OS-5725 allow swap to be "disabled" in LX

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

6 files changed, 111 insertions, 25 deletions

diff --git a/usr/src/lib/brand/lx/testing/ltp_skiplist b/usr/src/lib/brand/lx/testing/ltp_skiplist
index d4a732c33d..571717b23a 100644
--- a/usr/src/lib/brand/lx/testing/ltp_skiplist
+++ b/usr/src/lib/brand/lx/testing/ltp_skiplist
@@ -72,9 +72,7 @@ setfsuid04
 setfsuid04_16
 settimeofday01
 stime01
-swapoff01
 swapoff02
-swapon01
 swapon02
 swapon03
 switch01
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 750b81eed2..832686f17a 100644
--- a/usr/src/uts/common/brand/lx/os/lx_syscall.c
+++ b/usr/src/uts/common/brand/lx/os/lx_syscall.c
@@ -606,7 +606,7 @@ lx_sysent_t lx_sysent32[] = {
 	{"oldlstat",	NULL,			NOSYS_OBSOLETE,	0}, /* 84 */
 	{"readlink",	lx_readlink,		0,		3}, /* 85 */
 	{"uselib",	NULL,			NOSYS_KERNEL,	0}, /* 86 */
-	{"swapon",	NULL,			NOSYS_KERNEL,	0}, /* 87 */
+	{"swapon",	lx_swapon,		0,		2}, /* 87 */
 	{"reboot",	lx_reboot,		0,		4}, /* 88 */
 	{"readdir",	NULL,			0,		3}, /* 89 */
 	{"mmap",	NULL,			0,		6}, /* 90 */
@@ -634,7 +634,7 @@ lx_sysent_t lx_sysent32[] = {
 	{"idle",	NULL,			NOSYS_NO_EQUIV,	0}, /* 112 */
 	{"vm86old",	NULL,			NOSYS_OBSOLETE,	0}, /* 113 */
 	{"wait4",	lx_wait4,		0,		4}, /* 114 */
-	{"swapoff",	NULL,			NOSYS_KERNEL,	0}, /* 115 */
+	{"swapoff",	lx_swapoff,		0,		1}, /* 115 */
 	{"sysinfo",	lx_sysinfo32,		0,		1}, /* 116 */
 	{"ipc",		NULL,			0,		5}, /* 117 */
 	{"fsync",	NULL,			0,		1}, /* 118 */
@@ -1057,8 +1057,8 @@ lx_sysent_t lx_sysent64[] = {
 	{"settimeofday", NULL,			0,		2}, /* 164 */
 	{"mount",	lx_mount,		0,		5}, /* 165 */
 	{"umount2",	lx_umount2,		0,		2}, /* 166 */
-	{"swapon",	NULL,			NOSYS_KERNEL,	0}, /* 167 */
-	{"swapoff",	NULL,			NOSYS_KERNEL,	0}, /* 168 */
+	{"swapon",	lx_swapon,		0,		2}, /* 167 */
+	{"swapoff",	lx_swapoff,		0,		1}, /* 168 */
 	{"reboot",	lx_reboot,		0,		4}, /* 169 */
 	{"sethostname",	lx_sethostname,		0,		2}, /* 170 */
 	{"setdomainname", lx_setdomainname,	0,		2}, /* 171 */
diff --git a/usr/src/uts/common/brand/lx/procfs/lx_prvnops.c b/usr/src/uts/common/brand/lx/procfs/lx_prvnops.c
index ea518f646b..5f85c1440e 100644
--- a/usr/src/uts/common/brand/lx/procfs/lx_prvnops.c
+++ b/usr/src/uts/common/brand/lx/procfs/lx_prvnops.c
@@ -3713,12 +3713,16 @@ static void
 lxpr_read_meminfo(lxpr_node_t *lxpnp, lxpr_uiobuf_t *uiobuf)
 {
 	zone_t *zone = LXPTOZ(lxpnp);
-	int global = zone == global_zone;
+	lx_zone_data_t *lxzd = ztolxzd(zone);
 	long total_mem, free_mem, total_swap;
+	boolean_t swap_disabled;
 
