path: root/drivers/block/skd_main.c

/* Copyright 2012 STEC, Inc.
 *
 * This file is licensed under the terms of the 3-clause
 * BSD License (http://opensource.org/licenses/BSD-3-Clause)
 * or the GNU GPL-2.0 (http://www.gnu.org/licenses/gpl-2.0.html),
 * at your option. Both licenses are also available in the LICENSE file
 * distributed with this project. This file may not be copied, modified,
 * or distributed except in accordance with those terms.
 * Gordoni Waidhofer <gwaidhofer@stec-inc.com>
 * Initial Driver Design!
 * Thomas Swann <tswann@stec-inc.com>
 * Interrupt handling.
 * Ramprasad Chinthekindi <rchinthekindi@stec-inc.com>
 * biomode implementation.
 * Akhil Bhansali <abhansali@stec-inc.com>
 * Added support for DISCARD / FLUSH and FUA.
 */

#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/init.h>
#include <linux/pci.h>
#include <linux/slab.h>
#include <linux/spinlock.h>
#include <linux/blkdev.h>
#include <linux/sched.h>
#include <linux/interrupt.h>
#include <linux/compiler.h>
#include <linux/workqueue.h>
#include <linux/bitops.h>
#include <linux/delay.h>
#include <linux/time.h>
#include <linux/hdreg.h>
#include <linux/dma-mapping.h>
#include <linux/completion.h>
#include <linux/scatterlist.h>
#include <linux/version.h>
#include <linux/err.h>
#include <linux/scatterlist.h>
#include <linux/aer.h>
#include <linux/ctype.h>
#include <linux/wait.h>
#include <linux/uio.h>
#include <scsi/scsi.h>
#include <scsi/sg.h>
#include <linux/io.h>
#include <linux/uaccess.h>
#include <asm/unaligned.h>

#include "skd_s1120.h"

static int skd_dbg_level;
static int skd_isr_comp_limit = 4;

enum {
	STEC_LINK_2_5GTS = 0,
	STEC_LINK_5GTS = 1,
	STEC_LINK_8GTS = 2,
	STEC_LINK_UNKNOWN = 0xFF
};

enum {
	SKD_FLUSH_INITIALIZER,
	SKD_FLUSH_ZERO_SIZE_FIRST,
	SKD_FLUSH_DATA_SECOND,
};

#define SKD_ASSERT(expr) \
	do { \
		if (unlikely(!(expr))) { \
			pr_err("Assertion failed! %s,%s,%s,line=%d\n",	\
			       # expr, __FILE__, __func__, __LINE__); \
		} \
	} while (0)

#define DRV_NAME "skd"
#define DRV_VERSION "2.2.1"
#define DRV_BUILD_ID "0260"
#define PFX DRV_NAME ": "
#define DRV_BIN_VERSION 0x100
#define DRV_VER_COMPL   "2.2.1." DRV_BUILD_ID

MODULE_AUTHOR("bug-reports: support@stec-inc.com");
MODULE_LICENSE("Dual BSD/GPL");

MODULE_DESCRIPTION("STEC s1120 PCIe SSD block driver (b" DRV_BUILD_ID ")");
MODULE_VERSION(DRV_VERSION "-" DRV_BUILD_ID);

#define PCI_VENDOR_ID_STEC      0x1B39
#define PCI_DEVICE_ID_S1120     0x0001

#define SKD_FUA_NV		(1 << 1)
#define SKD_MINORS_PER_DEVICE   16

#define SKD_MAX_QUEUE_DEPTH     200u

#define SKD_PAUSE_TIMEOUT       (5 * 1000)

#define SKD_N_FITMSG_BYTES      (512u)

#define SKD_N_SPECIAL_CONTEXT   32u
#define SKD_N_SPECIAL_FITMSG_BYTES      (128u)

/* SG elements are 32 bytes, so we can make this 4096 and still be under the
 * 128KB limit.  That allows 4096*4K = 16M xfer size
 */
#define SKD_N_SG_PER_REQ_DEFAULT 256u
#define SKD_N_SG_PER_SPECIAL    256u

