Skip to contentSkip to conversation

/proc/kmsg permissions, take four

PATCH In review Zack Weinberg 3 participants 7 messages Dec 24, 2006 – Dec 27, 2006

Subsystems (4)
lore
Share & tools
lore
#6 lore raw

PATCH 4/4 Distinguish /proc/kmsg access from sys_syslog

In reply to Zack Weinberg

Finally, add a new security class for access to /proc/kmsg, distinct from the class used for the "read current messages" operations on sys_syslog. The dummy and capability modules permit access to /proc/kmsg to any user (who has somehow acquired an open fd on it); SELinux is unchanged. This accomplishes what I was trying to do in the first place, i.e. enable running klogd unprivileged without a root shim, in a non-SELinux installation. Please remember that the default DAC permissions for /proc/kmsg restrict it to root, so unless you chmod it in your installation or modify klogd to open the file and then drop privs, the actual restrictions are unchanged.

zw

fs/proc/kmsg.c +3 -3
Index: linux-2.6/fs/proc/kmsg.c
===================================================================
@@ -15,7 +15,7 @@
 
 static int kmsg_open(struct inode * inode, struct file * file)
 {
-	int error = security_syslog(LSM_KLOG_READ);
+	int error = security_syslog(LSM_KLOG_READ_PROC);
 	if (error)
 		return error;
 	return nonseekable_open(inode, file);
@@ -29,7 +29,7 @@
 static ssize_t kmsg_read(struct file *file, char __user *buf,
 			 size_t count, loff_t *ppos)
 {
-	int error = security_syslog(LSM_KLOG_READ);
+	int error = security_syslog(LSM_KLOG_READ_PROC);
 	if (error)
 		return error;
 	return klog_read(buf, count, !(file->f_flags & O_NONBLOCK));
@@ -37,7 +37,7 @@
 
 static unsigned int kmsg_poll(struct file *file, poll_table *wait)
 {
-	int error = security_syslog(LSM_KLOG_READ);
+	int error = security_syslog(LSM_KLOG_READ_PROC);
 	if (error)
 		return error;
 	return klog_poll(file, wait);
security/commoncap.c +8 -1
Index: linux-2.6/security/commoncap.c
===================================================================
@@ -311,7 +311,14 @@
 
 int cap_syslog (int type)
 {
-	if ((type & ~LSM_KLOG_READHIST) && !capable(CAP_SYS_ADMIN))
+ 	/*
+ 	 * Reading history is allowed to any user, and so is reading
+ 	 * current messages via /proc/kmsg (by default that file is
+ 	 * only readable by root, but root is allowed to change that,
+ 	 * or open it and hand the fd to an unprivileged process).
+ 	 */
+	if ((type & ~(LSM_KLOG_READHIST|LSM_KLOG_READ_PROC))
+	    && !capable(CAP_SYS_ADMIN))
 		return -EPERM;
 	return 0;
 }
security/selinux/hooks.c +8 -2
Index: linux-2.6/security/selinux/hooks.c
===================================================================
@@ -1505,14 +1505,20 @@
 	int rc;
 
 	/* if there are any codes we don't know about, there's a bug */
-	BUG_ON(type & ~(LSM_KLOG_READ|LSM_KLOG_READHIST
+	BUG_ON(type & ~(LSM_KLOG_READ|LSM_KLOG_READ_PROC|LSM_KLOG_READHIST
 			|LSM_KLOG_CLEARHIST|LSM_KLOG_CONSOLE));
 
 	rc = secondary_ops->syslog(type);
 	if (rc)
 		return rc;
 
-	if (type & (LSM_KLOG_READ|LSM_KLOG_CLEARHIST))
+	/*
+	 * N.B. Unlike the default security model, with
+	 * SELinux active you have to have SYSTEM__SYSLOG_MOD
+	 * privilege to read current messages either with the
+	 * system call or from /proc/kmsg.
+	 */
+	if (type & (LSM_KLOG_READ|LSM_KLOG_READ_PROC|LSM_KLOG_CLEARHIST))
 		rc = task_has_system(current, SYSTEM__SYSLOG_MOD);
 	if (rc)
 		return rc;
include/linux/security.h +2 -0
Index: linux-2.6/include/linux/security.h
===================================================================
@@ -97,6 +97,8 @@
 #define LSM_KLOG_READHIST  2  /* read message history (dmesg) */
 #define LSM_KLOG_CLEARHIST 4  /* clear message history (dmesg -c) */
 #define LSM_KLOG_CONSOLE   8  /* set or query console log level */
+#define LSM_KLOG_READ_PROC 16 /* read current messages, but from /proc/kmsg
+				 rather than the system call */
 
