Re: [PATCH 3/5] Extend e820_table to hold information about memory encryption

[Dave Hansen <dave.hansen@intel.com>]

On 11/8/21 10:40 AM, Martin Fernandez wrote:
> On 11/5/21, Dave Hansen <dave.hansen@intel.com> wrote:
>> On 11/5/21 2:27 PM, Martin Fernandez wrote:
>>> +void __init e820__mark_regions_as_crypto_capable(u64 start, u64 size)
>>> +{
>>> +	int i;
>>> +	u64 end = start + size;
>>> +
>>> +	for (i = 0; i < e820_table->nr_entries; i++) {
>>> +		struct e820_entry *const entry = &e820_table->entries[i];
>>> +
>>> +		if (entry->addr >= start && entry->addr + entry->size <= end)
>>> +			entry->crypto_capable = true;
>>> +	}
>>> +}
>>
>> Looking at this in isolation, this is really tricky.  I have no idea
>> what this is _supposed_ to or expected to be doing.  It also makes me
>> wonder what happens when start/size don't line up exactly on an e820 entry.
> 
> Do you think it's better to just add new entries, just as they are in
> the EFI memmap and then let e820__update_table handle them?
> 
> Although, as it is it's faster, the other way would be clearer in the
> code (since efi_mark_e820_regions_as_crypto_capable in part 4/5 isn't
> also the nicest of the functions and with this change it would be very
> straightforward), but it would require one e820__update_table. Also,
> it would more accurate, since if you call this with a start and size
> that doesn't cover at least one e820_entry then it will do nothing.

I was actually trying to make a comment about this function's lack of
documentation.

But, you make a good point about the alternate approach.  I don't have a
preference either way.


>>> @@ -327,6 +343,8 @@ int __init e820__update_table(struct e820_table
>>> *table)
>>>  	unsigned long long last_addr;
>>>  	u32 new_nr_entries, overlap_entries;
>>>  	u32 i, chg_idx, chg_nr;
>>> +	bool current_crypto;
>>> +	bool last_crypto = false;
>>>
>>>  	/* If there's only one memory region, don't bother: */
>>>  	if (table->nr_entries < 2)
>>> @@ -388,13 +406,17 @@ int __init e820__update_table(struct e820_table
>>> *table)
>>>  		 * 1=usable, 2,3,4,4+=unusable)
>>>  		 */
>>>  		current_type = 0;
>>> +		current_crypto = false;
>>>  		for (i = 0; i < overlap_entries; i++) {
>>> +			current_crypto = current_crypto || overlap_list[i]->crypto_capable;
>>
>> No comment, eh?
>>
>> This seems backwards to me.  If there are overlapping region and only
>> one is crypto-capable, shouldn't the whole thing become non-crypto-capable?
> 
> The reason for that is that right now, if a region is mark as
> crypto_capable is because EFI memmap says so, and again, right now
> that's the only source to fill the crypto_capable value, so I have to
> "believe" it. Now that I think about it, yes I should have a least put
> a comment on it.

My concern was if:

	current_crypto=0
and
	overlap_list[i]->crypto_capable=1

Doesn't that mean a non-crypto entry is being parsed, but current_crypto
will end up as 1?

>>>  		/* Continue building up new map based on this information: */
>>> -		if (current_type != last_type || e820_nomerge(current_type)) {
>>> +		if (current_type != last_type ||
>>> +		    current_crypto != last_crypto ||
>>> +		    e820_nomerge(current_type)) {
>>>  			if (last_type != 0)	 {
>>>  				new_entries[new_nr_entries].size = change_point[chg_idx]->addr -
>>> last_addr;
>>>  				/* Move forward only if the new size was non-zero: */
>>> @@ -406,6 +428,9 @@ int __init e820__update_table(struct e820_table
>>> *table)
>>>  			if (current_type != 0)	{
>>>  				new_entries[new_nr_entries].addr = change_point[chg_idx]->addr;
>>>  				new_entries[new_nr_entries].type = current_type;
>>> +				new_entries[new_nr_entries].crypto_capable = current_crypto;
>>> +
>>> +				last_crypto = current_crypto;
>>>  				last_addr = change_point[chg_idx]->addr;
>>>  			}
>>>  			last_type = current_type;
>>
>> The "current_crypto != last_crypto" checks seem to go with the
>> current_type/last_type checks.  I'm naively surprised that the
>> last_crypto assignment wasn't paired with the last_type assignment.
>>
>> I kinda get the impression this was just quickly hacked in here.  It
>> seems like "crypto" and "type" are very closely related in how they are
>> being handled.  It's a shame they're not being managed in a common way.
> 
> Yes, "crypto" and "type" seems really close, but to be honest, this
> function has a very weird flow, or something that I couldn't
> completely understand. After a while thinking about it I came up with
> that.
> 
> Again, this function is a pain, but I'll dedicate it some more time to
> see if I can come up with something better.

Please do.  Like you said, there's probably more context that it would
be helpful to understand.