Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Greg Kroah-Hartman <gregkh@linuxfoundation.org>
To:     linux-kernel@vger.kernel.org
Cc:     Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
        stable@vger.kernel.org, Joerg Roedel <jroedel@suse.de>,
        Borislav Petkov <bp@suse.de>
Subject: [PATCH 5.12 072/127] x86/sev-es: Use __put_user()/__get_user() for data accesses
Date:   Mon, 24 May 2021 17:26:29 +0200
Message-Id: <20210524152337.290079245@linuxfoundation.org>
In-Reply-To: <20210524152334.857620285@linuxfoundation.org>
References: <20210524152334.857620285@linuxfoundation.org>
User-Agent: quilt/0.66
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Precedence: bulk

From: Joerg Roedel <jroedel@suse.de>

commit 4954f5b8ef0baf70fe978d1a99a5f70e4dd5c877 upstream.

The put_user() and get_user() functions do checks on the address which is
passed to them. They check whether the address is actually a user-space
address and whether its fine to access it. They also call might_fault()
to indicate that they could fault and possibly sleep.

All of these checks are neither wanted nor needed in the #VC exception
handler, which can be invoked from almost any context and also for MMIO
instructions from kernel space on kernel memory. All the #VC handler
wants to know is whether a fault happened when the access was tried.

This is provided by __put_user()/__get_user(), which just do the access
no matter what. Also add comments explaining why __get_user() and
__put_user() are the best choice here and why it is safe to use them
in this context. Also explain why copy_to/from_user can't be used.

In addition, also revert commit

  7024f60d6552 ("x86/sev-es: Handle string port IO to kernel memory properly")

because using __get_user()/__put_user() fixes the same problem while
the above commit introduced several problems:

  1) It uses access_ok() which is only allowed in task context.

  2) It uses memcpy() which has no fault handling at all and is
     thus unsafe to use here.

  [ bp: Fix up commit ID of the reverted commit above. ]

Fixes: f980f9c31a92 ("x86/sev-es: Compile early handler code into kernel image")
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Signed-off-by: Borislav Petkov <bp@suse.de>
Cc: stable@vger.kernel.org # v5.10+
Link: https://lkml.kernel.org/r/20210519135251.30093-4-joro@8bytes.org
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
---
 arch/x86/kernel/sev-es.c |   66 ++++++++++++++++++++++++++++++++---------------
 1 file changed, 46 insertions(+), 20 deletions(-)

--- a/arch/x86/kernel/sev-es.c
+++ b/arch/x86/kernel/sev-es.c
@@ -288,31 +288,44 @@ static enum es_result vc_write_mem(struc
 	u16 d2;
 	u8  d1;
 
-	/* If instruction ran in kernel mode and the I/O buffer is in kernel space */
-	if (!user_mode(ctxt->regs) && !access_ok(target, size)) {
-		memcpy(dst, buf, size);
-		return ES_OK;
-	}
-
+	/*
+	 * This function uses __put_user() independent of whether kernel or user
+	 * memory is accessed. This works fine because __put_user() does no
+	 * sanity checks of the pointer being accessed. All that it does is
+	 * to report when the access failed.
+	 *
+	 * Also, this function runs in atomic context, so __put_user() is not
+	 * allowed to sleep. The page-fault handler detects that it is running
+	 * in atomic context and will not try to take mmap_sem and handle the
+	 * fault, so additional pagefault_enable()/disable() calls are not
+	 * needed.
+	 *
+	 * The access can't be done via copy_to_user() here because
+	 * vc_write_mem() must not use string instructions to access unsafe
+	 * memory. The reason is that MOVS is emulated by the #VC handler by
+	 * splitting the move up into a read and a write and taking a nested #VC
+	 * exception on whatever of them is the MMIO access. Using string
+	 * instructions here would cause infinite nesting.
+	 */
 	switch (size) {
 	case 1:
 		memcpy(&d1, buf, 1);
-		if (put_user(d1, target))
+		if (__put_user(d1, target))
 			goto fault;
 		break;
 	case 2:
 		memcpy(&d2, buf, 2);
-		if (put_user(d2, target))
+		if (__put_user(d2, target))
 			goto fault;
 		break;
 	case 4:
 		memcpy(&d4, buf, 4);
-		if (put_user(d4, target))
+		if (__put_user(d4, target))
 			goto fault;
 		break;
 	case 8:
 		memcpy(&d8, buf, 8);
-		if (put_user(d8, target))
+		if (__put_user(d8, target))
 			goto fault;
 		break;
 	default:
@@ -343,30 +356,43 @@ static enum es_result vc_read_mem(struct
 	u16 d2;
 	u8  d1;
 
-	/* If instruction ran in kernel mode and the I/O buffer is in kernel space */
-	if (!user_mode(ctxt->regs) && !access_ok(s, size)) {
-		memcpy(buf, src, size);
-		return ES_OK;
-	}
-
+	/*
+	 * This function uses __get_user() independent of whether kernel or user
+	 * memory is accessed. This works fine because __get_user() does no
+	 * sanity checks of the pointer being accessed. All that it does is
+	 * to report when the access failed.
+	 *
+	 * Also, this function runs in atomic context, so __get_user() is not
+	 * allowed to sleep. The page-fault handler detects that it is running
+	 * in atomic context and will not try to take mmap_sem and handle the
+	 * fault, so additional pagefault_enable()/disable() calls are not
+	 * needed.
+	 *
+	 * The access can't be done via copy_from_user() here because
+	 * vc_read_mem() must not use string instructions to access unsafe
+	 * memory. The reason is that MOVS is emulated by the #VC handler by
+	 * splitting the move up into a read and a write and taking a nested #VC
+	 * exception on whatever of them is the MMIO access. Using string
+	 * instructions here would cause infinite nesting.
+	 */
 	switch (size) {
 	case 1:
-		if (get_user(d1, s))
+		if (__get_user(d1, s))
 			goto fault;
 		memcpy(buf, &d1, 1);
 		break;
 	case 2:
-		if (get_user(d2, s))
+		if (__get_user(d2, s))
 			goto fault;
 		memcpy(buf, &d2, 2);
 		break;
 	case 4:
-		if (get_user(d4, s))
+		if (__get_user(d4, s))
 			goto fault;
 		memcpy(buf, &d4, 4);
 		break;
 	case 8:
-		if (get_user(d8, s))
+		if (__get_user(d8, s))
 			goto fault;
 		memcpy(buf, &d8, 8);
 		break;