGSIZE))
> > +
> > +static int fdt_prop_len(const char *prop_name, int len)
> > +{
> > + return (strlen(prop_name) + 1) +
> > + sizeof(struct fdt_property) +
> > + FDT_TAGALIGN(len);
> > +}
>
> This stuff should really be in libfdt.h Those macros come from
> libfdt_internal.h, so we're probably doing something wrong here.
>
>
> > +static bool cells_size_fitted(unsigned long base, unsigned long size)
> > +{
> > + /* if *_cells >= 2, cells can hold 64-bit values anyway */
> > + if ((__dt_root_addr_cells == 1) && (base >= (1ULL << 32)))
> > + return false;
> > +
> > + if ((__dt_root_size_cells == 1) && (size >= (1ULL << 32)))
> > + return false;
>
> Using '> U32_MAX' here may be more readable.

OK

>
> > + return true;
> > +}
> > +
> > +static void fill_property(void *buf, u64 val64, int cells)
> > +{
> > + u32 val32;
> > +
> > + if (cells == 1) {
> > + val32 = cpu_to_fdt32((u32)val64);
> > + memcpy(buf, &val32, sizeof(val32));
> > + } else {
>
> > + memset(buf, 0, cells * sizeof(u32) - sizeof(u64));
> > + buf += cells * sizeof(u32) - sizeof(u64);
>
> Is this trying to clear the 'top' cells and shuffle the pointer to point at the
> 'bottom' 2? I'm pretty sure this isn't endian safe.
>
> Do we really expect a system to have #address-cells > 2?

I don't know, but just for safety.

>
> > + val64 = cpu_to_fdt64(val64);
> > + memcpy(buf, &val64, sizeof(val64));
> > + }
> > +}
> > +
> > +static int fdt_setprop_range(void *fdt, int nodeoffset, const char *name,
> > + unsigned long addr, unsigned long size)
>
> (the device-tree spec describes a 'ranges' property, which had me confused. This
> is encoding a prop-encoded-array)

Should we rename it to, say, fdt_setprop_reg()?


> > +{
> > + void *buf, *prop;
> > + size_t buf_size;
> > + int result;
> > +
> > + buf_size = (__dt_root_addr_cells + __dt_root_size_cells) * sizeof(u32);
> > + prop = buf = vmalloc(buf_size);
>
> virtual memory allocation for something less than PAGE_SIZE?

I've never cared about that. Let me think again.

>
> > + if (!buf)
> > + return -ENOMEM;
> > +
> > + fill_property(prop, addr, __dt_root_addr_cells);
> > + prop += __dt_root_addr_cells * sizeof(u32);
> > +
> > + fill_property(prop, size, __dt_root_size_cells);
> > +
> > + result = fdt_setprop(fdt, nodeoffset, name, buf, buf_size);
> > +
> > + vfree(buf);
> > +
> > + return result;
> > +}
>
> Doesn't this stuff belong in libfdt? I guess there is no 'add array element' api
> because this the first time we've wanted to create a node with more than
> key=fixed-size-value.
>
> I don't think this belongs in arch C code. Do we have a plan for getting libfdt
> to support encoding prop-arrays? Can we put it somewhere anyone else duplicating
> this will find it, until we can (re)move it?

I will temporarily move all fdt-related stuff to a separate file, but

> I have no idea how that happens... it looks like the devicetree list is the
> place to ask.

should we always sync with the original dtc/libfdt repository?

>
> > static int setup_dtb(struct kimage *image,
> > unsigned long initrd_load_addr, unsigned long initrd_len,
> > char *cmdline, unsigned long cmdline_len,
> > @@ -88,10 +165,26 @@ static int setup_dtb(struct kimage *image,
> > int range_len;
> > int ret;
> >
> > + /* check ranges against root's #address-cells and #size-cells */
> > + if (image->type == KEXEC_TYPE_CRASH &&
> > + (!cells_size_fitted(image->arch.elf_load_addr,
> > + image->arch.elf_headers_sz) ||
> > + !cells_size_fitted(crashk_res.start,
> > + crashk_res.end - crashk_res.start + 1))) {
> > + pr_err("Crash memory region doesn't fit into DT's root cell sizes.\n");
> > + ret = -EINVAL;
> > + goto out_err;
> > + }
>
> To check I've understood this properly: This can happen if the firmware provided
> a DTB with 32bit address/size cells, but at least some of the memory requires 64
> bit address/size cells. This could only happen on a UEFI system where the
> firmware-DTB doesn't describe memory. ACPI-only systems would have the EFIstub DT.

