RE: [RFC PATCH] efi: arm64: Introduce /sys/firmware/efi/memreserve to tell the persistent pages

[d.hatayama@xxxxxxxxxxx]

> -----Original Message-----
> From: linux-kernel-owner@xxxxxxxxxxxxxxx
> [mailto:linux-kernel-owner@xxxxxxxxxxxxxxx] On Behalf Of Masayoshi Mizuma
> Sent: Wednesday, November 13, 2019 1:53 AM
> To: Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx>;
> linux-arm-kernel@xxxxxxxxxxxxxxxxxxx; linux-efi@xxxxxxxxxxxxxxx
> Cc: Masayoshi Mizuma <msys.mizuma@xxxxxxxxx>; Mizuma, Masayoshi/水間 理仁
> <m.mizuma@xxxxxxxxxxxxxx>; linux-kernel@xxxxxxxxxxxxxxx;
> kexec@xxxxxxxxxxxxxxxxxxx
> Subject: [RFC PATCH] efi: arm64: Introduce /sys/firmware/efi/memreserve to
> tell the persistent pages
> 
> From: Masayoshi Mizuma <m.mizuma@xxxxxxxxxxxxxx>
> 
> kexec reboot stucks because efi_config_parse_tables() refers garbage
>  (with memblock=debug):
> 
>   efi:  ACPI 2.0=0x9821790014  PROP=0x8757f5c0  SMBIOS 3.0=0x9820740000
> MEMRESERVE=0x9820bfdc58
>   memblock_reserve: [0x0000009820bfdc58-0x0000009820bfdc67]
> efi_config_parse_tables+0x228/0x278
>   memblock_reserve: [0x0000000082760000-0x00000000324d07ff]
> efi_config_parse_tables+0x228/0x278
>   memblock_reserve: [0xcc4f84ecc0511670-0x5f6e5214a7fd91f9]
> efi_config_parse_tables+0x244/0x278
>   memblock_reserve: [0xd2fd4144b9af693d-0xad0c1db1086f40a2]
> efi_config_parse_tables+0x244/0x278
>   memblock_reserve: [0x0c719bb159b1fadc-0x5aa6e62a1417ce12]
> efi_config_parse_tables+0x244/0x278
>   ...
> 
> That happens because 0x82760000, struct linux_efi_memreserve, is destroyed.
> 0x82760000 is pointed from efi.mem_reseve, and efi.mem_reserve points the
> head page of pending table and prop table which are allocated by
> gic_reserve_range().
> 
> The destroyer is kexec. kexec locates the inird to the area:
> 
> # kexec -d -l /boot/vmlinuz-5.4.0-rc7 /boot/initramfs-5.4.0-rc7.img
> --reuse-cmdline
> ...
> initrd: base 82290000, size 388dd8ah (59301258)
> ...
> 
> From dynamic debug log:
>   machine_kexec_prepare:70:
>     kexec kimage info:
>       type:        0
>       start:       85b30680
>       head:        0
>       nr_segments: 4
>         segment[0]: 0000000080480000 - 0000000082290000, 0x1e10000 bytes, 481
> pages
>         segment[1]: 0000000082290000 - 0000000085b20000, 0x3890000 bytes, 905
> pages
>         segment[2]: 0000000085b20000 - 0000000085b30000, 0x10000 bytes, 1
> pages
>         segment[3]: 0000000085b30000 - 0000000085b40000, 0x10000 bytes, 1
> pages
> 
> kexec searches the appropriate memory region to locate initrd through "System
> RAM"
> in /proc/iomem. The pending tables are included in "System RAM" because they
> are
> allocated by alloc_pages(), so kexec destroys the pending tables.
> 
> Introduce /sys/firmware/efi/memreserve to tell the pages pointed by
> efi.mem_reserve
> so that kexec can avoid the area to locate initrd.
> 
> Signed-off-by: Masayoshi Mizuma <m.mizuma@xxxxxxxxxxxxxx>
> ---
>  drivers/firmware/efi/efi.c | 32 +++++++++++++++++++++++++++++++-
>  1 file changed, 31 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/firmware/efi/efi.c b/drivers/firmware/efi/efi.c
> index e98bbf8e5..67b21ae7a 100644
> --- a/drivers/firmware/efi/efi.c
> +++ b/drivers/firmware/efi/efi.c
> @@ -141,6 +141,36 @@ static ssize_t systab_show(struct kobject *kobj,
> 
>  static struct kobj_attribute efi_attr_systab = __ATTR_RO_MODE(systab, 0400);
> 
> +static struct linux_efi_memreserve *efi_memreserve_root __ro_after_init;
> +static ssize_t memreserve_show(struct kobject *kobj,
> +			   struct kobj_attribute *attr, char *buf)
> +{
> +	struct linux_efi_memreserve *rsv;
> +	unsigned long prsv;
> +	char *str = buf;
> +	int index, i;
> +
> +	if (!kobj || !buf)
> +		return -EINVAL;
> +
> +	if (!efi_memreserve_root)
> +		return -ENODEV;

Other functions use different conditions.
The latter efi_memreserve_root == (void *)ULONG_MAX is correct?

static int __init efi_memreserve_map_root(void)
{
        if (efi.mem_reserve == EFI_INVALID_TABLE_ADDR)
                return -ENODEV;

int __ref efi_mem_reserve_persistent(phys_addr_t addr, u64 size)
{
        struct linux_efi_memreserve *rsv;
        unsigned long prsv;
        int rc, index;

        if (efi_memreserve_root == (void *)ULONG_MAX)
                return -ENODEV;

> +
> +	for (prsv = efi_memreserve_root->next; prsv; prsv = rsv->next) {
> +		rsv = memremap(prsv, sizeof(*rsv), MEMREMAP_WB);

memremap() could fail with NULL as a return value.
You need to deal with such case.

It looks to me efi_mem_reserve_persistent() also doesn't deal with this.
Maybe you should fix this, too.

> +		index = atomic_read(&rsv->count);
> +		for (i = 0; i < index; i++)
> +			str += sprintf(str, "%llx-%llx\n",
> +				rsv->entry[i].base,
> +				rsv->entry[i].base + rsv->entry[i].size -
> 1);

Is memreserve supported on 32-bit system?
If so, phy_addr_t could have a type of 4-byte length in such system
(not so if with PAE) and then %llx could lead to inconsistent type error.
It's enough to add a cast to unsigned long long.

> +		memunmap(rsv);
> +	}
> +
> +	return str - buf;
> +}
> +
> +static struct kobj_attribute efi_attr_memreserve =
> __ATTR_RO_MODE(memreserve, 0444);
> +
>  #define EFI_FIELD(var) efi.var
> 
>  #define EFI_ATTR_SHOW(name) \
> @@ -172,6 +202,7 @@ static struct attribute *efi_subsys_attrs[] = {
>  	&efi_attr_runtime.attr,
>  	&efi_attr_config_table.attr,
>  	&efi_attr_fw_platform_size.attr,
> +	&efi_attr_memreserve.attr,
>  	NULL,
>  };
> 
> @@ -955,7 +986,6 @@ int efi_status_to_err(efi_status_t status)
>  }
> 
>  static DEFINE_SPINLOCK(efi_mem_reserve_persistent_lock);
> -static struct linux_efi_memreserve *efi_memreserve_root __ro_after_init;
> 
>  static int __init efi_memreserve_map_root(void)
>  {
> --
> 2.18.1