 /* forward declares to avoid warnings */
 struct nfsctl_arg;
security/dummy.c +7 -1
Index: linux-2.6/security/dummy.c
===================================================================
@@ -96,7 +96,13 @@
 
 static int dummy_syslog (int type)
 {
-	if ((type & ~LSM_KLOG_READHIST) && current->euid)
+ 	/*
+ 	 * Reading history is allowed to any user, and so is reading
+ 	 * current messages via /proc/kmsg (by default that file is
+ 	 * only readable by root, but root is allowed to change that,
+ 	 * or open it and hand the fd to an unprivileged process).
+ 	 */
+	if ((type & ~(LSM_KLOG_READHIST|LSM_KLOG_READ_PROC)) && current->euid)
 		return -EPERM;
 	return 0;
 }

--

#4 lore raw

PATCH 2/4 permission mapping for sys_syslog operations

In reply to Zack Weinberg

As suggested by Stephen Smalley: map the various sys_syslog operations to a smaller set of privilege codes before calling security modules. This patch changes the security module interface! There should be no change in the actual security semantics enforced by dummy, capability, nor SELinux (with one exception, clearly marked in sys_syslog).

The privilege codes are now in linux/security.h instead of linux/klog.h, and use the LSM_* naming convention used for other such constants in that file.

I'm now using a static table to map operations to privilege codes. This makes it very easy to see the mapping as a whole, and prevents security operations from cluttering up the body of do_syslog, but does necessitate some checks to ensure that when new KLOG_* operations are added, people don't forget to add entries to the table. If people are not fans of the idiom I used, I will see about getting a feature added to gcc to make it less ugly.

Note that after this patch, close() on /proc/kmsg and sys_syslog(KLOG_CLOSE) do not do security checks; they always succeed. (IMO close should never fail.) This was always in this patch series, but used to be in a different stage.

zw

kernel/printk.c +37 -2
Index: linux-2.6/kernel/printk.c
===================================================================
@@ -164,6 +164,39 @@
 
 __setup("log_buf_len=", log_buf_len_setup);
 
+/*
+ * This table maps KLOG_* operation codes to LSM_KLOG_* privilege classes.
+ * "unsigned char" is used just to keep it small.
+ */
+
+static const unsigned char klog_privclass[] = {
+	[KLOG_CLOSE]           = 0,  /* close always succeeds */
+	[KLOG_OPEN]            = LSM_KLOG_READ,
+	[KLOG_READ]            = LSM_KLOG_READ,
+	[KLOG_GET_UNREAD]      = LSM_KLOG_READ,
+
+	[KLOG_READ_HIST]       = LSM_KLOG_READHIST,
+	[KLOG_GET_SIZE]        = LSM_KLOG_READHIST,
+	[KLOG_READ_CLEAR_HIST] = LSM_KLOG_READHIST|LSM_KLOG_CLEARHIST,
+	[KLOG_CLEAR_HIST]      = LSM_KLOG_CLEARHIST,
+
+	[KLOG_DISABLE_CONSOLE] = LSM_KLOG_CONSOLE,
+	[KLOG_ENABLE_CONSOLE]  = LSM_KLOG_CONSOLE,
+	[KLOG_SET_CONSOLE_LVL] = LSM_KLOG_CONSOLE,
+};
+
+/* It is essential that there be an entry in the above table for every
+ * KLOG_* code.  The following, er, odd declarations cause compilation
+ * of this file to fail if that is not true.  They do not correspond
+ * to real data objects.
+ */
+extern char klog_privclass_is_missing_an_entry[
+	ARRAY_SIZE(klog_privclass) < KLOG_OPCODE_MAX ? -1 : 1
+];
+extern char klog_privclass_has_too_many_entries[
+	ARRAY_SIZE(klog_privclass) > KLOG_OPCODE_MAX ? -1 : 1
+];
+
 /**
  * do_syslog - operate on the log of kernel messages
  * @type: operation to perform
@@ -185,9 +218,11 @@
 	unsigned long i, j, limit, count;
 	int do_clear = 0;
 	char c;
-	int error = 0;
+	int error;
 