Probably, yes. I assumed the case where #address-cells and #size-cells
were just missing in fdt.

>
> > /* duplicate dt blob */
> > buf_size = fdt_totalsize(initial_boot_params);
> > range_len = (__dt_root_addr_cells + __dt_root_size_cells) * sizeof(u32);
> >
> > + if (image->type == KEXEC_TYPE_CRASH)
> > + buf_size += fdt_prop_len("linux,elfcorehdr", range_len)
> > + + fdt_prop_len("linux,usable-memory-range",
> > + range_len);

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

> > +
> > if (initrd_load_addr)
> > buf_size += fdt_prop_len("linux,initrd-start", sizeof(u64))
> > + fdt_prop_len("linux,initrd-end", sizeof(u64));
> > @@ -113,6 +206,23 @@ static int setup_dtb(struct kimage *image,
> > if (nodeoffset < 0)
> > goto out_err;
> >
> > + if (image->type == KEXEC_TYPE_CRASH) {
> > + /* add linux,elfcorehdr */
> > + ret = fdt_setprop_range(buf, nodeoffset, "linux,elfcorehdr",
> > + image->arch.elf_load_addr,
> > + image->arch.elf_headers_sz);
> > + if (ret)
> > + goto out_err;
> > +
> > + /* add linux,usable-memory-range */
> > + ret = fdt_setprop_range(buf, nodeoffset,
> > + "linux,usable-memory-range",
> > + crashk_res.start,
> > + crashk_res.end - crashk_res.start + 1);
>
> Don't you need to add "linux,usable-memory-range" to the buf_size estimate?

I think the code exists. See above.

>
> > + if (ret)
> > + goto out_err;
> > + }
>
> > @@ -148,17 +258,109 @@ static int setup_dtb(struct kimage *image,
>
> > +static struct crash_mem *get_crash_memory_ranges(void)
> > +{
> > + unsigned int nr_ranges;
> > + struct crash_mem *cmem;
> > +
> > + nr_ranges = 1; /* for exclusion of crashkernel region */
> > + walk_system_ram_res(0, -1, &nr_ranges, get_nr_ranges_callback);
> > +
> > + cmem = vmalloc(sizeof(struct crash_mem) +
> > + sizeof(struct crash_mem_range) * nr_ranges);
> > + if (!cmem)
> > + return NULL;
> > +
> > + cmem->max_nr_ranges = nr_ranges;
> > + cmem->nr_ranges = 0;
> > + walk_system_ram_res(0, -1, cmem, add_mem_range_callback);
> > +
> > + /* Exclude crashkernel region */
> > + if (crash_exclude_mem_range(cmem, crashk_res.start, crashk_res.end)) {
> > + vfree(cmem);
> > + return NULL;
> > + }
> > +
> > + return cmem;
> > +}
>
> Could this function be included in prepare_elf_headers() so that the alloc() and
> free() occur together.


Or aiming that arm64 and x86 have similar-look code?

>
> > +static int prepare_elf_headers(void **addr, unsigned long *sz)
> > +{
> > + struct crash_mem *cmem;
> > + int ret = 0;
> > +
> > + cmem = get_crash_memory_ranges();
> > + if (!cmem)
> > + return -ENOMEM;
> > +
> > + ret = crash_prepare_elf64_headers(cmem, true, addr, sz);
> > +
> > + vfree(cmem);
>
> > + return ret;
> > +}
>
> All this is moving memory-range information from core-code's
> walk_system_ram_res() into core-code's struct crash_mem, and excluding
> crashk_res, which again is accessible to the core code.
>
> It looks like this is duplicated in arch/x86 and arch/arm64 because arm64
> doesn't have a second 'crashk_low_res' region, and always wants elf64, instead
> of when IS_ENABLED(CONFIG_X86_64).
> If we can abstract just those two, more of this could be moved to core code
> where powerpc can make use of it if they want to support kdump with
> kexec_file_load().
>
> But, its getting late for cross-architecture dependencies, lets put that on the
> for-later list. (assuming there isn't a powerpc-kdump series out there adding a
> third copy of this)

Sure. X86 code has so many exceptional lines in the code :)

Thanks,
-Takahiro AKASHI


>
> Thanks,
>
> James