OS-6211 support Linux accounting

Reviewed by: Patrick Mooney <patrick.mooney@joyent.com>
Approved by: Patrick Mooney <patrick.mooney@joyent.com>

14 files changed, 243 insertions, 11 deletions

diff --git a/usr/src/uts/common/brand/lx/os/lx_acct.c b/usr/src/uts/common/brand/lx/os/lx_acct.c
new file mode 100644
index 0000000000..7f38a240ab
--- /dev/null
+++ b/usr/src/uts/common/brand/lx/os/lx_acct.c
@@ -0,0 +1,198 @@
+/*
+ * This file and its contents are supplied under the terms of the
+ * Common Development and Distribution License ("CDDL"), version 1.0.
+ * You may only use this file in accordance with the terms of version
+ * 1.0 of the CDDL.
+ *
+ * A full copy of the text of the CDDL should have accompanied this
+ * source.  A copy of the CDDL is also available via the Internet at
+ * http://www.illumos.org/license/CDDL.
+ */
+
+/*
+ * Copyright 2017 Joyent, Inc.
+ */
+
+#include <sys/types.h>
+#include <sys/sysmacros.h>
+#include <sys/systm.h>
+#include <sys/acct.h>
+#include <sys/proc.h>
+#include <sys/user.h>
+#include <sys/cred.h>
+#include <sys/file.h>
+#include <sys/vnode.h>
+#include <sys/session.h>
+#include <sys/wait.h>
+#include <sys/ddi.h>
+#include <sys/zone.h>
+#include <sys/lx_types.h>
+
+/*
+ * Based on the Linux acct(5) man page, their comp_t definition is the same
+ * as ours. lxac_etime is encoded as a float for v3 accounting records.
+ */
+
+#define	LX_ACCT_VERSION	3
+
+/*
+ * Bit flags in lxac_flag. The Linux AFORK and ASU match native. The rest of
+ * the flags diverge.
+ */
+#define	LX_AFORK	0x01	/* executed fork, but no exec */
+#define	LX_ASU		0x02	/* used superuser privileges */
+#define	LX_ACORE	0x08	/* dumped core */
+#define	LX_AXSIG	0x10	/* killed by a signal */
+
+typedef struct lx_acct {
+	char		lxac_flag;
+	char		lxac_version;
+	uint16_t	lxac_tty;
+	uint32_t	lxac_exitcode;
+	uint32_t	lxac_uid;
+	uint32_t	lxac_gid;
+	uint32_t	lxac_pid;
+	uint32_t	lxac_ppid;
+	uint32_t	lxac_btime;	/* seconds since the epoch */
+	uint32_t	lxac_etime;	/* float representation of ticks */
+	comp_t		lxac_utime;
+	comp_t		lxac_stime;
+	comp_t		lxac_mem;	/* kb */
+	comp_t		lxac_io;	/* unused */
+	comp_t		lxac_rw;	/* unused */
+	comp_t		lxac_minflt;
+	comp_t		lxac_majflt;
+	comp_t		lxac_swaps;	/* unused */
+	char		lxac_comm[16];
+} lx_acct_t;
+
+/*
+ * Same functionality as acct_compress(). Produce a pseudo-floating point
+ * representation with 3 bits base-8 exponent, 13 bits fraction.
+ */
+static comp_t
+lx_acct_compt(ulong_t t)
+{
+	int exp = 0, round = 0;
+
+	while (t >= 8192) {
+		exp++;
+		round = t & 04;
+		t >>= 3;
+	}
+	if (round) {
+		t++;
+		if (t >= 8192) {
+			t >>= 3;
+			exp++;
+		}
+	}
+#ifdef _LP64
+	if (exp > 7) {
+		/* prevent wraparound */
+		t = 8191;
+		exp = 7;
+	}
+#endif
+	return ((exp << 13) + t);
+}
+
+/*
+ * 32-bit IEEE float encoding as-per Linux.
+ */
+static uint32_t
+lx_acct_float(int64_t t)
+{
+	uint32_t val, exp = 190;
