[PATCH] xen/dt: Skip non-RAM reserved-memory regions in make_memory_node()

[PATCH] xen/dt: Skip non-RAM reserved-memory regions in make_memory_node()

[Kuan-Wei Chiu]

Xen creates a second /memory node for Dom0 from
bootinfo_get_reserved_mem() without checking whether those regions
belong to host RAM. However, /reserved-memory can also describe
on-chip SRAM or firmware regions outside system RAM. Listing them in a
/memory node causes Dom0 to register non-RAM addresses as system RAM,
conflicting with device MMIO mappings.

Skip reserved-memory banks outside the host RAM banks reported by
bootinfo_get_mem() in make_memory_node().

Fixes: 248faa637d2f ("xen/arm: add reserved-memory regions to the dom0 memory node")
Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>
---
Found while trying to bring up Xen on new hardware.
Tested on real silicon.

 xen/common/device-tree/domain-build.c | 42 +++++++++++++++++++++++++--
 1 file changed, 39 insertions(+), 3 deletions(-)

diff --git a/xen/common/device-tree/domain-build.c b/xen/common/device-tree/domain-build.c
index 30a59abfa7..33baf150a6 100644
--- a/xen/common/device-tree/domain-build.c
+++ b/xen/common/device-tree/domain-build.c
@@ -504,6 +504,34 @@ int __init make_chosen_node(const struct kernel_info *kinfo)
     return res;
 }
 
+static bool __init is_in_host_ram(const struct membank *bank)
+{
+    const struct membanks *ram = bootinfo_get_mem();
+    paddr_t start = bank->start;
+    paddr_t end = start + bank->size;
+    unsigned int i;
+
+    if ( !bank->size || end < start )
+        return false;
+
+    for ( i = 0; i < ram->nr_banks; i++ )
+    {
+        paddr_t ram_start = ram->bank[i].start;
+        paddr_t ram_end = ram_start + ram->bank[i].size;
+
+        if ( start < ram_start )
+            return false;
+
+        if ( end <= ram_end )
+            return true;
+
+        if ( start < ram_end )
+            start = ram_end;
+    }
+
+    return false;
+}
+
 int __init make_memory_node(const struct kernel_info *kinfo, int addrcells,
                             int sizecells, const struct membanks *mem)
 {
@@ -513,14 +541,20 @@ int __init make_memory_node(const struct kernel_info *kinfo, int addrcells,
     int nr_cells = 0;
     __be32 reg[DT_MEM_NODE_REG_RANGE_SIZE];
     __be32 *cells;
+    bool is_resv = (mem == bootinfo_get_reserved_mem());
 
     if ( mem->nr_banks == 0 )
         return -ENOENT;
 
     /* find the first memory range that is reserved for device (or firmware) */
-    for ( i = 0; i < mem->nr_banks &&
-                 (mem->bank[i].type != MEMBANK_DEFAULT); i++ )
-        ;
+    for ( i = 0; i < mem->nr_banks; i++ )
+    {
+        if ( mem->bank[i].type != MEMBANK_DEFAULT )
+            continue;
+        if ( is_resv && !is_in_host_ram(&mem->bank[i]) )
+            continue;
+        break;
+    }
 
     if ( i == mem->nr_banks )
         return 0;
@@ -544,6 +578,8 @@ int __init make_memory_node(const struct kernel_info *kinfo, int addrcells,
         if ( (mem->bank[i].type == MEMBANK_STATIC_DOMAIN) ||
              (mem->bank[i].type == MEMBANK_FDT_RESVMEM) )
             continue;
+        if ( is_resv && !is_in_host_ram(&mem->bank[i]) )
+            continue;
 
         nr_cells += reg_size;
         BUG_ON(nr_cells > ARRAY_SIZE(reg));
-- 
2.56.0.385.gac671c819a-goog

Re: [PATCH] xen/dt: Skip non-RAM reserved-memory regions in make_memory_node()

[Orzel, Michal]

On 28-Oct-26 22:17, Kuan-Wei Chiu wrote:
> Xen creates a second /memory node for Dom0 from
> bootinfo_get_reserved_mem() without checking whether those regions
> belong to host RAM. However, /reserved-memory can also describe
> on-chip SRAM or firmware regions outside system RAM. Listing them in a
> /memory node causes Dom0 to register non-RAM addresses as system RAM,
> conflicting with device MMIO mappings.
Hmm, I always thought that /reserved-memory is to exclude regions from memory
described under /memory. That's what Xen assumes too. map_range_to_domain() for
example does not call iomem_permit_access() for such regions because we assume
it's RAM. Can you provide more details of your board e.g. the device tree? Is
there any documentation I could use to confirm that "reserved-memory can also
describe on-chip SRAM or firmware regions outside system RAM"?

