[AMD Official Use Only - General]
Hello Boris,
>> There is 1 64-bit RMP entry for every physical 4k page of DRAM, so
>> essentially every 4K page of DRAM is represented by a RMP entry.
>Before we get to the rest - this sounds wrong to me. My APM has:
>"PSMASH Page Smash
>Expands a 2MB-page RMP entry into a corresponding set of contiguous 4KB-page RMP entries. The 2MB page's system physical address is specified in the RAX register. The new entries inherit the attributes of the original entry. Upon completion, a >return code is stored in EAX.
>rFLAGS bits OF, ZF, AF, PF and SF are set based on this return code..."
>So there *are* 2M entries in the RMP table.
> So even if host page is 1G and underlying (smashed/split) RMP entries
> are 2M, the RMP table entry has to be indexed to a 4K entry
> corresponding to that.
>So if there are 2M entries in the RMP table, how does that indexing with 4K entries is supposed to work?
>Hell, even PSMASH pseudocode shows how you go and write all those 512 4K entries using the 2M entry as a template. So *before* you have smashed that 2M entry, it *is* an *actual* 2M entry.
>So if you fault on a page which is backed by that 2M RMP entry, you will get that 2M RMP entry.
> If it was simply a 2M entry in the RMP table, then pmd_index() will
> work correctly.
>Judging by the above text, it *can* *be* a 2M RMP entry!
>By reading your example you're trying to tell me that a RMP #PF will always need to work on 4K entries. Which would then need for a 2M entry as above to be PSMASHed in order to get the 4K thing. But that would be silly - RMP PFs will this way >gradually break all 2M pages and degrage performance for no real reason.
>So this still looks real wrong to me.
Please note that RMP table entries have only 2 page size indicators 4k and 2M, so it covers a max physical address range of 2MB.
In all cases, there is one RMP entry per 4K page and the index into the RMP table is basically address /PAGE_SIZE, and that does
not change for hugepages. Therefore we need to capture the address bits (from address) so that we index into the
4K entry in the RMP table.
An important point to note here is that RMPUPDATE instruction sets the Assigned bit for all the sub-page entries for
a hugepage mapping in RMP table, so we will get the correct "assigned" page information when we index into the 4K entry
in the RMP table and additionally, __snp_lookup_rmpentry() gets the 2MB aligned entry in the RMP table to get the correct Page size.
(as below)
static struct rmpentry *__snp_lookup_rmpentry(u64 pfn, int *level)
{
..
/* Read a large RMP entry to get the correct page level used in RMP entry. */
large_entry = rmptable_entry(paddr & PMD_MASK);
*level = RMP_TO_X86_PG_LEVEL(rmpentry_pagesize(large_entry));
..
Therefore, the 2M entry and it's subpages in the RMP table will always exist because of the RMPUPDATE instruction even
without smashing/splitting of the hugepage, so we really don't need the 2MB entry to be PSMASHed in order to get the 4K thing.
Thanks,
Ashish