#define SKD_N_COMPLETION_ENTRY  256u
#define SKD_N_READ_CAP_BYTES    (8u)

#define SKD_N_INTERNAL_BYTES    (512u)

/* 5 bits of uniqifier, 0xF800 */
#define SKD_ID_INCR             (0x400)
#define SKD_ID_TABLE_MASK       (3u << 8u)
#define  SKD_ID_RW_REQUEST      (0u << 8u)
#define  SKD_ID_INTERNAL        (1u << 8u)
#define  SKD_ID_SPECIAL_REQUEST (2u << 8u)
#define  SKD_ID_FIT_MSG         (3u << 8u)
#define SKD_ID_SLOT_MASK        0x00FFu
#define SKD_ID_SLOT_AND_TABLE_MASK 0x03FFu

#define SKD_N_TIMEOUT_SLOT      4u
#define SKD_TIMEOUT_SLOT_MASK   3u

#define SKD_N_MAX_SECTORS 2048u

#define SKD_MAX_RETRIES 2u

#define SKD_TIMER_SECONDS(seconds) (seconds)
#define SKD_TIMER_MINUTES(minutes) ((minutes) * (60))

#define INQ_STD_NBYTES 36
#define SKD_DISCARD_CDB_LENGTH	24

enum skd_drvr_state {
	SKD_DRVR_STATE_LOAD,
	SKD_DRVR_STATE_IDLE,
	SKD_DRVR_STATE_BUSY,
	SKD_DRVR_STATE_STARTING,
	SKD_DRVR_STATE_ONLINE,
	SKD_DRVR_STATE_PAUSING,
	SKD_DRVR_STATE_PAUSED,
	SKD_DRVR_STATE_DRAINING_TIMEOUT,
	SKD_DRVR_STATE_RESTARTING,
	SKD_DRVR_STATE_RESUMING,
	SKD_DRVR_STATE_STOPPING,
	SKD_DRVR_STATE_FAULT,
	SKD_DRVR_STATE_DISAPPEARED,
	SKD_DRVR_STATE_PROTOCOL_MISMATCH,
	SKD_DRVR_STATE_BUSY_ERASE,
	SKD_DRVR_STATE_BUSY_SANITIZE,
	SKD_DRVR_STATE_BUSY_IMMINENT,
	SKD_DRVR_STATE_WAIT_BOOT,
	SKD_DRVR_STATE_SYNCING,
};

#define SKD_WAIT_BOOT_TIMO      SKD_TIMER_SECONDS(90u)
#define SKD_STARTING_TIMO       SKD_TIMER_SECONDS(8u)
#define SKD_RESTARTING_TIMO     SKD_TIMER_MINUTES(4u)
#define SKD_DRAINING_TIMO       SKD_TIMER_SECONDS(6u)
#define SKD_BUSY_TIMO           SKD_TIMER_MINUTES(20u)
#define SKD_STARTED_BUSY_TIMO   SKD_TIMER_SECONDS(60u)
#define SKD_START_WAIT_SECONDS  90u

enum skd_req_state {
	SKD_REQ_STATE_IDLE,
	SKD_REQ_STATE_SETUP,
	SKD_REQ_STATE_BUSY,
	SKD_REQ_STATE_COMPLETED,
	SKD_REQ_STATE_TIMEOUT,
	SKD_REQ_STATE_ABORTED,
};

enum skd_fit_msg_state {
	SKD_MSG_STATE_IDLE,
	SKD_MSG_STATE_BUSY,
};

enum skd_check_status_action {
	SKD_CHECK_STATUS_REPORT_GOOD,
	SKD_CHECK_STATUS_REPORT_SMART_ALERT,
	SKD_CHECK_STATUS_REQUEUE_REQUEST,
	SKD_CHECK_STATUS_REPORT_ERROR,
	SKD_CHECK_STATUS_BUSY_IMMINENT,
};

struct skd_fitmsg_context {
	enum skd_fit_msg_state state;