-	error = security_syslog(type);
+	if (type < 0 || type >= KLOG_OPCODE_MAX)
+		return -EINVAL;
+	error = security_syslog(klog_privclass[type]);
 	if (error)
 		return error;
 
security/commoncap.c +1 -1
Index: linux-2.6/security/commoncap.c
===================================================================
@@ -311,7 +311,7 @@
 
 int cap_syslog (int type)
 {
-	if ((type != 3 && type != 10) && !capable(CAP_SYS_ADMIN))
+	if ((type & ~LSM_KLOG_READHIST) && !capable(CAP_SYS_ADMIN))
 		return -EPERM;
 	return 0;
 }
security/dummy.c +1 -1
Index: linux-2.6/security/dummy.c
===================================================================
@@ -96,7 +96,7 @@
 
 static int dummy_syslog (int type)
 {
-	if ((type != 3 && type != 10) && current->euid)
+	if ((type & ~LSM_KLOG_READHIST) && current->euid)
 		return -EPERM;
 	return 0;
 }
security/selinux/hooks.c +17 -19
Index: linux-2.6/security/selinux/hooks.c
===================================================================
@@ -1504,29 +1504,27 @@
 {
 	int rc;
 
+	/* if there are any codes we don't know about, there's a bug */
+	BUG_ON(type & ~(LSM_KLOG_READ|LSM_KLOG_READHIST
+			|LSM_KLOG_CLEARHIST|LSM_KLOG_CONSOLE));
+
 	rc = secondary_ops->syslog(type);
 	if (rc)
 		return rc;
 
-	switch (type) {
-		case 3:         /* Read last kernel messages */
-		case 10:        /* Return size of the log buffer */
-			rc = task_has_system(current, SYSTEM__SYSLOG_READ);
-			break;
-		case 6:         /* Disable logging to console */
-		case 7:         /* Enable logging to console */
-		case 8:		/* Set level of messages printed to console */
-			rc = task_has_system(current, SYSTEM__SYSLOG_CONSOLE);
-			break;
-		case 0:         /* Close log */
-		case 1:         /* Open log */
-		case 2:         /* Read from log */
-		case 4:         /* Read/clear last kernel messages */
-		case 5:         /* Clear ring buffer */
-		default:
-			rc = task_has_system(current, SYSTEM__SYSLOG_MOD);
-			break;
-	}
+	if (type & (LSM_KLOG_READ|LSM_KLOG_CLEARHIST))
+		rc = task_has_system(current, SYSTEM__SYSLOG_MOD);
+	if (rc)
+		return rc;
+
+	if (type & LSM_KLOG_READHIST)
+		rc = task_has_system(current, SYSTEM__SYSLOG_READ);
+	if (rc)
+		return rc;
+
+	if (type & LSM_KLOG_CONSOLE)
+		rc = task_has_system(current, SYSTEM__SYSLOG_CONSOLE);
+
 	return rc;
 }
 
include/linux/security.h +12 -0
Index: linux-2.6/include/linux/security.h
===================================================================
@@ -86,6 +86,18 @@
 /* setfsuid or setfsgid, id0 == fsuid or fsgid */
 #define LSM_SETID_FS	8
 
+/*
+ * Values passed to security_syslog to indicate which privilege class
+ * is requested.  requested.  Note these are bitmasks - callers may
+ * request more than one of the privilege classes at once (this is
+ * currently only used for READHIST|CLEARHIST) or no privilege at all
+ * (which should always succeed).
+ */
+#define LSM_KLOG_READ      1  /* read current messages (klogd) */
+#define LSM_KLOG_READHIST  2  /* read message history (dmesg) */
+#define LSM_KLOG_CLEARHIST 4  /* clear message history (dmesg -c) */
+#define LSM_KLOG_CONSOLE   8  /* set or query console log level */
+
 /* forward declares to avoid warnings */
 struct nfsctl_arg;
 struct sched_param;

