Skip to contentSkip to conversation

X86-32: Allocate 256 bytes for pgd in PAE paging

PATCH Superseded Fenghua Yu 4 participants 7 messages Dec 17, 2014 – Dec 18, 2014

Subsystems (2)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Fenghua Yu

author
#0 lore raw

PATCH v2 1/1 X86-32: Allocate 256 bytes for pgd in PAE paging

From: Fenghua Yu <[email protected]>

X86 32-bit machine and kernel use PAE paging, which currently wastes about 4K of memory per process on Linux where we have to reserve an entire page to support a single 256-byte PGD structure. It would be a very good thing if we could eliminate that wastage.

---
 arch/x86/mm/pgtable.c | 42 +++++++++++++++++++++++++++++++++++++++---
 1 file changed, 39 insertions(+), 3 deletions(-)
arch/x86/mm/pgtable.c +39 -3
index 6fb6927..695db92 100644
@@ -1,5 +1,6 @@
 #include <linux/mm.h>
 #include <linux/gfp.h>
+#include <linux/slab.h>
 #include <asm/pgalloc.h>
 #include <asm/pgtable.h>
 #include <asm/tlb.h>
@@ -271,12 +272,46 @@ static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
 	}
 }
 
+/*
+ * Xen paravirt assumes pgd table should be in one page. pgd in 64 bit also
+ * needs to be in one page.
+ *
+ * But PAE without Xen only needs to allocate 256 bytes for pgd.
+ *
+ * So if kernel is compiled as PAE model without Xen, we allocate 256 bytes
+ * for pgd entries to save memory space.
+ *
+ * In other cases, one page is allocated for pgd. In theory, a kernel
+ * in PAE mode not running in Xen could allocate 256 bytes for pgd
+ * as well. But that will make the allocation and free more complex
+ * but not useful in reality. To simplify the code and testing, we just
+ * allocate one page when CONFIG_XEN is enabled regardelss kernel is running
+ * in Xen or not.
+ */
+static inline pgd_t *_pgd_alloc(void)
+{
+#if defined(CONFIG_X86_PAE) && !defined(CONFIG_XEN)
+	return kmalloc(sizeof(pgdval_t) * PTRS_PER_PGD, PGALLOC_GFP);
+#else
+	return (pgd_t *)__get_free_page(PGALLOC_GFP);
+#endif
+}
+
+static inline void _pgd_free(pgd_t *pgd)
+{
+#if defined(CONFIG_X86_PAE) && !defined(CONFIG_XEN)
+	kfree(pgd);
+#else
+	free_page((unsigned long)pgd);
+#endif
+}
+
 pgd_t *pgd_alloc(struct mm_struct *mm)
 {
 	pgd_t *pgd;
 	pmd_t *pmds[PREALLOCATED_PMDS];
 
-	pgd = (pgd_t *)__get_free_page(PGALLOC_GFP);
+	pgd = _pgd_alloc();
 
 	if (pgd == NULL)
 		goto out;
@@ -306,7 +341,7 @@ pgd_t *pgd_alloc(struct mm_struct *mm)
 out_free_pmds:
 	free_pmds(pmds);
 out_free_pgd:
-	free_page((unsigned long)pgd);
+	_pgd_free(pgd);
 out:
 	return NULL;
 }
@@ -316,7 +351,8 @@ void pgd_free(struct mm_struct *mm, pgd_t *pgd)
 	pgd_mop_up_pmds(mm, pgd);
 	pgd_dtor(pgd);
 	paravirt_pgd_free(mm, pgd);
-	free_page((unsigned long)pgd);
+	_pgd_free(pgd);
+
 }
 
 /*
Signature · 1 line
1.8.1.2
#1 lore raw

arch/x86/mm/pgtable.c static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[]) · line 298

Quoted 8 lines · Fenghua Yu
arch/x86/mm/pgtable.c +8 -0
+static inline pgd_t *_pgd_alloc(void)
+{
+#if defined(CONFIG_X86_PAE) && !defined(CONFIG_XEN)
+	return kmalloc(sizeof(pgdval_t) * PTRS_PER_PGD, PGALLOC_GFP);
+#else
+	return (pgd_t *)__get_free_page(PGALLOC_GFP);
+#endif
+}

I'm looking at:

"Figure 4-7. Formats of CR3 and Paging-Structure Entries with
PAE Paging"

in the SDM. It makes it pretty clear that the lower 5 bits of cr3 are ignored in PAE mode. That means we have to be 32-byte (or greater) aligned, right? Does kmalloc() guarantee that?

IOW, do *ALL* of the sl*b allocators in all of their forms with all of their debugging options guarantee 32-byte alignment when allocating 256-byte objects?

I know we at least try to align to a cacheline, which would be good enough, but I'm fuzzy on what we *guarantee*.

#2 lore raw
Quoted 3 lines · Dave Hansen

IOW, do *ALL* of the sl*b allocators in all of their forms with all of their debugging options guarantee 32-byte alignment when allocating 256-byte objects?

No. For that the arch has to set a macro call ARCH_KMALLOC_MINALIGN or ARCH_DMA_MINALIGN. Default alignment is to a word boundary.

Quoted 2 lines

I know we at least try to align to a cacheline, which would be good enough, but I'm fuzzy on what we *guarantee*.

Sorry we only do this if requested for a slab cache.

#3 lore raw
Quoted 6 lines · Christoph Lameter
Quoted 3 lines · Dave Hansen
Quoted 3 lines

IOW, do *ALL* of the sl*b allocators in all of their forms with all of their debugging options guarantee 32-byte alignment when allocating 256-byte objects?

No. For that the arch has to set a macro call ARCH_KMALLOC_MINALIGN or ARCH_DMA_MINALIGN. Default alignment is to a word boundary.

OK, sounds like this is going to need its own slab. It will make the patch a bit bigger but shouldn't be all that much more complicated.

#4 lore raw
In reply to Fenghua Yu
Quoted 7 lines · Fenghua Yu

From: Fenghua Yu <[email protected]>

X86 32-bit machine and kernel use PAE paging, which currently wastes about 4K of memory per process on Linux where we have to reserve an entire page to support a single 256-byte PGD structure. It would be a very good thing if we could eliminate that wastage.

4*8 = 32 bytes, where did 256 bytes come from?

-hpa

#5 lore raw
Quoted 9 lines

Sent: Thursday, December 18, 2014 9:26 AM
To: Yu, Fenghua; Thomas Gleixner; H. Peter Anvin; Ingo Molnar; Williamson, Glenn P
Cc: linux-kernel; x86
Subject: Re: [PATCH v2] X86-32: Allocate 256 bytes for pgd in PAE paging

Quoted 2 lines · Fenghua Yu

From: Fenghua Yu <[email protected]>

Quoted 7 lines
Quoted 5 lines

X86 32-bit machine and kernel use PAE paging, which currently wastes about 4K of memory per process on Linux where we have to reserve an entire page to support a single 256-byte PGD structure. It would be a very good thing if we could eliminate that wastage.

4*8 = 32 bytes, where did 256 bytes come from?

You are right. It should be 32 bytes. I will change the wording in future patch. The real calculation in code is right though.

Thanks.

-Fenghua

#6 lore raw
Quoted 3 lines
Quoted 1 line

4*8 = 32 bytes, where did 256 bytes come from?

You are right. It should be 32 bytes. I will change the wording in future patch. The real calculation in code is right though.

I don't know if it makes sense to round up to a cache line. My suspicion is that it won't matter, as these fields will be read-mostly, on the other hand, wasting 32 bytes isn't exactly a problem, either.

-hpa