	struct skd_fitmsg_context *next;

	u32 id;
	u16 outstanding;

	u32 length;
	u32 offset;

	u8 *msg_buf;
	dma_addr_t mb_dma_address;
};

struct skd_request_context {
	enum skd_req_state state;

	struct skd_request_context *next;

	u16 id;
	u32 fitmsg_id;

	struct request *req;
	u8 flush_cmd;
	u8 discard_page;

	u32 timeout_stamp;
	u8 sg_data_dir;
	struct scatterlist *sg;
	u32 n_sg;
	u32 sg_byte_count;

	struct fit_sg_descriptor *sksg_list;
	dma_addr_t sksg_dma_address;

	struct fit_completion_entry_v1 completion;

	struct fit_comp_error_info err_info;

};
#define SKD_DATA_DIR_HOST_TO_CARD       1
#define SKD_DATA_DIR_CARD_TO_HOST       2
#define SKD_DATA_DIR_NONE		3	/* especially for DISCARD requests. */

struct skd_special_context {
	struct skd_request_context req;

	u8 orphaned;

	void *data_buf;
	dma_addr_t db_dma_address;

	u8 *msg_buf;
	dma_addr_t mb_dma_address;
};

struct skd_sg_io {
	fmode_t mode;
	void __user *argp;

	struct sg_io_hdr sg;

	u8 cdb[16];

	u32 dxfer_len;
	u32 iovcnt;
	struct sg_iovec *iov;
	struct sg_iovec no_iov_iov;

	struct skd_special_context *skspcl;
};

typedef enum skd_irq_type {
	SKD_IRQ_LEGACY,
	SKD_IRQ_MSI,
	SKD_IRQ_MSIX
} skd_irq_type_t;

#define SKD_MAX_BARS                    2

struct skd_device {
	volatile void __iomem *mem_map[SKD_MAX_BARS];
	resource_size_t mem_phys[SKD_MAX_BARS];
	u32 mem_size[SKD_MAX_BARS];

	skd_irq_type_t irq_type;
	u32 msix_count;
	struct skd_msix_entry *msix_entries;

	struct pci_dev *pdev;
	int pcie_error_reporting_is_enabled;

	spinlock_t lock;
	struct gendisk *disk;
	struct request_queue *queue;
	struct device *class_dev;
	int gendisk_on;
	int sync_done;

	atomic_t device_count;
	u32 devno;
	u32 major;
	char name[32];
	char isr_name[30];

	enum skd_drvr_state state;
	u32 drive_state;

	u32 in_flight;
	u32 cur_max_queue_depth;
	u32 queue_low_water_mark;
	u32 dev_max_queue_depth;

	u32 num_fitmsg_context;
	u32 num_req_context;

	u32 timeout_slot[SKD_N_TIMEOUT_SLOT];
	u32 timeout_stamp;
	struct skd_fitmsg_context *skmsg_free_list;
	struct skd_fitmsg_context *skmsg_table;

	struct skd_request_context *skreq_free_list;
	struct skd_request_context *skreq_table;

	struct skd_special_context *skspcl_free_list;
	struct skd_special_context *skspcl_table;

	struct skd_special_context internal_skspcl;
	u32 read_cap_blocksize;
	u32 read_cap_last_lba;
	int read_cap_is_valid;
	int inquiry_is_valid;
	u8 inq_serial_num[13];  /*12 chars plus null term */
	u8 id_str[80];          /* holds a composite name (pci + sernum) */

	u8 skcomp_cycle;
	u32 skcomp_ix;
	struct fit_completion_entry_v1 *skcomp_table;
	struct fit_comp_error_info *skerr_table;
	dma_addr_t cq_dma_address;

	wait_queue_head_t waitq;

	struct timer_list timer;
	u32 timer_countdown;
	u32 timer_substate;

	int n_special;
	int sgs_per_request;
	u32 last_mtd;

	u32 proto_ver;

	int dbg_level;
	u32 connect_time_stamp;
	int connect_retries;
#define SKD_MAX_CONNECT_RETRIES 16
	u32 drive_jiffies;

	u32 timo_slot;


	struct work_struct completion_worker;
};