~Michal

> 
> Skip reserved-memory banks outside the host RAM banks reported by
> bootinfo_get_mem() in make_memory_node().
> 
> Fixes: 248faa637d2f ("xen/arm: add reserved-memory regions to the dom0 memory node")
> Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>
> ---
> Found while trying to bring up Xen on new hardware.
> Tested on real silicon.
> 
>  xen/common/device-tree/domain-build.c | 42 +++++++++++++++++++++++++--
>  1 file changed, 39 insertions(+), 3 deletions(-)
> 
> diff --git a/xen/common/device-tree/domain-build.c b/xen/common/device-tree/domain-build.c
> index 30a59abfa7..33baf150a6 100644
> --- a/xen/common/device-tree/domain-build.c
> +++ b/xen/common/device-tree/domain-build.c
> @@ -504,6 +504,34 @@ int __init make_chosen_node(const struct kernel_info *kinfo)
>      return res;
>  }
>  
> +static bool __init is_in_host_ram(const struct membank *bank)
> +{
> +    const struct membanks *ram = bootinfo_get_mem();
> +    paddr_t start = bank->start;
> +    paddr_t end = start + bank->size;
> +    unsigned int i;
> +
> +    if ( !bank->size || end < start )
> +        return false;
> +
> +    for ( i = 0; i < ram->nr_banks; i++ )
> +    {
> +        paddr_t ram_start = ram->bank[i].start;
> +        paddr_t ram_end = ram_start + ram->bank[i].size;
> +
> +        if ( start < ram_start )
> +            return false;
> +
> +        if ( end <= ram_end )
> +            return true;
> +
> +        if ( start < ram_end )
> +            start = ram_end;
> +    }
> +
> +    return false;
> +}
> +
>  int __init make_memory_node(const struct kernel_info *kinfo, int addrcells,
>                              int sizecells, const struct membanks *mem)
>  {
> @@ -513,14 +541,20 @@ int __init make_memory_node(const struct kernel_info *kinfo, int addrcells,
>      int nr_cells = 0;
>      __be32 reg[DT_MEM_NODE_REG_RANGE_SIZE];
>      __be32 *cells;
> +    bool is_resv = (mem == bootinfo_get_reserved_mem());
>  
>      if ( mem->nr_banks == 0 )
>          return -ENOENT;
>  
>      /* find the first memory range that is reserved for device (or firmware) */
> -    for ( i = 0; i < mem->nr_banks &&
> -                 (mem->bank[i].type != MEMBANK_DEFAULT); i++ )
> -        ;
> +    for ( i = 0; i < mem->nr_banks; i++ )
> +    {
> +        if ( mem->bank[i].type != MEMBANK_DEFAULT )
> +            continue;
> +        if ( is_resv && !is_in_host_ram(&mem->bank[i]) )
> +            continue;
> +        break;
> +    }
>  
>      if ( i == mem->nr_banks )
>          return 0;
> @@ -544,6 +578,8 @@ int __init make_memory_node(const struct kernel_info *kinfo, int addrcells,
>          if ( (mem->bank[i].type == MEMBANK_STATIC_DOMAIN) ||
>               (mem->bank[i].type == MEMBANK_FDT_RESVMEM) )
>              continue;
> +        if ( is_resv && !is_in_host_ram(&mem->bank[i]) )
> +            continue;
>  
>          nr_cells += reg_size;
>          BUG_ON(nr_cells > ARRAY_SIZE(reg));

Re: [PATCH] xen/dt: Skip non-RAM reserved-memory regions in make_memory_node()

[Kuan-Wei Chiu]

Hi Michal,

On Fri, Oct 30, 2026 at 12:04:58PM +0100, Orzel, Michal wrote:
> 
> 
> On 28-Oct-26 22:17, Kuan-Wei Chiu wrote:
> > Xen creates a second /memory node for Dom0 from
> > bootinfo_get_reserved_mem() without checking whether those regions
> > belong to host RAM. However, /reserved-memory can also describe
> > on-chip SRAM or firmware regions outside system RAM. Listing them in a
> > /memory node causes Dom0 to register non-RAM addresses as system RAM,
> > conflicting with device MMIO mappings.
> Hmm, I always thought that /reserved-memory is to exclude regions from memory
> described under /memory. That's what Xen assumes too. map_range_to_domain() for
> example does not call iomem_permit_access() for such regions because we assume
> it's RAM. Can you provide more details of your board e.g. the device tree? Is
> there any documentation I could use to confirm that "reserved-memory can also
> describe on-chip SRAM or firmware regions outside system RAM"?
> 
Hmm, your reply initially made me wonder if this was caused by a
non-compliant downstream device tree on our side, so I checked the
DT spec and upstream Linux.

The devicetree spec does not seem to explicitly mention whether
/reserved-memory regions must fall within the range described by
/memory. From a quick look at early_init_dt_reserve_memory() in Linux,
the comment there also mentions not worrying if a no-map region is not
memory as it will not be mapped.

In upstream Linux device trees, there are cases like
arch/mips/boot/dts/mobileye/eyeq5.dtsi and eyeq5-epm5.dts, where
/memory starts at 0x40000000 and 0x802000000, while /reserved-memory
defines an on-chip sram region nvram@461fe00 with "nvmem-rmem" outside
/memory.

Our board's device tree is not upstream yet, but it follows a similar
pattern where /reserved-memory defines "nvmem-rmem" and dma pool
sub-regions inside on-chip sram outside the dram range described by
/memory.

Regards,
Kuan-Wei