[PATCH 4/4] riscv: Improve kasan population by using hugepages when possible

From: Alexandre Ghiti <alex@ghiti.fr>
To: Andrey Ryabinin <aryabinin@virtuozzo.com>,
	Alexander Potapenko <glider@google.com>,
	Dmitry Vyukov <dvyukov@google.com>,
	Paul Walmsley <paul.walmsley@sifive.com>,
	Palmer Dabbelt <palmer@dabbelt.com>,
	Albert Ou <aou@eecs.berkeley.edu>,
	kasan-dev@googlegroups.com, linux-riscv@lists.infradead.org,
	linux-kernel@vger.kernel.org
Cc: Alexandre Ghiti <alex@ghiti.fr>
Subject: [PATCH 4/4] riscv: Improve kasan population by using hugepages when possible
Date: Mon,  8 Feb 2021 14:30:17 -0500	[thread overview]
Message-ID: <20210208193017.30904-5-alex@ghiti.fr> (raw)
In-Reply-To: <20210208193017.30904-1-alex@ghiti.fr>

The kasan functions that populates the shadow regions used to allocate them
page by page and did not take advantage of hugepages, so fix this by
trying to allocate hugepages of 1GB and fallback to 2MB hugepages or 4K
pages in case it fails.

This reduces the page table memory consumption and improves TLB usage,
as shown below:

Before this patch:

---[ Kasan shadow start ]---
0xffffffc000000000-0xffffffc400000000    0x00000000818ef000        16G PTE     . A . . . . R V
0xffffffc400000000-0xffffffc447fc0000    0x00000002b7f4f000   1179392K PTE     D A . . . W R V
0xffffffc480000000-0xffffffc800000000    0x00000000818ef000        14G PTE     . A . . . . R V
---[ Kasan shadow end ]---

After this patch:

---[ Kasan shadow start ]---
0xffffffc000000000-0xffffffc400000000    0x00000000818ef000        16G PTE     . A . . . . R V
0xffffffc400000000-0xffffffc440000000    0x0000000240000000         1G PGD     D A . . . W R V
0xffffffc440000000-0xffffffc447e00000    0x00000002b7e00000       126M PMD     D A . . . W R V
0xffffffc447e00000-0xffffffc447fc0000    0x00000002b818f000      1792K PTE     D A . . . W R V
0xffffffc480000000-0xffffffc800000000    0x00000000818ef000        14G PTE     . A . . . . R V
---[ Kasan shadow end ]---

Signed-off-by: Alexandre Ghiti <alex@ghiti.fr>
---
 arch/riscv/mm/kasan_init.c | 24 ++++++++++++++++++++++++
 1 file changed, 24 insertions(+)

diff --git a/arch/riscv/mm/kasan_init.c b/arch/riscv/mm/kasan_init.c
index b7d4d9abd144..2b196f512f07 100644
--- a/arch/riscv/mm/kasan_init.c
+++ b/arch/riscv/mm/kasan_init.c
@@ -83,6 +83,15 @@ static void kasan_populate_pmd(pgd_t *pgd, unsigned long vaddr, unsigned long en
 
 	do {
 		next = pmd_addr_end(vaddr, end);
+
+		if (pmd_none(*pmdp) && IS_ALIGNED(vaddr, PMD_SIZE) && (next - vaddr) >= PMD_SIZE) {
+			phys_addr = memblock_phys_alloc(PMD_SIZE, PMD_SIZE);
+			if (phys_addr) {
+				set_pmd(pmdp, pfn_pmd(PFN_DOWN(phys_addr), PAGE_KERNEL));
+				continue;
+			}
+		}
+
 		kasan_populate_pte(pmdp, vaddr, next);
 	} while (pmdp++, vaddr = next, vaddr != end);
 
@@ -103,6 +112,21 @@ static void kasan_populate_pgd(unsigned long vaddr, unsigned long end)
 
 	do {
 		next = pgd_addr_end(vaddr, end);
+
+		/*
+		 * pgdp can't be none since kasan_early_init initialized all KASAN
+		 * shadow region with kasan_early_shadow_pmd: if this is stillthe case,
+		 * that means we can try to allocate a hugepage as a replacement.
+		 */
+		if (pgd_page_vaddr(*pgdp) == (unsigned long)lm_alias(kasan_early_shadow_pmd) &&
+		    IS_ALIGNED(vaddr, PGDIR_SIZE) && (next - vaddr) >= PGDIR_SIZE) {
+			phys_addr = memblock_phys_alloc(PGDIR_SIZE, PGDIR_SIZE);
+			if (phys_addr) {
+				set_pgd(pgdp, pfn_pgd(PFN_DOWN(phys_addr), PAGE_KERNEL));
+				continue;
+			}
+		}
+
 		kasan_populate_pmd(pgdp, vaddr, next);
 	} while (pgdp++, vaddr = next, vaddr != end);
 }
-- 
2.20.1

