7382 basic NVMe 1.1 support

Reviewed by: Josef 'Jeff' Sipek <josef.sipek@nexenta.com>
Reviewed by: Rick McNeal <rick.mcneal@nexenta.com>
Reviewed by: Robert Mustacchi <rm@joyent.com>
Approved by: Garrett D'Amore <garrett@damore.org>

5 files changed, 145 insertions, 40 deletions

diff --git a/usr/src/man/man7d/nvme.7d b/usr/src/man/man7d/nvme.7d
index 70ae723058..03420fd19d 100644
--- a/usr/src/man/man7d/nvme.7d
+++ b/usr/src/man/man7d/nvme.7d
@@ -48,7 +48,7 @@ to attach to devices supporting newer version of the NVMe
 specification.
 The default value is 1, limiting
 .Nm
-to work with devices up to specification version 1.0.
+to work with devices up to specification version 1.1.
 .It Va ignore-unknown-vendor-status
 This can be set to 1 to allow
 .Nm
@@ -95,8 +95,9 @@ A controller number, typically one for each
 device found.
 Controller numbers are dynamically assigned by the system.
 .It Va tn
-The target number, this corresponds to the namespace ID used by the
-hardware.
+The target number.
+This is the namespace EUI64 if available, or the namespace ID used by
+the hardware.
 Namespace ID 0 is reserved, hence target numbers start with 1.
 .It Va sn
 This is the
diff --git a/usr/src/pkg/manifests/driver-storage-nvme.mf b/usr/src/pkg/manifests/driver-storage-nvme.mf
index 3296a3beef..87b0954b7f 100644
--- a/usr/src/pkg/manifests/driver-storage-nvme.mf
+++ b/usr/src/pkg/manifests/driver-storage-nvme.mf
@@ -14,7 +14,7 @@
 #
 
 #
-# Copyright 2015 Nexenta Systems, Inc. All rights reserved.
+# Copyright 2016 Nexenta Systems, Inc. All rights reserved.
 #
 
 #
@@ -25,7 +25,7 @@
 <include global_zone_only_component>
 set name=pkg.fmri value=pkg:/driver/storage/nvme@$(PKGVERS)
 set name=pkg.description \
-    value="Driver for Intel NVMe 1.0e compliant storage devices"
+    value="Driver for Intel NVMe 1.1b compliant storage devices"
 set name=pkg.summary value="NVMe driver"
 set name=info.classification \
     value=org.opensolaris.category.2008:System/Hardware
@@ -37,7 +37,7 @@ dir path=usr group=sys
 dir path=usr/share
 dir path=usr/share/man
 dir path=usr/share/man/man7d
-driver name=nvme alias=pciex8086,953 class=disk perms="* 0600 root sys"
+driver name=nvme alias=pciexclass,010802 class=disk perms="* 0600 root sys"
 file path=kernel/drv/$(ARCH64)/nvme group=sys
 file path=kernel/drv/nvme group=sys
 file path=kernel/drv/nvme.conf group=sys
diff --git a/usr/src/uts/common/io/nvme/nvme.c b/usr/src/uts/common/io/nvme/nvme.c
index aaa3b5fb73..3fad056534 100644
--- a/usr/src/uts/common/io/nvme/nvme.c
+++ b/usr/src/uts/common/io/nvme/nvme.c
@@ -18,7 +18,7 @@
 /*
  * blkdev driver for NVMe compliant storage devices
  *
- * This driver was written to conform to version 1.0e of the NVMe specification.
+ * This driver was written to conform to version 1.1b of the NVMe specification.
  * It may work with newer versions, but that is completely untested and disabled
  * by default.
  *
@@ -78,6 +78,11 @@
  * thin provisioning and protection information. This driver does not support
  * any of this and ignores namespaces that have these attributes.
  *
+ * As of NVMe 1.1 namespaces can have an 64bit Extended Unique Identifier
+ * (EUI64). This driver uses the EUI64 if present to generate the devid and
+ * passes it to blkdev to use it in the device node names. As this is currently
+ * untested namespaces with EUI64 are ignored by default.
+ *
  *
  * Blkdev Interface:
  *
@@ -93,8 +98,9 @@
  * Blkdev also supports querying device/media information and generating a
  * devid. The driver reports the best block size as determined by the namespace
  * format back to blkdev as physical block size to support partition and block
- * alignment. The devid is composed using the device vendor ID, model number,
- * serial number, and the namespace ID.
+ * alignment. The devid is either based on the namespace EUI64, if present, or
+ * composed using the device vendor ID, model number, serial number, and the
+ * namespace ID.
  *
  *
  * Error Handling:
@@ -140,7 +146,7 @@
  * The following driver properties can be changed to control some aspects of the
  * drivers operation:
  * - strict-version: can be set to 0 to allow devices conforming to newer
- *   versions to be used
+ *   versions or namespaces with EUI64 to be used
  * - ignore-unknown-vendor-status: can be set to 1 to not handle any vendor
  *   specific command status as a fatal error leading device faulting
  * - admin-queue-len: the maximum length of the admin queue (16-4096)
@@ -163,6 +169,11 @@
  * - support for media formatting and hard partitioning into namespaces
  * - support for big-endian systems
  * - support for fast reboot
+ * - support for firmware updates
+ * - support for NVMe Subsystem Reset (1.1)
+ * - support for Scatter/Gather lists (1.1)
+ * - support for Reservations (1.1)
+ * - support for power management
  */
 