+
+	if (t == 0)
+		return (0);
+
+	while (t > 0) {
+		t <<= 1;
+		exp--;
+	}
+	val = (uint32_t)(t >> 40) & 0x7fffffu;
+
+	return (val | (exp << 23));
+}
+
+/*
+ * Write a Linux-formatted record to the accounting file.
+ */
+void
+lx_acct_out(vnode_t *vp, int exit_status)
+{
+	struct proc *p;
+	user_t *ua;
+	struct cred *cr;
+	dev_t d;
+	pid_t pid, ppid;
+	struct vattr va;
+	ssize_t resid = 0;
+	int err;
+	lx_acct_t a;
+
+	p = curproc;
+	ua = PTOU(p);
+	cr = CRED();
+
+	bzero(&a, sizeof (a));
+
+	a.lxac_flag = ua->u_acflag & (LX_AFORK | LX_ASU);
+	a.lxac_version = LX_ACCT_VERSION;
+	d = cttydev(p);
+	a.lxac_tty = LX_MAKEDEVICE(getmajor(d), getminor(d));
+	if (WIFEXITED(exit_status)) {
+		a.lxac_exitcode = WEXITSTATUS(exit_status);
+	} else if (WIFSIGNALED(exit_status)) {
+		a.lxac_flag |= LX_AXSIG;
+		if (WCOREDUMP(exit_status)) {
+			a.lxac_flag |= LX_ACORE;
+		}
+	}
+	a.lxac_uid = crgetruid(cr);
+	a.lxac_gid = crgetrgid(cr);
+	pid = p->p_pid;
+	ppid = p->p_ppid;
+	/* Perform pid translation ala lxpr_fixpid(). */
+	if (pid == curzone->zone_proc_initpid) {
+		pid = 1;
+		ppid = 0;
+	} else {
+		if (ppid == curzone->zone_proc_initpid) {
+			ppid = 1;
+		} else if (ppid == curzone->zone_zsched->p_pid ||
+		    (p->p_flag & SZONETOP) != 0) {
+			ppid = 1;
+		}
+	}
+	a.lxac_pid = pid;
+	a.lxac_ppid = ppid;
+	a.lxac_btime = ua->u_start.tv_sec;
+	/* For Linux v3 accounting record, this is an encoded float. */
+	a.lxac_etime = lx_acct_float(ddi_get_lbolt() - ua->u_ticks);
+	a.lxac_utime = lx_acct_compt(NSEC_TO_TICK(p->p_acct[LMS_USER]));
+	a.lxac_stime = lx_acct_compt(
+	    NSEC_TO_TICK(p->p_acct[LMS_SYSTEM] + p->p_acct[LMS_TRAP]));
+	a.lxac_mem = lx_acct_compt((ulong_t)(ptob(ua->u_mem) / 1024));
+	/* a.lxac_io		unused */
+	/* a.lxac_rw		unused */
+	a.lxac_minflt = lx_acct_compt((ulong_t)p->p_ru.minflt);
+	a.lxac_majflt = lx_acct_compt((ulong_t)p->p_ru.majflt);
+	/* a.lxac_swaps		unused */
+	bcopy(ua->u_comm, a.lxac_comm, sizeof (a.lxac_comm));
+
+	/*
+	 * As with the native acct() handling, we save the size so that if the
+	 * write fails, we can reset the size to avoid corrupting the accounting
+	 * file.
+	 */
+	va.va_mask = AT_SIZE;
+	if (VOP_GETATTR(vp, &va, 0, kcred, NULL) == 0) {
+		err = vn_rdwr(UIO_WRITE, vp, (caddr_t)&a, sizeof (a), 0LL,
+		    UIO_SYSSPACE, FAPPEND, (rlim64_t)MAXOFF_T, kcred, &resid);
+		if (err != 0 || resid != 0)
+			(void) VOP_SETATTR(vp, &va, 0, kcred, NULL);
+	}
+}
diff --git a/usr/src/uts/common/brand/lx/os/lx_brand.c b/usr/src/uts/common/brand/lx/os/lx_brand.c
index a8a710f088..3646df26e5 100644
--- a/usr/src/uts/common/brand/lx/os/lx_brand.c
+++ b/usr/src/uts/common/brand/lx/os/lx_brand.c
@@ -307,7 +307,8 @@ struct brand_ops lx_brops = {
 #endif
 	B_FALSE,			/* b_intp_parse_arg */
 	lx_clearbrand,			/* b_clearbrand */
-	lx_upcall_statd			/* b_rpc_statd */
+	lx_upcall_statd,		/* b_rpc_statd */
+	lx_acct_out			/* b_acct_out */
 };
 
