OS-8136 Add DISCARD/TRIM support to bhyve (illumos specific bits) (#270)

Reviewed by: Mike Zeller <mike.zeller@joyent.com>
Reviewed by: Mike Gerdts <mike.gerdts@joyent.com>
Approved by: Mike Zeller <mike.zeller@joyent.com>

2 files changed, 58 insertions, 9 deletions

diff --git a/usr/src/cmd/bhyve/block_if.c b/usr/src/cmd/bhyve/block_if.c
index 42fa8c6109..85ad55b05a 100644
--- a/usr/src/cmd/bhyve/block_if.c
+++ b/usr/src/cmd/bhyve/block_if.c
@@ -364,9 +364,36 @@ blockif_proc(struct blockif_ctxt *bc, struct blockif_elem *be, uint8_t *buf)
 			else
 				br->br_resid = 0;
 		}
-#endif
 		else
 			 err = EOPNOTSUPP;
+#else
+		else if (bc->bc_ischr) {
+			dkioc_free_list_t dfl = {
+				.dfl_num_exts = 1,
+				.dfl_offset = 0,
+				.dfl_flags = 0,
+				.dfl_exts[0].dfle_start = br->br_offset,
+				.dfl_exts[0].dfle_length = br->br_resid
+			};
+
+			if (ioctl(bc->bc_fd, DKIOCFREE, &dfl))
+				err = errno;
+			else
+				br->br_resid = 0;
+		} else {
+			struct flock fl = {
+				.l_whence = 0,
+				.l_type = F_WRLCK,
+				.l_start = br->br_offset,
+				.l_len = br->br_resid
+			};
+
+			if (fcntl(bc->bc_fd, F_FREESP, &fl))
+				err = errno;
+			else
+				br->br_resid = 0;
+		}
+#endif
 		break;
 	default:
 		err = EINVAL;
@@ -475,9 +502,7 @@ blockif_open(const char *optstr, const char *ident)
 	off_t size, psectsz, psectoff;
 	int extra, fd, i, sectsz;
 	int nocache, sync, ro, candelete, geom, ssopt, pssopt;
-#ifdef __FreeBSD__
 	int nodelete;
-#endif
 
 #ifndef WITHOUT_CAPSICUM
 	cap_rights_t rights;
@@ -491,9 +516,7 @@ blockif_open(const char *optstr, const char *ident)
 	nocache = 0;
 	sync = 0;
 	ro = 0;
-#ifdef __FreeBSD__
 	nodelete = 0;
-#endif
 
 	/*
 	 * The first element in the optstring is always a pathname.
@@ -506,10 +529,8 @@ blockif_open(const char *optstr, const char *ident)
 			continue;
 		else if (!strcmp(cp, "nocache"))
 			nocache = 1;
-#ifdef __FreeBSD__
 		else if (!strcmp(cp, "nodelete"))
 			nodelete = 1;
-#endif
 		else if (!strcmp(cp, "sync") || !strcmp(cp, "direct"))
 			sync = 1;
 		else if (!strcmp(cp, "ro"))
@@ -630,6 +651,10 @@ blockif_open(const char *optstr, const char *ident)
 				}
 			}
 		}
+
+		if (nodelete == 0 && ioctl(fd, DKIOC_CANFREE, &candelete))
+			candelete = 0;
+
 	} else {
 		int flags;
 
@@ -639,6 +664,19 @@ blockif_open(const char *optstr, const char *ident)
 				wce = WCE_FCNTL;
 			}
 		}
+
+		/*
+		 * We don't have a way to discover if a file supports the
+		 * FREESP fcntl cmd (other than trying it).  However,
+		 * zfs, ufs, tmpfs, and udfs all support the FREESP fcntl cmd.
+		 * Nfsv4 and nfsv4 also forward the FREESP request
+		 * to the server, so we always enable it for file based
+		 * volumes. Anyone trying to run volumes on an unsupported
+		 * configuration is on their own, and should be prepared
+		 * for the requests to fail.
+		 */
+		if (nodelete == 0)
+			candelete = 1;
 	}
 #endif
 
diff --git a/usr/src/lib/brand/bhyve/zone/boot.c b/usr/src/lib/brand/bhyve/zone/boot.c
index 23bbb88a34..dadcf4e96a 100644
--- a/usr/src/lib/brand/bhyve/zone/boot.c
+++ b/usr/src/lib/brand/bhyve/zone/boot.c
@@ -228,6 +228,7 @@ add_disk(char *disk, char *path, char *slotconf, size_t slotconf_len)
 	const char *model = "virtio-blk";
 	uint_t pcibus = 0, pcidev = 0, pcifn = 0;
 	const char *slotstr;
+	const char *nodelstr = "";
 	boolean_t isboot;
 
 	isboot = is_env_true("device", disk, "boot");
@@ -262,6 +263,16 @@ add_disk(char *disk, char *path, char *slotconf, size_t slotconf_len)
 
 	if (is_env_string("device", disk, "model", "virtio")) {
 		model = "virtio-blk";
+		/*
+		 * bhyve's blockif code refers to the UNMAP/DISCARD/TRIM
+		 * feature as 'delete' and so 'nodelete' is used by
+		 * bhyve to disable the feature. We use 'trim' for
+		 * interfaces we expose to the operator as that seems to
+		 * be the most familiar name for the operation (and less
+		 * likely to cause confusion).
+		 */
+		if (is_env_string("device", disk, "notrim", "true"))
+			nodelstr = ",nodelete";
 	} else if (is_env_string("device", disk, "model", "ahci")) {
 		if (is_env_string("device", disk, "media", "cdrom")) {
 			model = "ahci-cd";
@@ -273,8 +284,8 @@ add_disk(char *disk, char *path, char *slotconf, size_t slotconf_len)
 		return (-1);
 	}
 
-	if (snprintf(slotconf, slotconf_len, "%u:%u:%u,%s,%s",
-	    pcibus, pcidev, pcifn, model, path) >= slotconf_len) {
+	if (snprintf(slotconf, slotconf_len, "%u:%u:%u,%s,%s%s",
+	    pcibus, pcidev, pcifn, model, path, nodelstr) >= slotconf_len) {
 		(void) printf("Error: disk path '%s' too long\n", path);
 		return (-1);
 	}