 #include <sys/byteorder.h>
@@ -192,7 +203,7 @@
 
 /* NVMe spec version supported */
 static const int nvme_version_major = 1;
-static const int nvme_version_minor = 0;
+static const int nvme_version_minor = 1;
 
 /* tunable for admin command timeout in seconds, default is 1s */
 static volatile int nvme_admin_cmd_timeout = 1;
@@ -360,7 +371,7 @@ static struct dev_ops nvme_dev_ops = {
 
 static struct modldrv nvme_modldrv = {
 	.drv_modops	= &mod_driverops,
-	.drv_linkinfo	= "NVMe v1.0e",
+	.drv_linkinfo	= "NVMe v1.1b",
 	.drv_dev_ops	= &nvme_dev_ops
 };
 
@@ -1828,6 +1839,14 @@ nvme_shutdown(nvme_t *nvme, int mode, boolean_t quiesce)
 static void
 nvme_prepare_devid(nvme_t *nvme, uint32_t nsid)
 {
+	/*
+	 * Section 7.7 of the spec describes how to get a unique ID for
+	 * the controller: the vendor ID, the model name and the serial
+	 * number shall be unique when combined.
+	 *
+	 * If a namespace has no EUI64 we use the above and add the hex
+	 * namespace ID to get a unique ID for the namespace.
+	 */
 	char model[sizeof (nvme->n_idctl->id_model) + 1];
 	char serial[sizeof (nvme->n_idctl->id_serial) + 1];
 
@@ -1838,8 +1857,7 @@ nvme_prepare_devid(nvme_t *nvme, uint32_t nsid)
 	model[sizeof (nvme->n_idctl->id_model)] = '\0';
 	serial[sizeof (nvme->n_idctl->id_serial)] = '\0';
 
-	(void) snprintf(nvme->n_ns[nsid - 1].ns_devid,
-	    sizeof (nvme->n_ns[0].ns_devid), "%4X-%s-%s-%X",
+	nvme->n_ns[nsid - 1].ns_devid = kmem_asprintf("%4X-%s-%s-%X",
 	    nvme->n_idctl->id_vid, model, serial, nsid);
 }
 
@@ -1860,12 +1878,13 @@ nvme_init(nvme_t *nvme)
 
 	/* Check controller version */
 	vs.r = nvme_get32(nvme, NVME_REG_VS);
+	nvme->n_version.v_major = vs.b.vs_mjr;
+	nvme->n_version.v_minor = vs.b.vs_mnr;
 	dev_err(nvme->n_dip, CE_CONT, "?NVMe spec version %d.%d",
-	    vs.b.vs_mjr, vs.b.vs_mnr);
+	    nvme->n_version.v_major, nvme->n_version.v_minor);
 
-	if (nvme_version_major < vs.b.vs_mjr ||
-	    (nvme_version_major == vs.b.vs_mjr &&
-	    nvme_version_minor < vs.b.vs_mnr)) {
+	if (NVME_VERSION_HIGHER(&nvme->n_version,
+	    nvme_version_major, nvme_version_minor)) {
 		dev_err(nvme->n_dip, CE_WARN, "!no support for version > %d.%d",
 		    nvme_version_major, nvme_version_minor);
 		if (nvme->n_strict_version)
@@ -2164,7 +2183,26 @@ nvme_init(nvme_t *nvme)
 		    1 << idns->id_lbaf[idns->id_flbas.lba_format].lbaf_lbads;
 		nvme->n_ns[i].ns_best_block_size = nvme->n_ns[i].ns_block_size;
 