 struct brand_mach_ops lx_mops = {
diff --git a/usr/src/uts/common/brand/lx/os/lx_syscall.c b/usr/src/uts/common/brand/lx/os/lx_syscall.c
index 7191a52f77..b8088e2b5d 100644
--- a/usr/src/uts/common/brand/lx/os/lx_syscall.c
+++ b/usr/src/uts/common/brand/lx/os/lx_syscall.c
@@ -570,7 +570,7 @@ lx_sysent_t lx_sysent32[] = {
 	{"signal",	NULL,			0,		2}, /* 48 */
 	{"geteuid16",	lx_geteuid16,		0,		0}, /* 49 */
 	{"getegid16",	lx_getegid16,		0,		0}, /* 50 */
-	{"acct",	NULL,			NOSYS_NO_EQUIV,	0}, /* 51 */
+	{"acct",	lx_acct,		0,		1}, /* 51 */
 	{"umount2",	lx_umount2,		0,		2}, /* 52 */
 	{"lock",	NULL,			NOSYS_OBSOLETE,	0}, /* 53 */
 	{"ioctl",	lx_ioctl,		0,		3}, /* 54 */
@@ -1053,7 +1053,7 @@ lx_sysent_t lx_sysent64[] = {
 	{"setrlimit",	lx_setrlimit,		0,		2}, /* 160 */
 	{"chroot",	lx_chroot,		0,		1}, /* 161 */
 	{"sync",	lx_sync,		0,		0}, /* 162 */
-	{"acct",	NULL,			NOSYS_NO_EQUIV,	0}, /* 163 */
+	{"acct",	lx_acct,		0,		1}, /* 163 */
 	{"settimeofday", NULL,			0,		2}, /* 164 */
 	{"mount",	lx_mount,		0,		5}, /* 165 */
 	{"umount2",	lx_umount2,		0,		2}, /* 166 */
diff --git a/usr/src/uts/common/brand/lx/sys/lx_brand.h b/usr/src/uts/common/brand/lx/sys/lx_brand.h
index 6912f7a51c..ec5fedcff8 100644
--- a/usr/src/uts/common/brand/lx/sys/lx_brand.h
+++ b/usr/src/uts/common/brand/lx/sys/lx_brand.h
@@ -710,6 +710,7 @@ extern int lx_lpid_lock(pid_t, zone_t *, lx_pid_flag_t, proc_t **,
 extern pid_t lx_lwp_ppid(klwp_t *, pid_t *, id_t *);
 extern void lx_pid_init(void);
 extern void lx_pid_fini(void);
+extern void lx_acct_out(vnode_t *, int);
 
 extern uint_t lx_pipe_max_limit;
 extern uint_t lx_pipe_max_default;
diff --git a/usr/src/uts/common/brand/lx/sys/lx_syscalls.h b/usr/src/uts/common/brand/lx/sys/lx_syscalls.h
index 988230bbd0..e658a241e1 100644
--- a/usr/src/uts/common/brand/lx/sys/lx_syscalls.h
+++ b/usr/src/uts/common/brand/lx/sys/lx_syscalls.h
@@ -37,6 +37,7 @@ extern "C" {
 extern long lx_accept();
 extern long lx_accept4();
 extern long lx_access();
+extern long lx_acct();
 extern long lx_alarm();
 extern long lx_arch_prctl();
 extern long lx_bind();
diff --git a/usr/src/uts/common/brand/lx/syscall/lx_miscsys.c b/usr/src/uts/common/brand/lx/syscall/lx_miscsys.c
index 5ae59327e5..b7057d3958 100644
--- a/usr/src/uts/common/brand/lx/syscall/lx_miscsys.c
+++ b/usr/src/uts/common/brand/lx/syscall/lx_miscsys.c
@@ -10,7 +10,7 @@
  */
 
