Re: [PATCH 4/6] KVM: selftests: memslot_perf_test: Support variable guest page size
From: Gavin Shan
Date: Thu Oct 20 2022 - 03:20:19 EST
On 10/20/22 4:18 AM, Maciej S. Szmigiero wrote:
On 19.10.2022 02:26, Gavin Shan wrote:
On 10/18/22 11:56 PM, Maciej S. Szmigiero wrote:
On 18.10.2022 02:51, Gavin Shan wrote:
On 10/18/22 8:46 AM, Gavin Shan wrote:
On 10/18/22 5:31 AM, Maciej S. Szmigiero wrote:
On 14.10.2022 09:19, Gavin Shan wrote:
The test case is obviously broken on aarch64 because non-4KB guest
page size is supported. The guest page size on aarch64 could be 4KB,
16KB or 64KB.
This supports variable guest page size, mostly for aarch64.
- The host determines the guest page size when virtual machine is
created. The value is also passed to guest through the synchronization
area.
- The number of guest pages are unknown until the virtual machine
is to be created. So all the related macros are dropped. Instead,
their values are dynamically calculated based on the guest page
size.
- The static checks on memory sizes and pages becomes dependent
on guest page size, which is unknown until the virtual machine
is about to be created. So all the static checks are converted
to dynamic checks, done in check_memory_sizes().
- As the address passed to madvise() should be aligned to host page,
the size of page chunk is automatically selected, other than one
page.
- All other changes included in this patch are almost mechanical
replacing '4096' with 'guest_page_size'.
Signed-off-by: Gavin Shan <gshan@xxxxxxxxxx>
---
.../testing/selftests/kvm/memslot_perf_test.c | 191 +++++++++++-------
1 file changed, 115 insertions(+), 76 deletions(-)
diff --git a/tools/testing/selftests/kvm/memslot_perf_test.c b/tools/testing/selftests/kvm/memslot_perf_test.c
index d5aa9148f96f..d587bd952ff9 100644
--- a/tools/testing/selftests/kvm/memslot_perf_test.c
+++ b/tools/testing/selftests/kvm/memslot_perf_test.c
(...)
@@ -77,8 +61,7 @@ static_assert(MEM_TEST_UNMAP_SIZE_PAGES %
* for the total size of 25 pages.
* Hence, the maximum size here is 50 pages.
*/
-#define MEM_TEST_MOVE_SIZE_PAGES (50)
-#define MEM_TEST_MOVE_SIZE (MEM_TEST_MOVE_SIZE_PAGES * 4096)
+#define MEM_TEST_MOVE_SIZE 0x32000
The above number seems less readable than an explicit value of 50 pages.
In addition to that, it's 50 pages only with 4k page size, so at least
the comment above needs to be updated to reflect this fact.
Yeah, I will change the comments like below in next revision.
/*
* When running this test with 32k memslots, actually 32763 excluding
* the reserved memory slot 0, the memory for each slot is 0x4000 bytes.
* The last slot contains 0x19000 bytes memory. Hence, the maximum size
* here is 0x32000 bytes.
*/
I will replace those numbers with readable ones like below :)
/*
* When running this test with 32k memslots, actually 32763 excluding
* the reserved memory slot 0, the memory for each slot is 16KB. The
* last slot contains 100KB memory with the remaining 84KB. Hence,
* the maximum size is double of that (200KB)
*/
Still, these numbers are for x86, which has KVM_INTERNAL_MEM_SLOTS
defined as 3.
As far as I can see aarch64 has KVM_INTERNAL_MEM_SLOTS equal to 0, so
this arch has 32766 slot available for the test memory.
Quick calculations show that this will result in 112 KiB of memory in
the last slot for 4 KiB page size (while for 64 KiB page size the
maximum slot count for this test is 8192 anyway - not counting slot 0).
It seems your calculation had (512MB+64KB), instead of (512MB+4KB).
In this particular patch, we still have (512MB+4KB). How about to change
like below in this patch. In next patch, it's adjusted accordingly after
we have (512MB+64KB).
My review comment above referred to the final MEM_SIZE value after the
whole series, so 512 MiB + 64 KiB.
I placed that review comment on patch 4 since it's the only patch in this
series that modified the code comment about MEM_TEST_MOVE_SIZE.
(1) In this patch, the comment is changed to as below
/*
* We have different number of memory slots, excluding the reserved
* memory slot 0, on various architectures and configurations. The
* memory size in this test is calculated by doubling the maximal
* memory size in last memory slot, with alignment to the largest
* supported page size (64KB).
*
* architecture slots memory-per-slot memory-on-last-slot
* --------------------------------------------------------------
* x86-4KB 32763 16KB 100KB
* arm64-4KB 32766 16KB 52KB
* arm64-64KB 8192 64KB 64KB
*/
#define MEM_TEST_MOVE_SIZE 0x40000 /* 256KB */
(2) In the next patch, where we have (512MB+64KB) after the various
memory sizes are consolidated, It is adjusted accordingly as below.
/*
* We have different number of memory slots, excluding the reserved
* memory slot 0, on various architectures and configurations. The
* memory size in this test is calculated by doubling the maximal
* memory size in last memory slot, with alignment to the largest
* supported page size (64KB).
*
* architecture slots memory-per-slot memory-on-last-slot
* --------------------------------------------------------------
* x86-4KB 32763 16KB 160KB
* arm64-4KB 32766 16KB 112KB
* arm64-64KB 8192 64KB 128KB
*/
#define MEM_TEST_MOVE_SIZE 0x50000 /* 320KB */
Now MEM_TEST_MOVE_SIZE is too high for arm64-4KB and arm64-64KB cases
(it needs 160 KiB in the last slot but has less available in these two
cases).
Using a test size of 192 KiB instead seems like a small difference
from the original size of 200 KiB, while still being aligned to
64 KiB.
The move benchmarks runtime difference on x86-4KB with this size
(compared to sizes of 200 KiB and 320 KiB) seems to be negligible.
Since it's an odd number of 64 KiB pages (3) the code that halves
this number of pages will need to be adjusted to operate on raw
sizes instead.
I can see a single block of code that will need such adjustment:
if (lastpages < move_pages / 2) {
*maxslots = 0;
return false;
}
Similar remark goes for the case (1) above, where you'll probably need
to use 64 KiB test area size (it's only an intermediate form of code
before the final patch changes this value so it's fine if it doesn't
perform as well as the final form of the code).
Maciej, all your comments make sense to me. It really took me some times
to do the calculation. I just posted v3 to address all your comments.
Hopefully, there is nothing missed. Please go ahead to review v3 directly
when you get a chance.
v3: https://lore.kernel.org/kvmarm/20221020071209.559062-1-gshan@xxxxxxxxxx/T/#t
In v3, the comments about MEM_TEST_MOVE_SIZE is fixed in PATCH[v3 4/6],
but it's 64KB. In PATCH[v3 5/6], it's fixed up to 192KB and memory size
is used for the comparison in test_memslot_move_prepare().
Thanks,
Gavin