Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.33 as permitted sender) client-ip=23.128.96.33;
Date:   Wed, 20 Sep 2023 15:04:43 +0000
Mime-Version: 1.0
Message-ID: <20230920150443.1789000-1-jackmanb@google.com>
Subject: [PATCH v2] x86/entry: Avoid redundant CR3 write on paranoid returns
From:   Brendan Jackman <jackmanb@google.com>
To:     luto@kernel.org, tglx@linutronix.de
Cc:     mingo@redhat.com, bp@alien8.de, dave.hansen@linux.intel.com,
        x86@kernel.org, hpa@zytor.com, linux-kernel@vger.kernel.org,
        laijs@linux.alibaba.com, yosryahmed@google.com, reijiw@google.com,
        oweisse@google.com, peterz@infradead.org,
        Brendan Jackman <jackmanb@google.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

From: Lai Jiangshan <laijs@linux.alibaba.com>

This path gets used called from:

1. #NMI return.
2. paranoid_exit (i.e. #MCE, #VC, #DB and #DF return)

Contrary to the implication in commit 21e94459110252 ("x86/mm: Optimize
RESTORE_CR3"), we never modify CR3 in any of these exceptions, except
for switching from user to kernel pagetables under PTI. That means that
most of the time when returning from an exception that interrupted the
kernel no CR3 restore is necessary. Writing CR3 is expensive on some
machines, so this commit avoids redundant writes.

I said "most of the time" because we might have interrupted the kernel
entry before the user->kernel CR3 switch or the exit after the
kernel->user switch. In the former case skipping the restore might
actually be be fine, but definitely not the latter. So we do still need
to check the saved CR3 and restore it if it's a user CR3.

To reflect the new behaviour RESTORE_CR3 is given a longer name, and a
comment that was describing its behaviour at the call site is removed.
We can also simplify the code around the SET_NOFLUSH_BIT invocation
as we no longer need to branch to it from above.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
[Rewrote commit message; responded to review comments]
Signed-off-by: Brendan Jackman <jackmanb@google.com>
---

Notes:
    V1: https://lore.kernel.org/lkml/20230817121513.1382800-1-jackmanb@google.com/
    
    v1->v2: Rewrote some comments, added a proper commit message, cleaned up
    	the code per tglx's suggestion.
    
    	I've kept Lai as the Author. If you prefer for the blame to
    	record the last person that touched it then that's also fine
    	though, I can credit Lai as Co-developed-by.

 arch/x86/entry/calling.h  | 25 +++++++++----------------
 arch/x86/entry/entry_64.S |  7 +++----
 2 files changed, 12 insertions(+), 20 deletions(-)

diff --git a/arch/x86/entry/calling.h b/arch/x86/entry/calling.h
index f6907627172b..84b1e32c27a1 100644
--- a/arch/x86/entry/calling.h
+++ b/arch/x86/entry/calling.h
@@ -233,17 +233,18 @@ For 32-bit we have the following conventions - kernel is built with
 .Ldone_\@:
 .endm
 
-.macro RESTORE_CR3 scratch_reg:req save_reg:req
+.macro RESTORE_CR3_IF_USER scratch_reg:req save_reg:req
 	ALTERNATIVE "jmp .Lend_\@", "", X86_FEATURE_PTI
 
-	ALTERNATIVE "jmp .Lwrcr3_\@", "", X86_FEATURE_PCID
-
 	/*
-	 * KERNEL pages can always resume with NOFLUSH as we do
-	 * explicit flushes.
+	 * If CR3 contained the kernel page tables at the paranoid exception
+	 * entry, then there is nothing to restore as CR3 is not modified while
+	 * handling the exception.
 	 */
 	bt	$PTI_USER_PGTABLE_BIT, \save_reg
-	jnc	.Lnoflush_\@
+	jnc	.Lend_\@
+
+	ALTERNATIVE "jmp .Lwrcr3_\@", "", X86_FEATURE_PCID
 
 	/*
 	 * Check if there's a pending flush for the user ASID we're
@@ -251,20 +252,12 @@ For 32-bit we have the following conventions - kernel is built with
 	 */
 	movq	\save_reg, \scratch_reg
 	andq	$(0x7FF), \scratch_reg
-	bt	\scratch_reg, THIS_CPU_user_pcid_flush_mask
-	jnc	.Lnoflush_\@
-
 	btr	\scratch_reg, THIS_CPU_user_pcid_flush_mask
-	jmp	.Lwrcr3_\@
+	jc	.Lwrcr3_\@
 
-.Lnoflush_\@:
 	SET_NOFLUSH_BIT \save_reg
 
 .Lwrcr3_\@:
-	/*
-	 * The CR3 write could be avoided when not changing its value,
-	 * but would require a CR3 read *and* a scratch register.
-	 */
 	movq	\save_reg, %cr3
 .Lend_\@:
 .endm
@@ -279,7 +272,7 @@ For 32-bit we have the following conventions - kernel is built with
 .endm
 .macro SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg:req save_reg:req
 .endm
-.macro RESTORE_CR3 scratch_reg:req save_reg:req
+.macro RESTORE_CR3_IF_USER scratch_reg:req save_reg:req
 .endm
 
 #endif
diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S
index 43606de22511..ff73767b5d1f 100644
--- a/arch/x86/entry/entry_64.S
+++ b/arch/x86/entry/entry_64.S
@@ -1019,14 +1019,14 @@ SYM_CODE_START_LOCAL(paranoid_exit)
 	IBRS_EXIT save_reg=%r15
 
 	/*
-	 * The order of operations is important. RESTORE_CR3 requires
+	 * The order of operations is important. RESTORE_CR3_IF_USER requires
 	 * kernel GSBASE.
 	 *
 	 * NB to anyone to try to optimize this code: this code does
 	 * not execute at all for exceptions from user mode. Those
 	 * exceptions go through error_return instead.
 	 */
-	RESTORE_CR3	scratch_reg=%rax save_reg=%r14
+	RESTORE_CR3_IF_USER scratch_reg=%rax save_reg=%r14
 
 	/* Handle the three GSBASE cases */
 	ALTERNATIVE "jmp .Lparanoid_exit_checkgs", "", X86_FEATURE_FSGSBASE
@@ -1457,8 +1457,7 @@ end_repeat_nmi:
 	/* Always restore stashed SPEC_CTRL value (see paranoid_entry) */
 	IBRS_EXIT save_reg=%r15
 
-	/* Always restore stashed CR3 value (see paranoid_entry) */
-	RESTORE_CR3 scratch_reg=%r15 save_reg=%r14
+	RESTORE_CR3_IF_USER scratch_reg=%r15 save_reg=%r14
 
 	/*
 	 * The above invocation of paranoid_entry stored the GSBASE
-- 
2.42.0.459.ge4e396fd5e-goog