Re: [PATCH v8 4/4] hugetlb: allow to free gigantic pages regardless of the configuration

From: Alexandre Ghiti
Date: Wed Mar 27 2019 - 05:49:24 EST


On 03/27/2019 09:55 AM, Aneesh Kumar K.V wrote:
On 3/27/19 2:14 PM, Alexandre Ghiti wrote:


On 03/27/2019 08:01 AM, Aneesh Kumar K.V wrote:
On 3/27/19 12:06 PM, Alexandre Ghiti wrote:
On systems without CONTIG_ALLOC activated but that support gigantic pages,
boottime reserved gigantic pages can not be freed at all. This patch
simply enables the possibility to hand back those pages to memory
allocator.

Signed-off-by: Alexandre Ghiti <alex@xxxxxxxx>
Acked-by: David S. Miller <davem@xxxxxxxxxxxxx> [sparc]

diff --git a/arch/powerpc/include/asm/book3s/64/hugetlb.h b/arch/powerpc/include/asm/book3s/64/hugetlb.h
index ec2a55a553c7..7013284f0f1b 100644
--- a/arch/powerpc/include/asm/book3s/64/hugetlb.h
+++ b/arch/powerpc/include/asm/book3s/64/hugetlb.h
@@ -36,8 +36,8 @@ static inline int hstate_get_psize(struct hstate *hstate)
ÂÂÂÂÂ }
 }
 -#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
-static inline bool gigantic_page_supported(void)
+#define __HAVE_ARCH_GIGANTIC_PAGE_RUNTIME_SUPPORTED
+static inline bool gigantic_page_runtime_supported(void)
 {
ÂÂÂÂÂ /*
ÂÂÂÂÂÂ * We used gigantic page reservation with hypervisor assist in some case.
@@ -49,7 +49,6 @@ static inline bool gigantic_page_supported(void)
 Â return true;
 }
-#endif
  /* hugepd entry valid bit */
 #define HUGEPD_VAL_BITS (0x8000000000000000UL)

Is that correct when CONTIG_ALLOC is not enabled? I guess we want

gigantic_page_runtime_supported to return false when CONTIG_ALLOC is not enabled on all architectures and on POWER when it is enabled we want it to be conditional as it is now.

-aneesh


CONFIG_ARCH_HAS_GIGANTIC_PAGE is set by default when an architecture supports gigantic
pages: on its own, it allows to allocate boottime gigantic pages AND to free them at runtime
(this is the goal of this series), but not to allocate runtime gigantic pages.
If CONTIG_ALLOC is set, it allows in addition to allocate runtime gigantic pages.

I re-introduced the runtime checks because we can't know at compile time if powerpc can
or not support gigantic pages.

So for all architectures, gigantic_page_runtime_supported only depends on
CONFIG_ARCH_HAS_GIGANTIC_PAGE enabled or not. The possibility to allocate runtime
gigantic pages is dealt with after those runtime checks.


you removed that #ifdef in the patch above. ie we had
#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
static inline bool gigantic_page_supported(void)
{
ÂÂÂÂ/*
ÂÂÂÂ * We used gigantic page reservation with hypervisor assist in some case.
ÂÂÂÂ * We cannot use runtime allocation of gigantic pages in those platforms
ÂÂÂÂ * This is hash translation mode LPARs.
ÂÂÂÂ */
ÂÂÂÂif (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
ÂÂÂÂÂÂÂ return false;

ÂÂÂÂreturn true;
}
#endif

Yes, I removed the #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE because it was defined only
if CONTIG_ALLOC was set. But now, CONFIG_ARCH_HAS_GIGANTIC_PAGE is inconditionally
set for powerpc so I think we don't need it anymore.
Actually I have doubts now, is this true for all configurations ? I see that it is only set for
PPC_RADIX_MMU. I think the problem is here: instead of returning true, it should do like
the generic version, ie return IS_ENABLED(CONFIG_ARCH_HAS_GIGANTIC_PAGE).
Do you agree ?



This is now
#define __HAVE_ARCH_GIGANTIC_PAGE_RUNTIME_SUPPORTED
static inline bool gigantic_page_runtime_supported(void)
{
if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
ÂÂÂÂÂÂÂ return false;

ÂÂÂÂreturn true;
}


I am wondering whether it should be

#define __HAVE_ARCH_GIGANTIC_PAGE_RUNTIME_SUPPORTED
static inline bool gigantic_page_runtime_supported(void)
{

ÂÂ if (!IS_ENABLED(CONFIG_CONTIG_ALLOC))
ÂÂÂÂÂÂÂ return false;

I don't think this test should happen here, CONFIG_CONTIG_ALLOC only allows
to allocate gigantic pages, doing that check here would prevent powerpc
to free boottime gigantic pages when not a guest. Note that this check
is actually done in set_max_huge_pages.



if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
ÂÂÂÂÂÂÂ return false;

Maybe I did not understand this check: I understood that, in the case the system
is virtualized, we do not want it to hand back gigantic pages. Does this check
test if the system is currently being virtualized ?
If yes, I think the patch is correct: it prevents freeing gigantic pages when the system
is virtualized but allows a 'normal' system to free gigantic pages.



ÂÂÂÂreturn true;
}

or add that #ifdef back.

By the way, I forgot to ask you why you think that if an arch cannot allocate runtime gigantic
pages, it should not be able to free boottime gigantic pages ?


on virtualized platforms like powervm which use a paravirtualized page table update mechanism (we dont' have two level table). The ability to map a page huge depends on how hypervisor allocated the guest ram. Hypervisor also allocates the guest specific page table of a specific size depending on how many pages are going to be mapped by what page size.

on POWER we indicate possible guest real address that can be mapped via hugepage (in this case 16G) using a device tree node (ibm,expected#pages) . It is expected that we will map these pages only as 16G pages. Hence we cannot free them back to the buddy where it could get mapped via 64K page size.

Thanks for the explanations.

Alex

-aneesh