--

#5 lore raw

PATCH 3/4 Refactor do_syslog interface

In reply to Zack Weinberg

This patch breaks out the read operations in do_syslog() into their own functions (klog_read, klog_readhist) and adds a klog_poll. klog_read grows the ability to do a nonblocking read, which I expose in the sys_syslog interface because there doesn't seem to be any reason not to. do_syslog itself is folded into sys_syslog. The security checks remain there, not in the subfunctions. I've added kerneldoc comments for all the new functions in printk.c. The control flow in sys_syslog is different in style but not in substance -- I like direct returns rather than setting a local variable and returning at the end; it lets me put a BUG() after the switch statement to catch missed cases. [gcc4 at least can optimize it out when there aren't any.]

kmsg.c is then changed to use those functions instead of calling do_syslog and/or poll_wait itself.. This entails that it must call security_syslog as appropriate itself. In this patch I preserve the security checks exactly as they were after the changes in the "map permissions" patch (i.e. kmsg_close() doesn't do a permission check).

Finally, I fixed some minor bugs. Error and partial read handling in klog_read/klog_readhist were slightly off: if __put_user returns an error, that character should not be consumed from the kernel buffer; if it returns an error after some characters have already been copied successfully, the read operation should return the count of already-copied characters, not the error code; if less than the entire buffer is read with KLOG_READ_CLEAR_HIST, we should only clear the part that was successfully read. Seeking on /proc/kmsg has never been meaningful, so kmsg_open() should call nonseekable_open() to enforce that.

proc/kmsg.c declares the kernel/printk.c interfaces itself, instead of getting them from klog.h which people want to be purely userspace- visible constants. kmsg.c has always had private declarations of printk.c functions (before, there were declarations of do_syslog and a wait queue there); as it is unlikely that more users of these functions will appear, I think this will do fine. (It might be reasonable to put declarations in console.h.)

It was possible to prune down the set of headers included by kmsg.c quite substantially.

zw

fs/proc/kmsg.c +17 -22
Index: linux-2.6/fs/proc/kmsg.c
===================================================================
@@ -5,47 +5,42 @@
  *
  */
 
-#include <linux/types.h>
-#include <linux/errno.h>
-#include <linux/time.h>
-#include <linux/kernel.h>
-#include <linux/poll.h>
 #include <linux/fs.h>
-#include <linux/klog.h>
-
-#include <asm/uaccess.h>
-#include <asm/io.h>
-
-extern wait_queue_head_t log_wait;
+#include <linux/poll.h>
+#include <linux/security.h>
 
-extern int do_syslog(int type, char __user *bug, int count);
+/* interfaces from kernel/printk.c */
+extern int klog_read(char __user *, int, int);
+extern unsigned int klog_poll(struct file *, poll_table *);
 
 static int kmsg_open(struct inode * inode, struct file * file)
 {
-	return do_syslog(KLOG_OPEN, NULL, 0);
+	int error = security_syslog(LSM_KLOG_READ);
+	if (error)
+		return error;
+	return nonseekable_open(inode, file);
 }
 
 static int kmsg_release(struct inode * inode, struct file * file)
 {
-	(void) do_syslog(KLOG_CLOSE, NULL, 0);
 	return 0;
 }
 
 static ssize_t kmsg_read(struct file *file, char __user *buf,
 			 size_t count, loff_t *ppos)
 {
-	if ((file->f_flags & O_NONBLOCK)
-	    && !do_syslog(KLOG_GET_UNREAD, NULL, 0))
-		return -EAGAIN;
-	return do_syslog(KLOG_READ, buf, count);
+	int error = security_syslog(LSM_KLOG_READ);
+	if (error)
+		return error;
+	return klog_read(buf, count, !(file->f_flags & O_NONBLOCK));
 }
 
 static unsigned int kmsg_poll(struct file *file, poll_table *wait)
 {
-	poll_wait(file, &log_wait, wait);
-	if (do_syslog(KLOG_GET_UNREAD, NULL, 0))
-		return POLLIN | POLLRDNORM;
-	return 0;
+	int error = security_syslog(LSM_KLOG_READ);
+	if (error)
+		return error;
+	return klog_poll(file, wait);
 }
 
 
include/linux/klog.h +2 -0
Index: linux-2.6/include/linux/klog.h
===================================================================
@@ -21,6 +21,8 @@
 