 	ASSERT(lxpnp->lxpr_type == LXPR_MEMINFO);
+	ASSERT(zone->zone_brand == &lx_brand);
+	ASSERT(lxzd != NULL);
+	swap_disabled = lxzd->lxzd_swap_disabled;
 
-	if (global || zone->zone_phys_mem_ctl == UINT64_MAX) {
+	if (zone->zone_phys_mem_ctl == UINT64_MAX) {
 		total_mem = physmem * PAGESIZE;
 		free_mem = freemem * PAGESIZE;
 	} else {
@@ -3728,12 +3732,16 @@ lxpr_read_meminfo(lxpr_node_t *lxpnp, lxpr_uiobuf_t *uiobuf)
 			free_mem = 0;
 	}
 
-	if (global || zone->zone_max_swap_ctl == UINT64_MAX) {
-		total_swap = k_anoninfo.ani_max * PAGESIZE;
+	if (swap_disabled) {
+		total_swap = 0;
 	} else {
-		mutex_enter(&zone->zone_mem_lock);
-		total_swap = zone->zone_max_swap_ctl;
-		mutex_exit(&zone->zone_mem_lock);
+		if (zone->zone_max_swap_ctl == UINT64_MAX) {
+			total_swap = k_anoninfo.ani_max * PAGESIZE;
+		} else {
+			mutex_enter(&zone->zone_mem_lock);
+			total_swap = zone->zone_max_swap_ctl;
+			mutex_exit(&zone->zone_mem_lock);
+		}
 	}
 
 	/*
@@ -4318,6 +4326,9 @@ lxpr_read_stat(lxpr_node_t *lxpnp, lxpr_uiobuf_t *uiobuf)
  * our entire swap cap as one swap partition. See lxpr_read_meminfo for an
  * explanation on why we report 0 used swap.
  *
+ * The zone's lxzd_swap_disabled boolean controls whether or not we pretend
+ * swap space is configured.
+ *
  * It is important to use formatting identical to the Linux implementation
  * so that consumers do not break. See swap_show() in mm/swapfile.c.
  */
@@ -4326,22 +4337,32 @@ static void
 lxpr_read_swaps(lxpr_node_t *lxpnp, lxpr_uiobuf_t *uiobuf)
 {
 	zone_t *zone = LXPTOZ(lxpnp);
-	uint64_t totswap, usedswap;
+	boolean_t swap_enabled;
+	lx_zone_data_t *lxzd = ztolxzd(zone);
 
-	if (zone->zone_max_swap_ctl == UINT64_MAX) {
-		totswap = (k_anoninfo.ani_max * PAGESIZE) >> 10;
-	} else {
-		mutex_enter(&zone->zone_mem_lock);
-		/* Uses units of 1 kb (2^10). */
-		totswap = zone->zone_max_swap_ctl >> 10;
-		mutex_exit(&zone->zone_mem_lock);
-	}
-	usedswap = 0;
+	ASSERT(zone->zone_brand == &lx_brand);
+	ASSERT(lxzd != NULL);
+	swap_enabled = !lxzd->lxzd_swap_disabled;
 
 	lxpr_uiobuf_printf(uiobuf,
 	    "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
-	lxpr_uiobuf_printf(uiobuf, "%-40s%s\t%llu\t%llu\t%d\n",
-	    "/dev/swap", "partition", totswap, usedswap, -1);
+
+	if (swap_enabled) {
+		uint64_t totswap, usedswap;
+
+		if (zone->zone_max_swap_ctl == UINT64_MAX) {
+			totswap = (k_anoninfo.ani_max * PAGESIZE) >> 10;
+		} else {
+			mutex_enter(&zone->zone_mem_lock);
+			/* Uses units of 1 kb (2^10). */
+			totswap = zone->zone_max_swap_ctl >> 10;
+			mutex_exit(&zone->zone_mem_lock);
+		}
+		usedswap = 0;
+
+		lxpr_uiobuf_printf(uiobuf, "%-40s%s\t%llu\t%llu\t%d\n",
+		    "/dev/swap", "partition", totswap, usedswap, -1);
+	}
 }
 
