Re: [PATCH v11 11/15] arm64: kexec_file: add crash dump support

[AKASHI Takahiro]

Hi James,

On Wed, Jul 18, 2018 at 05:50:22PM +0100, James Morse wrote:
> Hi Akashi,
> 
> On 11/07/18 08:41, 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 respectively a memory range
> >   to be used by crash dump kernel and the header's location
> 
> > diff --git a/arch/arm64/include/asm/kexec.h b/arch/arm64/include/asm/kexec.h
> > index 69333694e3e2..eeb5766928b0 100644
> > --- a/arch/arm64/include/asm/kexec.h
> > +++ b/arch/arm64/include/asm/kexec.h
> > @@ -99,6 +99,10 @@ static inline void crash_post_resume(void) {}
> >  struct kimage_arch {
> >  	phys_addr_t dtb_mem;
> >  	void *dtb_buf;
> > +	/* Core ELF header buffer */
> 
> > +	void *elf_headers;
> 
> Shouldn't this be a phys_addr_t if it comes from kbuf.mem?

Do you mean elf_load_addr? You're right.
But kexec_buf defined mem as unsigned long and so I'd rather change
dtb_mem to unsigned long instead of elf_load_addr, which will also be
renamed to elf_headers_mem for clarification.

> (dtb_mem is, and they type tells us which way round the runtime/kexec-time
> pointers are)
> 
> 
> > +	unsigned long elf_headers_sz;
> > +	unsigned long elf_load_addr;
> >  };
> >  
> >  /**
> 
> 
> > diff --git a/arch/arm64/kernel/machine_kexec_file.c b/arch/arm64/kernel/machine_kexec_file.c
> > index a0b44fe18b95..261564df7210 100644
> > --- a/arch/arm64/kernel/machine_kexec_file.c
> > +++ b/arch/arm64/kernel/machine_kexec_file.c
> > @@ -132,6 +173,45 @@ static int setup_dtb(struct kimage *image,
> >  	return ret;
> >  }
> >  
> > +static int prepare_elf_headers(void **addr, unsigned long *sz)
> > +{
> > +	struct crash_mem *cmem;
> > +	unsigned int nr_ranges;
> > +	int ret;
> > +	u64 i;
> > +	phys_addr_t start, end;
> 
> > +	nr_ranges = 1; /* for exclusion of crashkernel region */
> > +	for_each_mem_range(i, &memblock.memory, NULL, NUMA_NO_NODE, 0,
> > +							&start, &end, NULL)
> 
> Nit: flags = MEMBLOCK_NONE? Just to make it obvious this is how MEMBLOCK_NOMAP
> regions are weeded out.

OK.

> This is going to get interesting if we ever support hotpluggable memory... but
> it works for now and implicitly removes the nomap regions.
> 
> 
> > +		nr_ranges++;
> 
> > +
> > +	cmem = kmalloc(sizeof(struct crash_mem) +
> > +			sizeof(struct crash_mem_range) * nr_ranges, GFP_KERNEL);
> > +	if (!cmem)
> > +		return -ENOMEM;
> > +
> > +	cmem->max_nr_ranges = nr_ranges;
> > +	cmem->nr_ranges = 0;
> > +	for_each_mem_range(i, &memblock.memory, NULL, NUMA_NO_NODE, 0,
> > +							&start, &end, NULL) {
> > +		cmem->ranges[cmem->nr_ranges].start = start;
> > +		cmem->ranges[cmem->nr_ranges].end = end - 1;
> > +		cmem->nr_ranges++;
> > +	}
> > +
> > +	/* Exclude crashkernel region */
> > +	ret = crash_exclude_mem_range(cmem, crashk_res.start, crashk_res.end);
> 
> 
> > +	if (ret)
> > +		goto out;
> > +
> > +	ret =  crash_prepare_elf64_headers(cmem, true, addr, sz);
> > +
> > +out:
> 
> Nit: You could save the goto if you wrote this as:
> |	if (!ret)
> |		ret = crash_prepare_elf64_headers(cmem, true, addr, sz);

OK.

> > +	kfree(cmem);
> > +	return ret;
> > +}
> > +
> >  int load_other_segments(struct kimage *image,
> >  			unsigned long kernel_load_addr,
> >  			unsigned long kernel_size,
> > @@ -139,11 +219,43 @@ int load_other_segments(struct kimage *image,
> >  			char *cmdline, unsigned long cmdline_len)
> >  {
> >  	struct kexec_buf kbuf;
> > +	void *hdrs_addr;
> > +	unsigned long hdrs_sz;
> >  	unsigned long initrd_load_addr = 0;
> >  	char *dtb = NULL;
> >  	unsigned long dtb_len = 0;
> >  	int ret = 0;
> >  
> > +	/* load elf core header */
> > +	if (image->type == KEXEC_TYPE_CRASH) {
> > +		ret = prepare_elf_headers(&hdrs_addr, &hdrs_sz);
> > +		if (ret) {
> > +			pr_err("Preparing elf core header failed\n");
> > +			goto out_err;
> > +		}
> > +
> > +		kbuf.image = image;
> > +		kbuf.buffer = hdrs_addr;
> > +		kbuf.bufsz = hdrs_sz;
> > +		kbuf.memsz = hdrs_sz;
> 
> > +		kbuf.buf_align = PAGE_SIZE;
> 
> Whose PAGE_SIZE?
> 
> Won't this break if the kdump kernel is 64K pages, but the first kernel uses 4K?
> Should we change this to the largest supported PAGE_SIZE: SZ_64K?

Ah, yes.

> > +		kbuf.buf_min = crashk_res.start;
> > +		kbuf.buf_max = crashk_res.end + 1;
> > +		kbuf.top_down = true;
> > +
> > +		ret = kexec_add_buffer(&kbuf);
> > +		if (ret) {
> > +			vfree(hdrs_addr);
> > +			goto out_err;
> > +		}
> > +		image->arch.elf_headers = hdrs_addr;
> > +		image->arch.elf_headers_sz = hdrs_sz;
> > +		image->arch.elf_load_addr = kbuf.mem;
> > +
> > +		pr_debug("Loaded elf core header at 0x%lx bufsz=0x%lx memsz=0x%lx\n",
> > +				 image->arch.elf_load_addr, hdrs_sz, hdrs_sz);
> > +	}
> > +
> >  	kbuf.image = image;
> >  	/* not allocate anything below the kernel */
> >  	kbuf.buf_min = kernel_load_addr + kernel_size;
> 
> 
> I think the initramfs can escape the crash kernel range because you add to the
> buf_max region:
> |	/* within 1GB-aligned window of up to 32GB in size */
> |	kbuf.buf_max = round_down(kernel_load_addr, SZ_1G)
> |				â+ (unsigned long)SZ_1G * 32;

No worries.
kexec_add_buffer() will limit the search only within crashk_res anyway.

On the other hand, the code:
> > +	if (image->type == KEXEC_TYPE_CRASH) {
                (snip)
> > +		kbuf.buf_min = crashk_res.start;
> > +		kbuf.buf_max = crashk_res.end + 1;
can be misleading. I will fix it as follows:
|		kbuf.buf_min = kernel_load_addr + kernel_size;
|		kbuf.buf_max = ULONG_MAX;
(and likewise, will fix image_load().)

Thank you again for your valuable comments.
Are you reviewing other patches in my v11?
If not, I will post v12 tomorrow.

-Takahiro AKASHI

> 
> I think we need a helper to clamp these min/max ranges to within the crash
> kernel range, as its needs doing in a few places.
> 
> 
> Thanks,
> 
> James