On Sat, Nov 25, 2023 at 03:49:53AM +0530, Mukesh Ojha wrote:
> The reserved memory region for ramoops is assumed to be at a fixed
> and known location when read from the devicetree. This may not be
> required for something like Qualcomm's minidump which is interested
> in knowing addresses of ramoops region but it does not put hard
> requirement of address being fixed as most of it's SoC does not
> support warm reset and does not use pstorefs at all instead it has
> firmware way of collecting ramoops region if it gets to know the
> address and register it with apss minidump table which is sitting
> in shared memory region in DDR and firmware will have access to
> these table during reset and collects it on crash of SoC.
> 
> So, add the support of reserving ramoops region to be dynamically
> allocated early during boot if it is request through command line
> via 'dyn_ramoops_size=<size>' and fill up reserved resource structure
> and export the structure, so that it can be read by ramoops driver.
> 
> Signed-off-by: Mukesh Ojha <quic_mo...@quicinc.com>
> ---
>  Documentation/admin-guide/ramoops.rst |  7 ++++
>  fs/pstore/Kconfig                     | 15 +++++++++
>  fs/pstore/ram.c                       | 62 
> ++++++++++++++++++++++++++++++++---
>  include/linux/pstore_ram.h            |  5 +++
>  init/main.c                           |  2 ++
>  5 files changed, 87 insertions(+), 4 deletions(-)
> 
> diff --git a/Documentation/admin-guide/ramoops.rst 
> b/Documentation/admin-guide/ramoops.rst
> index e9f85142182d..af737adbf079 100644
> --- a/Documentation/admin-guide/ramoops.rst
> +++ b/Documentation/admin-guide/ramoops.rst
> @@ -33,6 +33,13 @@ 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.
>  
> +Ramoops memory region can also be allocated dynamically for a special case 
> where
> +there is no requirement to access the logs from pstorefs on next boot 
> instead there
> +is separate backend mechanism like minidump present which has awareness 
> about the
> +dynamic ramoops region and can recover the logs. This is enabled via command 
> line
> +parameter ``dyn_ramoops_size=<size>`` and should not be used in absence of
> +separate backend which knows how to recover this dynamic region.
> +
>  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
>  information.
> diff --git a/fs/pstore/Kconfig b/fs/pstore/Kconfig
> index 3acc38600cd1..e13e53d7a225 100644
> --- a/fs/pstore/Kconfig
> +++ b/fs/pstore/Kconfig
> @@ -81,6 +81,21 @@ config PSTORE_RAM
>  
>         For more information, see Documentation/admin-guide/ramoops.rst.
>  
> +config PSTORE_DYNAMIC_RAMOOPS_REGION_RESERVATION
> +     bool "Reserve ramoops region dynamically"
> +     select PSTORE_RAM
> +     help
> +       This enables the dynamic reservation of ramoops region for a special 
> case
> +       where there is no requirement to access the logs from pstorefs on 
> next boot
> +       instead there is separate backend mechanism like minidump present 
> which has
> +       awareness about the dynamic ramoops region and can recover the logs. 
> This is
> +       enabled via command line parameter dyn_ramoops_size=<size> and should 
> not be
> +       used in absence of separate backend which knows how to recover this 
> dynamic
> +       region.
> +
> +       Note whenever this config is selected ramoops driver will be build 
> statically
> +       into kernel.
> +

Is there any advantage if we decouple this memory reservation from
pstore ram so that pstore ram can still be compiled as module? Asking
because you explicitly mentioned this limitation.

>  config PSTORE_ZONE
>       tristate
>       depends on PSTORE
> diff --git a/fs/pstore/ram.c b/fs/pstore/ram.c
> index 88b34fdbf759..a6c0da8cfdd4 100644
> --- a/fs/pstore/ram.c
> +++ b/fs/pstore/ram.c
> @@ -20,6 +20,7 @@
>  #include <linux/compiler.h>
>  #include <linux/of.h>
>  #include <linux/of_address.h>
> +#include <linux/memblock.h>
>  #include <linux/mm.h>
>  
>  #include "internal.h"
> @@ -103,6 +104,55 @@ struct ramoops_context {
>  };
>  
>  static struct platform_device *dummy;
> +static int dyn_ramoops_size;
> +/* Location of the reserved area for the dynamic ramoops */
> +static struct resource dyn_ramoops_res = {
> +     .name  = "ramoops",
> +     .start = 0,
> +     .end   = 0,
> +     .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
> +     .desc  = IORES_DESC_NONE,
> +};
> +
> +static int __init parse_dyn_ramoops_size(char *p)
> +{
> +     char *tmp;
> +
> +     dyn_ramoops_size = memparse(p, &tmp);
> +     if (p == tmp) {
> +             pr_err("ramoops: memory size expected\n");
> +             return -EINVAL;
> +     }
> +
> +     return 0;
> +}
> +early_param("dyn_ramoops_size", parse_dyn_ramoops_size);

should not this code be under
CONFIG_PSTORE_DYNAMIC_RAMOOPS_REGION_RESERVATION?

> +
> +#ifdef CONFIG_PSTORE_DYNAMIC_RAMOOPS_REGION_RESERVATION
> +/*
> + * setup_dynamic_ramoops() - reserves memory for dynamic ramoops
> + *
> + * This enable dynamic reserve memory support for ramoops through
> + * command line.
> + */
> +void __init setup_dynamic_ramoops(void)
> +{
> +     unsigned long long ramoops_base;
> +     unsigned long long ramoops_size;
> +
> +     ramoops_base = memblock_phys_alloc_range(dyn_ramoops_size, 
> SMP_CACHE_BYTES,
> +                                              0, MEMBLOCK_ALLOC_NOLEAKTRACE);
> +     if (!ramoops_base) {
> +             pr_err("cannot allocate ramoops dynamic memory 
> (size:0x%llx).\n",
> +                     ramoops_size);
> +             return;
> +     }

This error needs to be propagated to ramoops_register_dummy() since it
rely on !dyn_ramoops_size . one way is to set dyn_ramoops_size to 0.

> +
> +     dyn_ramoops_res.start = ramoops_base;
> +     dyn_ramoops_res.end = ramoops_base + dyn_ramoops_size - 1;
> +     insert_resource(&iomem_resource, &dyn_ramoops_res);
> +}
> +#endif
>  
>  static int ramoops_pstore_open(struct pstore_info *psi)
>  {
> @@ -915,14 +965,18 @@ static void __init ramoops_register_dummy(void)
>  
>       /*
>        * Prepare a dummy platform data structure to carry the module
> -      * parameters. If mem_size isn't set, then there are no module
> -      * parameters, and we can skip this.
> +      * parameters. If mem_size isn't set, check for dynamic ramoops
> +      * size and use if it is set.
>        */
> -     if (!mem_size)
> +     if (!mem_size && !dyn_ramoops_size)
>               return;
>  

If mem_size and dyn_ramoops_size are set, you are taking
dyn_ramoops_size precedence here. The comment is a bit confusing, pls
review it once.

> -     pr_info("using module parameters\n");
> +     if (dyn_ramoops_size) {
> +             mem_size = dyn_ramoops_size;
> +             mem_address = dyn_ramoops_res.start;
> +     }
>  

Overall it Looks good to me. Thanks.

Reply via email to