 /* ARGSUSED */
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 ec5fedcff8..9230876254 100644
--- a/usr/src/uts/common/brand/lx/sys/lx_brand.h
+++ b/usr/src/uts/common/brand/lx/sys/lx_brand.h
@@ -622,6 +622,7 @@ typedef struct lx_zone_data {
 	dev_t lxzd_zfs_dev;			/* major num for zfs */
 	uint_t lxzd_aio_nr;			/* see lx_aio.c */
 	uint_t lxzd_pipe_max_sz;		/* pipe-max-size sysctl val */
+	boolean_t lxzd_swap_disabled;		/* no fake swap in zone? */
 } lx_zone_data_t;
 
 /* LWP br_lwp_flags values */
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 31ceb0da08..c998adbca7 100644
--- a/usr/src/uts/common/brand/lx/sys/lx_syscalls.h
+++ b/usr/src/uts/common/brand/lx/sys/lx_syscalls.h
@@ -242,6 +242,8 @@ extern long lx_socketpair();
 extern long lx_stat32();
 extern long lx_stat64();
 extern long lx_stime();
+extern long lx_swapoff();
+extern long lx_swapon();
 extern long lx_sync();
 extern long lx_sync_file_range();
 extern long lx_syncfs();
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 23c9d78628..0a24ef9678 100644
--- a/usr/src/uts/common/brand/lx/syscall/lx_miscsys.c
+++ b/usr/src/uts/common/brand/lx/syscall/lx_miscsys.c
@@ -40,6 +40,17 @@
 #define	LX_RUSAGE_BOTH			(-2)
 #define	LX_RUSAGE_THREAD		1
 
+#define	LX_SWAP_PRIOMASK		0x7fff
+#define	LX_SWAP_PREFER			0x8000
+#define	LX_SWAP_DISCARD			0x10000
+#define	LX_SWAP_DISCARD_ONCE		0x20000
+#define	LX_SWAP_DISCARD_PAGES		0x40000
+
+#define	LX_SWAP_ALL			(LX_SWAP_DISCARD_PAGES | \
+					LX_SWAP_DISCARD_ONCE | \
+					LX_SWAP_DISCARD | \
+					LX_SWAP_PREFER | LX_SWAP_PRIOMASK)
+
 /* From uts/common/fs/vfs.c */
 extern void vfs_sync(int);
 /* From uts/common/os/grow.c */
@@ -435,3 +446,56 @@ lx_unshare(int flags)
 {
 	return (set_errno(EPERM));
 }
+
+/*
+ * The whole idea of "swap space" within a zone is a complete fabrication.
+ * However, some apps expect to be able to see swap space data in the /proc
+ * files, while other apps actually don't want there to be any swap space
+ * configured. We use the swapon/off syscalls to allow this visibility to be
+ * controlled from within the zone iself. Note that the "swapon" CLI tends to
+ * do a lot of additional validation which will fail within a zone.
+ *
+ * Once we have better Linux capabilities(7) support we should check
+ * CAP_SYS_ADMIN instead of uid == 0.
+ */
+long
+lx_swapoff(char *path)
+{
+	char buf[MAXPATHLEN];
+	size_t len;
+	lx_zone_data_t *lxzd;
+
+	/* Simple validaton of the argument */
+	if (copyinstr(path, buf, sizeof (buf), &len) != 0)
+		return (set_errno(EFAULT));
+	if (crgetuid(CRED()) != 0)
+		return (set_errno(EPERM));
+
+	lxzd = ztolxzd(curzone);
+	ASSERT(lxzd != NULL);
+
+	lxzd->lxzd_swap_disabled = B_TRUE;
+	return (0);
+}
+
+long
+lx_swapon(char *path, int flags)
+{
+	char buf[MAXPATHLEN];
+	size_t len;
+	lx_zone_data_t *lxzd;
+
+	/* Simple validaton of the arguments */
+	if (copyinstr(path, buf, sizeof (buf), &len) != 0)
+		return (set_errno(EFAULT));
+	if (flags & ~LX_SWAP_ALL)
+		return (set_errno(EINVAL));
+	if (crgetuid(CRED()) != 0)
+		return (set_errno(EPERM));
+
+	lxzd = ztolxzd(curzone);
+	ASSERT(lxzd != NULL);
+
+	lxzd->lxzd_swap_disabled = B_FALSE;
+	return (0);
+}