 /*
- * Copyright 2016 Joyent, Inc.
+ * Copyright 2017 Joyent, Inc.
  */
 
 #include <sys/systeminfo.h>
@@ -75,6 +75,8 @@ extern int chdir_proc(proc_t *, vnode_t *, boolean_t, boolean_t);
 extern int lookupname(char *, enum uio_seg, int, vnode_t **, vnode_t **);
 /* From uts/common/fs/fs_subr.c */
 extern int fs_need_estale_retry(int);
+/* From uts/common/os/acct.c */
+extern int sysacct(char *);
 
 /* The callback arguments when handling a FS clone group. */
 typedef struct {
@@ -411,3 +413,9 @@ lx_vhangup(void)
 	 */
 	return (0);
 }
+
+long
+lx_acct(char *p)
+{
+	return (sysacct(p));
+}
diff --git a/usr/src/uts/common/brand/sn1/sn1_brand.c b/usr/src/uts/common/brand/sn1/sn1_brand.c
index 92c0aec1e0..a2383ca076 100644
--- a/usr/src/uts/common/brand/sn1/sn1_brand.c
+++ b/usr/src/uts/common/brand/sn1/sn1_brand.c
@@ -104,7 +104,8 @@ struct brand_ops sn1_brops = {
 	NULL,				/* b_pagefault */
 	B_TRUE,				/* b_intp_parse_arg */
 	NULL,				/* b_clearbrand */
-	NULL				/* b_rpc_statd */
+	NULL,				/* b_rpc_statd */
+	NULL				/* b_acct_out */
 };
 
 #ifdef	sparc
diff --git a/usr/src/uts/common/brand/solaris10/s10_brand.c b/usr/src/uts/common/brand/solaris10/s10_brand.c
index c02fcdaef6..9da0855cc3 100644
--- a/usr/src/uts/common/brand/solaris10/s10_brand.c
+++ b/usr/src/uts/common/brand/solaris10/s10_brand.c
@@ -109,7 +109,8 @@ struct brand_ops s10_brops = {
 	NULL,				/* b_pagefault */
 	B_TRUE,				/* b_intp_parse_arg */
 	NULL,				/* b_clearbrand */
-	NULL				/* b_rpc_statd */
+	NULL,				/* b_rpc_statd */
+	NULL				/* b_acct_out */
 };
 
 #ifdef	sparc
diff --git a/usr/src/uts/common/os/acct.c b/usr/src/uts/common/os/acct.c
index e598e0d08d..891c4e0836 100644
--- a/usr/src/uts/common/os/acct.c
+++ b/usr/src/uts/common/os/acct.c
@@ -20,6 +20,7 @@
  */
 /*
  * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright 2017, Joyent, Inc.
  */
 
 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
@@ -47,6 +48,7 @@
 #include <sys/time.h>
 #include <sys/msacct.h>
 #include <sys/zone.h>
+#include <sys/brand.h>
 
 /*
  * Each zone has its own accounting settings (on or off) and associated
@@ -373,7 +375,7 @@ acct_compress(ulong_t t)
  * On exit, write a record on the accounting file.
  */
 void
-acct(char st)
+acct(int st)
 {
 	struct vnode *vp;
 	struct cred *cr;
@@ -402,6 +404,21 @@ acct(char st)
 	 * This only gets called from exit after all lwp's have exited so no
 	 * cred locking is needed.
 	 */
+
+	/* If there is a brand-specific hook, use it instead */
+	if (ZONE_IS_BRANDED(curzone) && ZBROP(curzone)->b_acct_out != NULL) {
+		ZBROP(curzone)->b_acct_out(vp, st);
+		mutex_exit(&ag->aclock);
+		return;
+	}
+
+	/*
+	 * The 'st' status value was traditionally masked this way by our
+	 * caller, but we now accept the unmasked value for brand handling.
+	 * Zones not using the brand hook mask the status here.
+	 */
+	st &= 0xff;
+
 	p = curproc;
 	ua = PTOU(p);
 	bcopy(ua->u_comm, ag->acctbuf.ac_comm, sizeof (ag->acctbuf.ac_comm));
diff --git a/usr/src/uts/common/os/exit.c b/usr/src/uts/common/os/exit.c
index 41f5f29eee..3edddcf61f 100644
--- a/usr/src/uts/common/os/exit.c
+++ b/usr/src/uts/common/os/exit.c
@@ -21,7 +21,7 @@
 