-		nvme_prepare_devid(nvme, nvme->n_ns[i].ns_id);
+		/*
+		 * Get the EUI64 if present. If not present prepare the devid
+		 * from other device data.
+		 */
+		if (NVME_VERSION_ATLEAST(&nvme->n_version, 1, 1))
+			bcopy(idns->id_eui64, nvme->n_ns[i].ns_eui64,
+			    sizeof (nvme->n_ns[i].ns_eui64));
+
+		/*LINTED: E_BAD_PTR_CAST_ALIGN*/
+		if (*(uint64_t *)nvme->n_ns[i].ns_eui64 == 0) {
+			nvme_prepare_devid(nvme, nvme->n_ns[i].ns_id);
+		} else {
+			/*
+			 * Until EUI64 support is tested on real hardware we
+			 * will ignore namespaces with an EUI64. This can
+			 * be overriden by setting strict-version=0 in nvme.conf
+			 */
+			if (nvme->n_strict_version)
+				nvme->n_ns[i].ns_ignore = B_TRUE;
+		}
 
 		/*
 		 * Find the LBA format with no metadata and the best relative
@@ -2662,6 +2700,8 @@ nvme_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
 			if (nvme->n_ns[i].ns_idns)
 				kmem_free(nvme->n_ns[i].ns_idns,
 				    sizeof (nvme_identify_nsid_t));
+			if (nvme->n_ns[i].ns_devid)
+				strfree(nvme->n_ns[i].ns_devid);
 		}
 
 		kmem_free(nvme->n_ns, sizeof (nvme_namespace_t) *
@@ -2889,6 +2929,7 @@ nvme_bd_driveinfo(void *arg, bd_drive_t *drive)
 	drive->d_removable = B_FALSE;
 	drive->d_hotpluggable = B_FALSE;
 
+	bcopy(ns->ns_eui64, drive->d_eui64, sizeof (drive->d_eui64));
 	drive->d_target = ns->ns_id;
 	drive->d_lun = 0;
 
@@ -2992,6 +3033,12 @@ nvme_bd_devid(void *arg, dev_info_t *devinfo, ddi_devid_t *devid)
 {
 	nvme_namespace_t *ns = arg;
 
-	return (ddi_devid_init(devinfo, DEVID_ENCAP, strlen(ns->ns_devid),
-	    ns->ns_devid, devid));
+	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
+	if (*(uint64_t *)ns->ns_eui64 != 0) {
+		return (ddi_devid_init(devinfo, DEVID_SCSI3_WWN,
+		    sizeof (ns->ns_eui64), ns->ns_eui64, devid));
+	} else {
+		return (ddi_devid_init(devinfo, DEVID_ENCAP,
+		    strlen(ns->ns_devid), ns->ns_devid, devid));
+	}
 }
diff --git a/usr/src/uts/common/io/nvme/nvme_reg.h b/usr/src/uts/common/io/nvme/nvme_reg.h
index 9c4259fa88..3e4b77079b 100644
--- a/usr/src/uts/common/io/nvme/nvme_reg.h
+++ b/usr/src/uts/common/io/nvme/nvme_reg.h
@@ -33,6 +33,22 @@ extern "C" {
 #define	NVME_MAX_ADMIN_QUEUE_LEN	4096
 
