linux-kernel - [PATCH] arm64: asid: Optimize cache

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Message-Id: <1561305869-18872-1-git-send-email-guoren@kernel.org>
Date:   Mon, 24 Jun 2019 00:04:29 +0800
From:   guoren@...nel.org
To:     julien.grall@....com, arnd@...db.de, linux-kernel@...r.kernel.org
Cc:     linux-csky@...r.kernel.org, Guo Ren <ren_guo@...ky.com>,
        Catalin Marinas <catalin.marinas@....com>
Subject: [PATCH] arm64: asid: Optimize cache_flush for SMT

From: Guo Ren <ren_guo@...ky.com>

The hardware threads of one core could share the same TLB for SMT+SMP
system. Assume hardware threads number sequence like this:

| 0 1 2 3 | 4 5 6 7 | 8 9 a b | c d e f |
   core1     core2     core3     core4

Current algorithm seems is correct for SMT+SMP, but it'll give some
duplicate local_tlb_flush. Because one hardware threads local_tlb_flush
will also flush other hardware threads' TLB entry in one core TLB.

So we can use bitmap to reduce local_tlb_flush for SMT.

C-SKY cores don't support SMT and the patch is no benefit for C-SKY.

Signed-off-by: Guo Ren <ren_guo@...ky.com>
Cc: Catalin Marinas <catalin.marinas@....com>
Cc: Julien Grall <julien.grall@....com>
---
 arch/csky/include/asm/asid.h |  4 ++++
 arch/csky/mm/asid.c          | 11 ++++++++++-
 arch/csky/mm/context.c       |  2 +-
 3 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/arch/csky/include/asm/asid.h b/arch/csky/include/asm/asid.h
index ac08b0f..f654492 100644
--- a/arch/csky/include/asm/asid.h
+++ b/arch/csky/include/asm/asid.h
@@ -23,6 +23,9 @@ struct asid_info
 	unsigned int		ctxt_shift;
 	/* Callback to locally flush the context. */
 	void			(*flush_cpu_ctxt_cb)(void);
+	/* To reduce duplicate tlb_flush for SMT */
+	unsigned int		harts_per_core;
+	unsigned int		harts_per_core_mask;
 };
 
 #define NUM_ASIDS(info)			(1UL << ((info)->bits))
@@ -73,6 +76,7 @@ static inline void asid_check_context(struct asid_info *info,
 
 int asid_allocator_init(struct asid_info *info,
 			u32 bits, unsigned int asid_per_ctxt,
+			unsigned int harts_per_core,
 			void (*flush_cpu_ctxt_cb)(void));
 
 #endif
diff --git a/arch/csky/mm/asid.c b/arch/csky/mm/asid.c
index b2e9147..50a983e 100644
--- a/arch/csky/mm/asid.c
+++ b/arch/csky/mm/asid.c
@@ -148,8 +148,13 @@ void asid_new_context(struct asid_info *info, atomic64_t *pasid,
 		atomic64_set(pasid, asid);
 	}
 
-	if (cpumask_test_and_clear_cpu(cpu, &info->flush_pending))
+	if (cpumask_test_cpu(cpu, &info->flush_pending)) {
+		unsigned int i;
+		unsigned int harts_base = cpu & info->harts_per_core_mask;
 		info->flush_cpu_ctxt_cb();
+		for (i = 0; i < info->harts_per_core; i++)
+			cpumask_clear_cpu(harts_base + i, &info->flush_pending);
+	}
 
 	atomic64_set(&active_asid(info, cpu), asid);
 	cpumask_set_cpu(cpu, mm_cpumask(mm));
@@ -162,15 +167,19 @@ void asid_new_context(struct asid_info *info, atomic64_t *pasid,
  * @info: Pointer to the asid allocator structure
  * @bits: Number of ASIDs available
  * @asid_per_ctxt: Number of ASIDs to allocate per-context. ASIDs are
+ * @harts_per_core: Number hardware threads per core, must be 1, 2, 4, 8, 16 ...
  * allocated contiguously for a given context. This value should be a power of
  * 2.
  */
 int asid_allocator_init(struct asid_info *info,
 			u32 bits, unsigned int asid_per_ctxt,
+			unsigned int harts_per_core,
 			void (*flush_cpu_ctxt_cb)(void))
 {
 	info->bits = bits;
 	info->ctxt_shift = ilog2(asid_per_ctxt);
+	info->harts_per_core = harts_per_core;
+	info->harts_per_core_mask = ~((1 << ilog2(harts_per_core)) - 1);
 	info->flush_cpu_ctxt_cb = flush_cpu_ctxt_cb;
 	/*
 	 * Expect allocation after rollover to fail if we don't have at least
diff --git a/arch/csky/mm/context.c b/arch/csky/mm/context.c
index 0d95bdd..b58523b 100644
--- a/arch/csky/mm/context.c
+++ b/arch/csky/mm/context.c
@@ -30,7 +30,7 @@ static int asids_init(void)
 {
 	BUG_ON(((1 << CONFIG_CPU_ASID_BITS) - 1) <= num_possible_cpus());
 
-	if (asid_allocator_init(&asid_info, CONFIG_CPU_ASID_BITS, 1,
+	if (asid_allocator_init(&asid_info, CONFIG_CPU_ASID_BITS, 1, 1,
 				asid_flush_cpu_ctxt))
 		panic("Unable to initialize ASID allocator for %lu ASIDs\n",
 		      NUM_ASIDS(&asid_info));
-- 
2.7.4