#define SKD_WRITEL(DEV, VAL, OFF) skd_reg_write32(DEV, VAL, OFF)
#define SKD_READL(DEV, OFF)      skd_reg_read32(DEV, OFF)
#define SKD_WRITEQ(DEV, VAL, OFF) skd_reg_write64(DEV, VAL, OFF)

static inline u32 skd_reg_read32(struct skd_device *skdev, u32 offset)
{
	u32 val;

	if (likely(skdev->dbg_level < 2))
		return readl(skdev->mem_map[1] + offset);
	else {
		barrier();
		val = readl(skdev->mem_map[1] + offset);
		barrier();
		pr_debug("%s:%s:%d offset %x = %x\n",
			 skdev->name, __func__, __LINE__, offset, val);
		return val;
	}

}

static inline void skd_reg_write32(struct skd_device *skdev, u32 val,
				   u32 offset)
{
	if (likely(skdev->dbg_level < 2)) {
		writel(val, skdev->mem_map[1] + offset);
		barrier();
	} else {
		barrier();
		writel(val, skdev->mem_map[1] + offset);
		barrier();
		pr_debug("%s:%s:%d offset %x = %x\n",
			 skdev->name, __func__, __LINE__, offset, val);
	}
}

static inline void skd_reg_write64(struct skd_device *skdev, u64 val,
				   u32 offset)
{
	if (likely(skdev->dbg_level < 2)) {
		writeq(val, skdev->mem_map[1] + offset);
		barrier();
	} else {
		barrier();
		writeq(val, skdev->mem_map[1] + offset);
		barrier();
		pr_debug("%s:%s:%d offset %x = %016llx\n",
			 skdev->name, __func__, __LINE__, offset, val);
	}
}


#define SKD_IRQ_DEFAULT SKD_IRQ_MSI
static int skd_isr_type = SKD_IRQ_DEFAULT;

module_param(skd_isr_type, int, 0444);
MODULE_PARM_DESC(skd_isr_type, "Interrupt type capability."
		 " (0==legacy, 1==MSI, 2==MSI-X, default==1)");

#define SKD_MAX_REQ_PER_MSG_DEFAULT 1
static int skd_max_req_per_msg = SKD_MAX_REQ_PER_MSG_DEFAULT;

module_param(skd_max_req_per_msg, int, 0444);
MODULE_PARM_DESC(skd_max_req_per_msg,
		 "Maximum SCSI requests packed in a single message."
		 " (1-14, default==1)");

#define SKD_MAX_QUEUE_DEPTH_DEFAULT 64
#define SKD_MAX_QUEUE_DEPTH_DEFAULT_STR "64"
static int skd_max_queue_depth = SKD_MAX_QUEUE_DEPTH_DEFAULT;

module_param(skd_max_queue_depth, int, 0444);
MODULE_PARM_DESC(skd_max_queue_depth,
		 "Maximum SCSI requests issued to s1120."
		 " (1-200, default==" SKD_MAX_QUEUE_DEPTH_DEFAULT_STR ")");

static int skd_sgs_per_request = SKD_N_SG_PER_REQ_DEFAULT;
module_param(skd_sgs_per_request, int, 0444);
MODULE_PARM_DESC(skd_sgs_per_request,
		 "Maximum SG elements per block request."
		 " (1-4096, default==256)");

static int skd_max_pass_thru = SKD_N_SPECIAL_CONTEXT;
module_param(skd_max_pass_thru, int, 0444);
MODULE_PARM_DESC(skd_max_pass_thru,
		 "Maximum SCSI pass-thru at a time." " (1-50, default==32)");

module_param(skd_dbg_level, int, 0444);
MODULE_PARM_DESC(skd_dbg_level, "s1120 debug level (0,1,2)");

module_param(skd_isr_comp_limit, int, 0444);
MODULE_PARM_DESC(skd_isr_comp_limit, "s1120 isr comp limit (0=none) default=4");