 /*
+ * NVMe version
+ */
+typedef struct {
+	uint16_t v_minor;
+	uint16_t v_major;
+} nvme_version_t;
+
+#define	NVME_VERSION_ATLEAST(v, maj, min) \
+	(((v)->v_major) > (maj) || \
+	((v)->v_major == (maj) && (v)->v_minor >= (min)))
+
+#define	NVME_VERSION_HIGHER(v, maj, min) \
+	(((v)->v_major) > (maj) || \
+	((v)->v_major == (maj) && (v)->v_minor > (min)))
+
+/*
  * NVMe registers and register fields
  */
 #define	NVME_REG_CAP	0x0		/* Controller Capabilities */
@@ -403,18 +419,21 @@ typedef struct {
 /* NVMe Power State Descriptor */
 typedef struct {
 	uint16_t psd_mp;		/* Maximum Power */
-	uint16_t psd_rsvd1;
+	uint8_t psd_rsvd1;
+	uint8_t psd_mps:1;		/* Max Power Scale (1.1) */
+	uint8_t psd_nops:1;		/* Non-Operational State (1.1) */
+	uint8_t psd_rsvd2:6;
 	uint32_t psd_enlat;		/* Entry Latency */
 	uint32_t psd_exlat;		/* Exit Latency */
 	uint8_t psd_rrt:5;		/* Relative Read Throughput */
-	uint8_t psd_rsvd2:3;
-	uint8_t psd_rrl:5;		/* Relative Read Latency */
 	uint8_t psd_rsvd3:3;
+	uint8_t psd_rrl:5;		/* Relative Read Latency */
+	uint8_t psd_rsvd4:3;
 	uint8_t psd_rwt:5;		/* Relative Write Throughput */
-	uint8_t	psd_rsvd4:3;
+	uint8_t	psd_rsvd5:3;
 	uint8_t psd_rwl:5;		/* Relative Write Latency */
-	uint8_t psd_rsvd5:3;
-	uint8_t psd_rsvd6[16];
+	uint8_t psd_rsvd6:3;
+	uint8_t psd_rsvd7[16];
 } nvme_idctl_psd_t;
 
 /* NVMe Identify Controller Data Structure */
@@ -428,11 +447,14 @@ typedef struct {
 	uint8_t id_rab;			/* Recommended Arbitration Burst */
 	uint8_t id_oui[3];		/* vendor IEEE OUI */
 	struct {			/* Multi-Interface Capabilities */
-		uint8_t m_multi:1;	/* HW has multiple PCIe interfaces */
-		uint8_t m_rsvd:7;
+		uint8_t m_multi_pci:1;	/* HW has multiple PCIe interfaces */
+		uint8_t m_multi_ctrl:1; /* HW has multiple controllers (1.1) */
+		uint8_t m_sr_iov:1;	/* controller is SR-IOV virt fn (1.1) */
+		uint8_t m_rsvd:5;
 	} id_mic;
 	uint8_t	id_mdts;		/* Maximum Data Transfer Size */
-	uint8_t id_rsvd_cc[256 - 78];
+	uint16_t id_cntlid;		/* Unique Controller Identifier (1.1) */
+	uint8_t id_rsvd_cc[256 - 80];
 
 	/* Admin Command Set Attributes */
 	struct {			/* Optional Admin Command Support */
@@ -458,7 +480,11 @@ typedef struct {
 		uint8_t av_spec:1;	/* use format from spec */
 		uint8_t av_rsvd:7;
 	} id_avscc;
-	uint8_t id_rsvd_ac[256 - 9];
+	struct {			/* Autonomous Power State Trans (1.1) */
+		uint8_t ap_sup:1;	/* APST supported (1.1) */
+		uint8_t ap_rsvd:7;
+	} id_apsta;
+	uint8_t id_rsvd_ac[256 - 10];
 
 	/* NVM Command Set Attributes */
 	nvme_idctl_qes_t id_sqes;	/* Submission Queue Entry Size */
@@ -469,7 +495,10 @@ typedef struct {
 		uint16_t on_compare:1;	/* Compare */
 		uint16_t on_wr_unc:1;	/* Write Uncorrectable */
 		uint16_t on_dset_mgmt:1; /* Dataset Management */
-		uint16_t on_rsvd:13;
+		uint16_t on_wr_zero:1;	/* Write Zeros (1.1) */
+		uint16_t on_save:1;	/* Save/Select in Get/Set Feat (1.1) */
+		uint16_t on_reserve:1;	/* Reservations (1.1) */
+		uint16_t on_rsvd:10;
 	} id_oncs;
 	struct {			/* Fused Operation Support */
 		uint16_t f_cmp_wr:1;	/* Compare and Write */
@@ -491,7 +520,16 @@ typedef struct {
 		uint8_t nv_spec:1;	/* use format from spec */
 		uint8_t nv_rsvd:7;
 	} id_nvscc;
-	uint8_t id_rsvd_nc_2[192 - 19];
+	uint8_t id_rsvd_nc_2;
+	uint16_t id_acwu;		/* Atomic Compare & Write Unit (1.1) */
+	uint16_t id_rsvd_nc_3;
+	struct {			/* SGL Support (1.1) */
+		uint16_t sgl_sup:1;	/* SGL Supported in NVM cmds (1.1) */
+		uint16_t sgl_rsvd1:15;
+		uint16_t sgl_bucket:1;	/* SGL Bit Bucket supported (1.1) */
+		uint16_t sgl_rsvd2:15;
+	} id_sgls;
+	uint8_t id_rsvd_nc_4[192 - 28];
 
