Re: [PATCH] lib/firmware_table: validate subtable bounds

[Dave Jiang <dave.jiang@intel.com>]

On 10/16/26 4:15 PM, Andrew Morton wrote:
> On Fri, 16 Oct 2026 22:41:54 +0800 Jiale Yao <yaojiale02@163.com> wrote:
> 
>> The generic firmware table walker dispatches matching entries after only
>> checking that a generic subtable header is present. A record whose
>> declared length is shorter than its header or extends past the table can
>> therefore reach a handler before its bounds have been established.
>>
>> Validate the complete generic header and the declared record range
>> before looking up or invoking a handler. This also rejects zero-length
>> records and prevents malformed lengths from advancing the cursor outside
>> the table.
> 
> Thanks.
> 
> get_maintainer is quite unreliable with lib/ files.  Please always use
> git-blame when determining who should be cc'ed.
> 
> Dave, can you please take a look?
> 
>> --- a/lib/fw_table.c
>> +++ b/lib/fw_table.c
>> @@ -172,7 +172,20 @@ acpi_parse_entries_array(char *id, unsigned long table_size,
>>  	    ((unsigned long)table_header + table_size);
>>  	subtable_len = acpi_get_subtable_header_length(&entry);
>>  
>> -	while (((unsigned long)entry.hdr) + subtable_len < table_end) {
>> +	while ((unsigned long)entry.hdr < table_end) {
>> +		if (table_end - (unsigned long)entry.hdr < subtable_len) {
>> +			pr_err("[%4.4s] Invalid subtable header\n", id);
>> +			return -EINVAL;
>> +		}
>> +
>> +		entry_len = acpi_get_entry_length(&entry);
>> +		if (entry_len < subtable_len ||
>> +		    entry_len > table_end - (unsigned long)entry.hdr) {
>> +			pr_err("[%4.4s:0x%02lx] Invalid length %lu\n", id,
>> +			       acpi_get_entry_type(&entry), entry_len);
>> +			return -EINVAL;
>> +		}
>> +
>>  		for (i = 0; i < proc_num; i++) {
>>  			if (acpi_get_entry_type(&entry) != proc[i].id)
>>  				continue;
>> @@ -186,16 +199,6 @@ acpi_parse_entries_array(char *id, unsigned long table_size,
>>  			break;
>>  		}
>>  
>> -		/*
>> -		 * If entry->length is 0, break from this loop to avoid
>> -		 * infinite loop.
>> -		 */
>> -		entry_len = acpi_get_entry_length(&entry);
>> -		if (entry_len == 0) {
>> -			pr_err("[%4.4s:0x%02x] Invalid zero length\n", id, proc->id);
>> -			return -EINVAL;
>> -		}
>> -
>>  		entry.hdr = (union acpi_subtable_headers *)
>>  		    ((unsigned long)entry.hdr + entry_len);
>>  	}
>> -- 
>> 2.34.1

Can this turn a previously accepted table into a fatal parse error? My
main concern is breaking previously tolerated behavior and we end up
breaking deployed systems. Given this code has been around a long time.
Maybe Rafael or Len can comment.

The old loop stopped quietly when subtable_len or fewer bytes remained
after the last entry. Here subtable_len is sizeof() the subtable header
structure, 2 bytes for MADT or 4 for CDAT, not the entry's length field.

Now fewer than subtable_len leftover bytes return -EINVAL. Exactly
subtable_len leftover bytes are read as a header, and fail the entry_len
check below unless they happen to hold a valid one. A table with a few
bytes of padding after the last entry now fails the whole parse.

On x86 that error is not contained to one entry type:

early_acpi_process_madt()
  early_acpi_parse_madt_lapic_addr_ovr()
    acpi_table_parse_madt(ACPI_MADT_TYPE_LOCAL_APIC_OVERRIDE, ...)
      -> -EINVAL

	if (error == -EINVAL) {
		...
		pr_err("Invalid BIOS MADT, disabling ACPI\n");
		disable_acpi();
	}

Would it be safer to keep tolerating leftover bytes too small to hold a
subtable header, and reject only entries whose declared length is
invalid? The loop condition can end the walk in that case without a
separate check:

	while ((unsigned long)entry.hdr + subtable_len <= table_end) {
		entry_len = acpi_get_entry_length(&entry);
		if (entry_len < subtable_len ||
		    entry_len > table_end - (unsigned long)entry.hdr) {
			...
			return -EINVAL;
		}