/* Major device number dynamically assigned. */
static u32 skd_major;

static void skd_destruct(struct skd_device *skdev);
static const struct block_device_operations skd_blockdev_ops;
static void skd_send_fitmsg(struct skd_device *skdev,
			    struct skd_fitmsg_context *skmsg);
static void skd_send_special_fitmsg(struct skd_device *skdev,
				    struct skd_special_context *skspcl);
static void skd_request_fn(struct request_queue *rq);
static void skd_end_request(struct skd_device *skdev,
			    struct skd_request_context *skreq, int error);
static int skd_preop_sg_list(struct skd_device *skdev,
			     struct skd_request_context *skreq);
static void skd_postop_sg_list(struct skd_device *skdev,
			       struct skd_request_context *skreq);

static void skd_restart_device(struct skd_device *skdev);
static int skd_quiesce_dev(struct skd_device *skdev);
static int skd_unquiesce_dev(struct skd_device *skdev);
static void skd_release_special(struct skd_device *skdev,
				struct skd_special_context *skspcl);
static void skd_disable_interrupts(struct skd_device *skdev);
static void skd_isr_fwstate(struct skd_device *skdev);
static void skd_recover_requests(struct skd_device *skdev, int requeue);
static void skd_soft_reset(struct skd_device *skdev);

static const char *skd_name(struct skd_device *skdev);
const char *skd_drive_state_to_str(int state);
const char *skd_skdev_state_to_str(enum skd_drvr_state state);
static void skd_log_skdev(struct skd_device *skdev, const char *event);
static void skd_log_skmsg(struct skd_device *skdev,
			  struct skd_fitmsg_context *skmsg, const char *event);
static void skd_log_skreq(struct skd_device *skdev,
			  struct skd_request_context *skreq, const char *event);

/*
 *****************************************************************************
 * READ/WRITE REQUESTS
 *****************************************************************************
 */
static void skd_fail_all_pending(struct skd_device *skdev)
{
	struct request_queue *q = skdev->queue;
	struct request *req;

	for (;; ) {
		req = blk_peek_request(q);
		if (req == NULL)
			break;
		blk_start_request(req);
		__blk_end_request_all(req, -EIO);
	}
}

static void
skd_prep_rw_cdb(struct skd_scsi_request *scsi_req,
		int data_dir, unsigned lba,
		unsigned count)
{
	if (data_dir == READ)
		scsi_req->cdb[0] = 0x28;
	else
		scsi_req->cdb[0] = 0x2a;

	scsi_req->cdb[1] = 0;
	scsi_req->cdb[2] = (lba & 0xff000000) >> 24;
	scsi_req->cdb[3] = (lba & 0xff0000) >> 16;
	scsi_req->cdb[4] = (lba & 0xff00) >> 8;
	scsi_req->cdb[5] = (lba & 0xff);
	scsi_req->cdb[6] = 0;
	scsi_req->cdb[7] = (count & 0xff00) >> 8;
	scsi_req->cdb[8] = count & 0xff;
	scsi_req->cdb[9] = 0;
}

static void
skd_prep_zerosize_flush_cdb(struct skd_scsi_request *scsi_req,
			    struct skd_request_context *skreq)
{
	skreq->flush_cmd = 1;

	scsi_req->cdb[0] = 0x35;
	scsi_req->cdb[1] = 0;
	scsi_req->cdb[2] = 0;
	scsi_req->cdb[3] = 0;
	scsi_req->cdb[4] = 0;
	scsi_req->cdb[5] = 0;
	scsi_req->cdb[6] = 0;
	scsi_req->cdb[7] = 0;
	scsi_req->cdb[8] = 0;
	scsi_req->cdb[9] = 0;
}

static void
skd_prep_discard_cdb(struct skd_scsi_request *scsi_req,
		     struct skd_request_context *skreq,
		     struct page *page,
		     u32 lba, u32 count)
{
	char *buf;
	unsigned long len;
	struct request *req;