 	/* I/O Command Set Attributes */
 	uint8_t id_rsvd_ioc[1344];
@@ -537,12 +575,29 @@ typedef struct {
 		uint8_t dp_type3:1;	/* Protection Information Type 3 */
 		uint8_t dp_first:1;	/* first 8 bytes of metadata */
 		uint8_t dp_last:1;	/* last 8 bytes of metadata */
+		uint8_t dp_rsvd:3;
 	} id_dpc;
 	struct {			/* Data Protection Settings */
 		uint8_t dp_pinfo:3;	/* Protection Information enabled */
 		uint8_t dp_first:1;	/* first 8 bytes of metadata */
+		uint8_t dp_rsvd:4;
 	} id_dps;
-	uint8_t id_rsvd1[128 - 30];
+	struct {			/* NS Multi-Path/Sharing Cap (1.1) */
+		uint8_t nm_shared:1;	/* NS is shared (1.1) */
+		uint8_t nm_rsvd:7;
+	} id_nmic;
+	struct {			/* Reservation Capabilities (1.1) */
+		uint8_t rc_persist:1;	/* Persist Through Power Loss (1.1) */
+		uint8_t rc_wr_excl:1;	/* Write Exclusive (1.1) */
+		uint8_t rc_excl:1;	/* Exclusive Access (1.1) */
+		uint8_t rc_wr_excl_r:1;	/* Wr Excl - Registrants Only (1.1) */
+		uint8_t rc_excl_r:1;	/* Excl Acc - Registrants Only (1.1) */
+		uint8_t rc_wr_excl_a:1;	/* Wr Excl - All Registrants (1.1) */
+		uint8_t rc_excl_a:1;	/* Excl Acc - All Registrants (1.1) */
+		uint8_t rc_rsvd:1;
+	} id_rescap;
+	uint8_t id_rsvd1[120 - 32];
+	uint8_t id_eui64[8];		/* IEEE Extended Unique Id (1.1) */
 	nvme_idns_lbaf_t id_lbaf[16];	/* LBA Formats */
 
 	uint8_t id_rsvd2[192];
@@ -577,6 +632,8 @@ typedef union {
 #define	NVME_FEAT_INTR_VECT	0x9	/* Interrupt Vector Configuration */
 #define	NVME_FEAT_WRITE_ATOM	0xa	/* Write Atomicity */
 #define	NVME_FEAT_ASYNC_EVENT	0xb	/* Asynchronous Event Configuration */
+#define	NVME_FEAT_AUTO_PST	0xc	/* Autonomous Power State Transition */
+					/* (1.1) */
 
 #define	NVME_FEAT_PROGRESS	0x80	/* Software Progress Marker */
 
diff --git a/usr/src/uts/common/io/nvme/nvme_var.h b/usr/src/uts/common/io/nvme/nvme_var.h
index 0ddb1649db..fd6f93af88 100644
--- a/usr/src/uts/common/io/nvme/nvme_var.h
+++ b/usr/src/uts/common/io/nvme/nvme_var.h
@@ -124,6 +124,7 @@ struct nvme {
 	char *n_product;
 	char *n_vendor;
 
+	nvme_version_t n_version;
 	boolean_t n_dead;
 	boolean_t n_strict_version;
 	boolean_t n_ignore_unknown_vendor_status;
@@ -215,6 +216,8 @@ struct nvme {
 
 struct nvme_namespace {
 	nvme_t *ns_nvme;
+	uint8_t ns_eui64[8];
+
 	bd_handle_t ns_bd_hdl;
 
 	uint32_t ns_id;
@@ -227,13 +230,10 @@ struct nvme_namespace {
 	nvme_identify_nsid_t *ns_idns;
 
 	/*
-	 * Section 7.7 of the spec describes how to get a unique ID for
-	 * the controller: the vendor ID, the model name and the serial
-	 * number shall be unique when combined.
-	 *
-	 * We add the hex namespace ID to get a unique ID for the namespace.
+	 * If a namespace has no EUI64, we create a devid in
+	 * nvme_prepare_devid().
 	 */
-	char ns_devid[4 + 1 + 20 + 1 + 40 + 1 + 8 + 1];
+	char *ns_devid;
 };
 
 struct nvme_task_arg {