Re: [RESEND PATCH v2 1/2] pstore: Add mem_type property DT parsing support

[Kees Cook <keescook@chromium.org>]

On Thu, Feb 25, 2021 at 09:30:16PM +0530, Mukesh Ojha wrote:
> There could be a sceanario where we define some region
> in normal memory and use them store to logs which is later
> retrieved by bootloader during warm reset.
> 
> In this scenario, we wanted to treat this memory as normal
> cacheable memory instead of default behaviour which
> is an overhead. Making it cacheable could improve
> performance.
> 
> This commit gives control to change mem_type from Device
> tree, and also documents the value for normal memory.
> 
> Signed-off-by: Mukesh Ojha <mojha@codeaurora.org>
> ---
> Changes in v2:
>  - if-else converted to switch case
>  - updated MODULE_PARM_DESC with new memory type.
>  - default setting is still intact.
> 
>  Documentation/admin-guide/ramoops.rst |  4 +++-
>  fs/pstore/ram.c                       |  3 ++-
>  fs/pstore/ram_core.c                  | 18 ++++++++++++++++--
>  3 files changed, 21 insertions(+), 4 deletions(-)
> 
> diff --git a/Documentation/admin-guide/ramoops.rst b/Documentation/admin-guide/ramoops.rst
> index b0a1ae7..8f107d8 100644
> --- a/Documentation/admin-guide/ramoops.rst
> +++ b/Documentation/admin-guide/ramoops.rst
> @@ -3,7 +3,7 @@ Ramoops oops/panic logger
>  
>  Sergiu Iordache <sergiu@chromium.org>
>  
> -Updated: 17 November 2011
> +Updated: 10 Feb 2021
>  
>  Introduction
>  ------------
> @@ -30,6 +30,8 @@ mapping to pgprot_writecombine. Setting ``mem_type=1`` attempts to use
>  depends on atomic operations. At least on ARM, pgprot_noncached causes the
>  memory to be mapped strongly ordered, and atomic operations on strongly ordered
>  memory are implementation defined, and won't work on many ARMs such as omaps.
> +Setting ``mem_type=2`` attempts to treat the memory region as normal memory,
> +which enables full cache on it. This can improve the performance.
>  
>  The memory area is divided into ``record_size`` chunks (also rounded down to
>  power of two) and each kmesg dump writes a ``record_size`` chunk of
> diff --git a/fs/pstore/ram.c b/fs/pstore/ram.c
> index ca6d8a8..af4ca6a4 100644
> --- a/fs/pstore/ram.c
> +++ b/fs/pstore/ram.c
> @@ -56,7 +56,7 @@ MODULE_PARM_DESC(mem_size,
>  static unsigned int mem_type;
>  module_param(mem_type, uint, 0400);
>  MODULE_PARM_DESC(mem_type,
> -		"set to 1 to try to use unbuffered memory (default 0)");
> +		"set to 1 to use unbuffered memory, 2 for cached memory (default 0)");

I'd like to be as explicit as possible (0 wasn't listed), so about this:

"memory type: 0=write-combined (default), 1=unbuffered, 2=cached"

>  static int ramoops_max_reason = -1;
>  module_param_named(max_reason, ramoops_max_reason, int, 0400);
> @@ -666,6 +666,7 @@ static int ramoops_parse_dt(struct platform_device *pdev,
>  		field = value;						\
>  	}
>  
> +	parse_u32("mem-type", pdata->record_size, pdata->mem_type);

Please update the documentation at:
Documentation/devicetree/bindings/reserved-memory/ramoops.txt
(and please move and update the language about "unbuffered" being
deprecated like "no-dump-oops", so that it's clear what happens when
both are specified -- "mem-type" overrides "unbuffered".)

>  	parse_u32("record-size", pdata->record_size, 0);
>  	parse_u32("console-size", pdata->console_size, 0);
>  	parse_u32("ftrace-size", pdata->ftrace_size, 0);
> diff --git a/fs/pstore/ram_core.c b/fs/pstore/ram_core.c
> index aa8e0b6..0da012f 100644
> --- a/fs/pstore/ram_core.c
> +++ b/fs/pstore/ram_core.c
> @@ -396,6 +396,10 @@ void persistent_ram_zap(struct persistent_ram_zone *prz)
>  	persistent_ram_update_header_ecc(prz);
>  }
>  
> +#define MEM_TYPE_WCOMBINE	0
> +#define MEM_TYPE_NONCACHED	1
> +#define MEM_TYPE_NORMAL		2
> +
>  static void *persistent_ram_vmap(phys_addr_t start, size_t size,
>  		unsigned int memtype)
>  {
> @@ -409,10 +413,20 @@ static void *persistent_ram_vmap(phys_addr_t start, size_t size,
>  	page_start = start - offset_in_page(start);
>  	page_count = DIV_ROUND_UP(size + offset_in_page(start), PAGE_SIZE);
>  
> -	if (memtype)
> +	switch (memtype) {
> +	case MEM_TYPE_NORMAL:
> +		prot = PAGE_KERNEL;
> +		break;
> +	case MEM_TYPE_NONCACHED:
>  		prot = pgprot_noncached(PAGE_KERNEL);
> -	else
> +		break;
> +	case MEM_TYPE_WCOMBINE:
>  		prot = pgprot_writecombine(PAGE_KERNEL);
> +		break;
> +	default:
> +		pr_err("invalid memory type\n");

This should be more verbose:

		pr_err("invalid mem_type=%d\n", memtype);

> +		return NULL;
> +	}
>  
>  	pages = kmalloc_array(page_count, sizeof(struct page *), GFP_KERNEL);
>  	if (!pages) {

With those changes, it looks good to me. Thanks!

-- 
Kees Cook