	buf = page_address(page);
	len = SKD_DISCARD_CDB_LENGTH;

	scsi_req->cdb[0] = UNMAP;
	scsi_req->cdb[8] = len;

	put_unaligned_be16(6 + 16, &buf[0]);
	put_unaligned_be16(16, &buf[2]);
	put_unaligned_be64(lba, &buf[8]);
	put_unaligned_be32(count, &buf[16]);

	req = skreq->req;
	blk_add_request_payload(req, page, len);
}

static void skd_request_fn_not_online(struct request_queue *q);

static void skd_request_fn(struct request_queue *q)
{
	struct skd_device *skdev = q->queuedata;
	struct skd_fitmsg_context *skmsg = NULL;
	struct fit_msg_hdr *fmh = NULL;
	struct skd_request_context *skreq;
	struct request *req = NULL;
	struct skd_scsi_request *scsi_req;
	struct page *page;
	unsigned long io_flags;
	int error;
	u32 lba;
	u32 count;
	int data_dir;
	u32 be_lba;
	u32 be_count;
	u64 be_dmaa;
	u64 cmdctxt;
	u32 timo_slot;
	void *cmd_ptr;
	int flush, fua;

	if (skdev->state != SKD_DRVR_STATE_ONLINE) {
		skd_request_fn_not_online(q);
		return;
	}

	if (blk_queue_stopped(skdev->queue)) {
		if (skdev->skmsg_free_list == NULL ||
		    skdev->skreq_free_list == NULL ||
		    skdev->in_flight >= skdev->queue_low_water_mark)
			/* There is still some kind of shortage */
			return;

		queue_flag_clear(QUEUE_FLAG_STOPPED, skdev->queue);
	}

	/*
	 * Stop conditions:
	 *  - There are no more native requests
	 *  - There are already the maximum number of requests in progress
	 *  - There are no more skd_request_context entries
	 *  - There are no more FIT msg buffers
	 */
	for (;; ) {

		flush = fua = 0;

		req = blk_peek_request(q);

		/* Are there any native requests to start? */
		if (req == NULL)
			break;

		lba = (u32)blk_rq_pos(req);
		count = blk_rq_sectors(req);
		data_dir = rq_data_dir(req);
		io_flags = req->cmd_flags;

		if (io_flags & REQ_FLUSH)
			flush++;

		if (io_flags & REQ_FUA)
			fua++;

		pr_debug("%s:%s:%d new req=%p lba=%u(0x%x) "
			 "count=%u(0x%x) dir=%d\n",
			 skdev->name, __func__, __LINE__,
			 req, lba, lba, count, count, data_dir);

		/* At this point we know there is a request */

		/* Are too many requets already in progress? */
		if (skdev->in_flight >= skdev->cur_max_queue_depth) {
			pr_debug("%s:%s:%d qdepth %d, limit %d\n",
				 skdev->name, __func__, __LINE__,
				 skdev->in_flight, skdev->cur_max_queue_depth);
			break;
		}

		/* Is a skd_request_context available? */
		skreq = skdev->skreq_free_list;
		if (skreq == NULL) {
			pr_debug("%s:%s:%d Out of req=%p\n",
				 skdev->name, __func__, __LINE__, q);
			break;
		}
		SKD_ASSERT(skreq->state == SKD_REQ_STATE_IDLE);
		SKD_ASSERT((skreq->id & SKD_ID_INCR) == 0);

		/* Now we check to see if we can get a fit msg */
		if (skmsg == NULL) {
			if (skdev->skmsg_free_list == NULL) {
				pr_debug("%s:%s:%d Out of msg\n",
					 skdev->name, __func__, __LINE__);
				break;
			}
		}

		skreq->flush_cmd = 0;
		skreq->n_sg = 0;
		skreq->sg_byte_count = 0;
		skreq->discard_page = 0;