 /*
  * Copyright (c) 1988, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright 2016 Joyent, Inc.
+ * Copyright 2017 Joyent, Inc.
  */
 
 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
@@ -601,7 +601,7 @@ proc_exit(int why, int what)
 		semexit(p);
 	rv = wstat(why, what);
 
-	acct(rv & 0xff);
+	acct(rv);
 	exacct_commit_proc(p, rv);
 
 	/*
diff --git a/usr/src/uts/common/sys/acct.h b/usr/src/uts/common/sys/acct.h
index f00884681b..e01ad61025 100644
--- a/usr/src/uts/common/sys/acct.h
+++ b/usr/src/uts/common/sys/acct.h
@@ -22,6 +22,7 @@
 /*
  * Copyright 2014 Garrett D'Amore <garrett@damore.org>
  * Copyright (c) 1989, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright 2017 Joyent, Inc.
  */
 
 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
@@ -88,7 +89,7 @@ extern int acct(const char *);
 
 #if defined(_KERNEL)
 
-void	acct(char);
+void	acct(int);
 int	sysacct(char *);
 
 struct vnode;
diff --git a/usr/src/uts/common/sys/brand.h b/usr/src/uts/common/sys/brand.h
index e7d4fb894e..e33de24757 100644
--- a/usr/src/uts/common/sys/brand.h
+++ b/usr/src/uts/common/sys/brand.h
@@ -151,6 +151,7 @@ struct execa;
  * b_intp_parse_arg - Controls interpreter argument handling (allow 1 or all)
  * b_clearbrand - Perform any actions necessary when clearing the brand.
  * b_rpc_statd - Upcall to rpc.statd running within the zone
+ * b_acct_out - Output properly formatted accounting record
  */
 struct brand_ops {
 	void	(*b_init_brand_data)(zone_t *, kmutex_t *);
@@ -202,6 +203,7 @@ struct brand_ops {
 	boolean_t b_intp_parse_arg;
 	void	(*b_clearbrand)(proc_t *, boolean_t);
 	void	(*b_rpc_statd)(int, void *, void *);
+	void	(*b_acct_out)(struct vnode *, int);
 };
 
 /*
diff --git a/usr/src/uts/intel/Makefile.files b/usr/src/uts/intel/Makefile.files
index 65e7dec2f0..962ac24afa 100644
--- a/usr/src/uts/intel/Makefile.files
+++ b/usr/src/uts/intel/Makefile.files
@@ -300,6 +300,7 @@ SN1_BRAND_OBJS	=	sn1_brand.o sn1_brand_asm.o
 S10_BRAND_OBJS	=	s10_brand.o s10_brand_asm.o
 LX_BRAND_OBJS  =		\
 	lx_access.o		\
+	lx_acct.o		\
 	lx_acl.o		\
 	lx_aio.o		\
 	lx_archdep.o		\
diff --git a/usr/src/uts/intel/lx_brand/Makefile b/usr/src/uts/intel/lx_brand/Makefile
index 819c22566c..4eff474a49 100644
--- a/usr/src/uts/intel/lx_brand/Makefile
+++ b/usr/src/uts/intel/lx_brand/Makefile
@@ -69,7 +69,7 @@ AS_INC_PATH	+= -I$(UTSBASE)/i86pc/genassym/$(OBJS_DIR)
 CFLAGS		+= $(CCVERBOSE)
 
 LDFLAGS		+= -dy -Nexec/elfexec -Nfs/fifofs -Nfs/sockfs -Ndrv/ip \
-		    -Nfs/zfs -Nmisc/klmmod
+		    -Nfs/zfs -Nmisc/klmmod -Nsys/sysacct
 
 #
 # For now, disable these lint checks; maintainers should endeavor