 	KLOG_GET_UNREAD      =  9, /* return number of unread characters */
 	KLOG_GET_SIZE        = 10, /* return size of log buffer */
+ 	KLOG_READ_NONBLOCK   = 11, /* read from log, don't block if empty
+ 				    * -- new in 2.6.20 */
 
 	KLOG_OPCODE_MAX
 };
kernel/printk.c +161 -112

335 lines not on this page · raw message

--

#0 lore raw

cover /proc/kmsg permissions, take four

Here's yet another revision of the /proc/kmsg permissions patch series. To recap, the point is to allow klogd to drop privileges and continue reading from /proc/kmsg (currently, even if klogd has a legitimately opened fd on /proc/kmsg, it cannot read from it unless it has CAP_SYS_ADMIN asserted). SELinux's pickier and finer-grained privilege rules for /proc/kmsg are unchanged.

There are two significant changes from the previous revision. First, in keeping with the recommended style, I have eliminated the security_syslog_or_fail() macro. Instead there is a static array mapping KLOG_* opcodes to LSM_KLOG_* privilege classes. This requires slightly different coding in the security hooks but I think it's clearer overall. Second, I've incorporated Vincent Legoll's kerneldoc comment for sys_syslog (nee do_syslog) with some wording improvements and expansion to cover the klog_* functions introduced part-way through the patch. I don't think proc/kmsg.c needs kerneldoc, it's very simple after this patch series.

I've been through Documentation/CodingStyle and satisfied myself that everything is now in the proper mode. I don't suppose anyone has comments on the *content* of the changes...?

zw

#1 lore raw

PATCH 1/4 Add <linux/klog.h>

This patch introduces <linux/klog.h> with symbolic constants for the various sys_syslog() opcodes, and changes all in-kernel references to those opcodes to use the constants. The header is added to the set of user/kernel interface headers.

zw

include/linux/Kbuild +1 -0
Index: linux-2.6/include/linux/Kbuild
===================================================================
@@ -100,6 +100,7 @@
 header-y += ixjuser.h
 header-y += jffs2.h
 header-y += keyctl.h
+header-y += klog.h
 header-y += limits.h
 header-y += lock_dlm_plock.h
 header-y += magic.h
include/linux/klog.h new file +28 -0
Index: linux-2.6/include/linux/klog.h
===================================================================
@@ -0,0 +1,28 @@
+#ifndef _LINUX_KLOG_H
+#define _LINUX_KLOG_H
+
+/*
+ * Constants for the first argument to the syslog() system call
+ * (aka klogctl()).  These numbers are part of the user space ABI!
+ */
+enum {
+	KLOG_CLOSE           =  0, /* close log */
+	KLOG_OPEN            =  1, /* open log */
+	KLOG_READ            =  2, /* read from log (klogd) */
+
+	KLOG_READ_HIST       =  3, /* read history of log messages (dmesg) */
+	KLOG_READ_CLEAR_HIST =  4, /* read and clear history */
+	KLOG_CLEAR_HIST      =  5, /* just clear history */
+
+	KLOG_DISABLE_CONSOLE =  6, /* disable printk to console */
+	KLOG_ENABLE_CONSOLE  =  7, /* enable printk to console */
+	KLOG_SET_CONSOLE_LVL =  8, /* set minimum severity of messages to be
+				    * printed to console */
+
+	KLOG_GET_UNREAD      =  9, /* return number of unread characters */
+	KLOG_GET_SIZE        = 10, /* return size of log buffer */
+
+	KLOG_OPCODE_MAX
+};
+
+#endif /* klog.h */
kernel/printk.c +26 -24

99 lines not on this page · raw message

fs/proc/kmsg.c +7 -5
Index: linux-2.6/fs/proc/kmsg.c
===================================================================
@@ -11,6 +11,7 @@
 #include <linux/kernel.h>
 #include <linux/poll.h>
 #include <linux/fs.h>
+#include <linux/klog.h>
 
 #include <asm/uaccess.h>
 #include <asm/io.h>
@@ -21,27 +22,28 @@
 
 static int kmsg_open(struct inode * inode, struct file * file)
 {
-	return do_syslog(1,NULL,0);
+	return do_syslog(KLOG_OPEN, NULL, 0);
 }
 
 static int kmsg_release(struct inode * inode, struct file * file)
 {
-	(void) do_syslog(0,NULL,0);
+	(void) do_syslog(KLOG_CLOSE, NULL, 0);
 	return 0;
 }
 
 static ssize_t kmsg_read(struct file *file, char __user *buf,
 			 size_t count, loff_t *ppos)
 {
-	if ((file->f_flags & O_NONBLOCK) && !do_syslog(9, NULL, 0))
+	if ((file->f_flags & O_NONBLOCK)
+	    && !do_syslog(KLOG_GET_UNREAD, NULL, 0))
 		return -EAGAIN;
-	return do_syslog(2, buf, count);
+	return do_syslog(KLOG_READ, buf, count);
 }
 
 static unsigned int kmsg_poll(struct file *file, poll_table *wait)
 {
 	poll_wait(file, &log_wait, wait);
-	if (do_syslog(9, NULL, 0))
+	if (do_syslog(KLOG_GET_UNREAD, NULL, 0))
 		return POLLIN | POLLRDNORM;
 	return 0;
 }

--

#2 lore raw

Re: [patch 1/4] Add <linux/klog.h>

Quoted 13 lines · Zack Weinberg

===================================================================

fs/proc/kmsg.c +1 -0
@@ -11,6 +11,7 @@
 #include <linux/kernel.h>
 #include <linux/poll.h>
 #include <linux/fs.h>
+#include <linux/klog.h>
 
 #include <asm/uaccess.h>
 #include <asm/io.h>
@@ -21,27 +22,28 @@
 

fs/proc/kmsg.c line 30

Quoted 10 lines
fs/proc/kmsg.c +2 -2
 static int kmsg_open(struct inode * inode, struct file * file)
 {
-	return do_syslog(1,NULL,0);
+	return do_syslog(KLOG_OPEN, NULL, 0);
 }
 
 static int kmsg_release(struct inode * inode, struct file * file)
 {
-	(void) do_syslog(0,NULL,0);
+	(void) do_syslog(KLOG_CLOSE, NULL, 0);
        ^^^^^^

I bet you can get rid of that in the process.

	-`J'
Acked-by on v1 1/4
#3 lore raw

Re: [patch 1/4] Add <linux/klog.h>

In reply to Zack Weinberg
Quoted 4 lines · Zack Weinberg

This patch introduces <linux/klog.h> with symbolic constants for the various sys_syslog() opcodes, and changes all in-kernel references to those opcodes to use the constants. The header is added to the set of user/kernel interface headers.

[...]

kernel/printk.c

Quoted 32 lines
kernel/printk.c +14 -13
-/*
- * Commands to do_syslog:
+/**
+ * do_syslog - operate on the log of kernel messages
+ * @type: operation to perform
+ * @buf: user-space buffer to copy data into
+ * @len: number of bytes of space available at @buf
+ *
+ * See include/linux/klog.h for the command numbers passed as @type.
+ * The @buf and @len parameters are used with the above meanings for
+ * @type values %KLOG_READ, %KLOG_READ_HIST and %KLOG_READ_CLEAR_HIST.
+ * @len is reused with a different meaning, and @buf ignored, for
+ * %KLOG_SET_CONSOLE_LVL.  Both @buf and @len are ignored for all
+ * other @type values.
  *
- * 	0 -- Close the log.  Currently a NOP.
- * 	1 -- Open the log. Currently a NOP.
- * 	2 -- Read from the log.
- * 	3 -- Read all messages remaining in the ring buffer.
- * 	4 -- Read and clear all messages remaining in the ring buffer
- * 	5 -- Clear ring buffer.
- * 	6 -- Disable printk's to console
- * 	7 -- Enable printk's to console
- *	8 -- Set level of messages printed to console
- *	9 -- Return number of unread characters in the log buffer
- *     10 -- Return size of the log buffer
+ * On failure, returns a negative errno code.  On success, returns a
+ * nonnegative integer whose meaning depends on @type.
  */
 int do_syslog(int type, char __user *buf, int len)
 {
@@ -190,11 +192,11 @@

That part looks good to me, and the kernel-doc hunks from "[patch 3/4] Refactor do_syslog interface"
too, are:

thanks for handling that.

Signature · 1 line
Vincent Legoll