		/*
		 * OK to now dequeue request from q.
		 *
		 * At this point we are comitted to either start or reject
		 * the native request. Note that skd_request_context is
		 * available but is still at the head of the free list.
		 */
		blk_start_request(req);
		skreq->req = req;
		skreq->fitmsg_id = 0;

		/* Either a FIT msg is in progress or we have to start one. */
		if (skmsg == NULL) {
			/* Are there any FIT msg buffers available? */
			skmsg = skdev->skmsg_free_list;
			if (skmsg == NULL) {
				pr_debug("%s:%s:%d Out of msg skdev=%p\n",
					 skdev->name, __func__, __LINE__,
					 skdev);
				break;
			}
			SKD_ASSERT(skmsg->state == SKD_MSG_STATE_IDLE);
			SKD_ASSERT((skmsg->id & SKD_ID_INCR) == 0);

			skdev->skmsg_free_list = skmsg->next;

			skmsg->state = SKD_MSG_STATE_BUSY;
			skmsg->id += SKD_ID_INCR;

			/* Initialize the FIT msg header */
			fmh = (struct fit_msg_hdr *)skmsg->msg_buf;
			memset(fmh, 0, sizeof(*fmh));
			fmh->protocol_id = FIT_PROTOCOL_ID_SOFIT;
			skmsg->length = sizeof(*fmh);
		}

		skreq->fitmsg_id = skmsg->id;

		/*
		 * Note that a FIT msg may have just been started
		 * but contains no SoFIT requests yet.
		 */

		/*
		 * Transcode the request, checking as we go. The outcome of
		 * the transcoding is represented by the error variable.
		 */
		cmd_ptr = &skmsg->msg_buf[skmsg->length];
		memset(cmd_ptr, 0, 32);

		be_lba = cpu_to_be32(lba);
		be_count = cpu_to_be32(count);
		be_dmaa = cpu_to_be64((u64)skreq->sksg_dma_address);
		cmdctxt = skreq->id + SKD_ID_INCR;

		scsi_req = cmd_ptr;
		scsi_req->hdr.tag = cmdctxt;
		scsi_req->hdr.sg_list_dma_address = be_dmaa;

		if (data_dir == READ)
			skreq->sg_data_dir = SKD_DATA_DIR_CARD_TO_HOST;
		else
			skreq->sg_data_dir = SKD_DATA_DIR_HOST_TO_CARD;

		if (io_flags & REQ_DISCARD) {
			page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
			if (!page) {
				pr_err("request_fn:Page allocation failed.\n");
				skd_end_request(skdev, skreq, -ENOMEM);
				break;
			}
			skreq->discard_page = 1;
			req->completion_data = page;
			skd_prep_discard_cdb(scsi_req, skreq, page, lba, count);

		} else if (flush == SKD_FLUSH_ZERO_SIZE_FIRST) {
			skd_prep_zerosize_flush_cdb(scsi_req, skreq);
			SKD_ASSERT(skreq->flush_cmd == 1);

		} else {
			skd_prep_rw_cdb(scsi_req, data_dir, lba, count);
		}

		if (fua)
			scsi_req->cdb[1] |= SKD_FUA_NV;

		if (!req->bio)
			goto skip_sg;

		error = skd_preop_sg_list(skdev, skreq);

		if (error != 0) {
			/*
			 * Complete the native request with error.
			 * Note that the request context is still at the
			 * head of the free list, and that the SoFIT request
			 * was encoded into the FIT msg buffer but the FIT
			 * msg length has not been updated. In short, the
			 * only resource that has been allocated but might
			 * not be used is that the FIT msg could be empty.
			 */
			pr_debug("%s:%s:%d error Out\n",
				 skdev->name, __func__, __LINE__);
			skd_end_request(skdev, skreq, error);
			continue;
		}

skip_sg:
		scsi_req->hdr.sg_list_len_bytes =
			cpu_to_be32(skreq->sg_byte_count);

		/* Complete resource allocations. */
		skdev->skreq_free_list = skreq->next;
		skreq->state = SKD_REQ_STATE_BUSY;
		skreq->id += SKD_ID_INCR;

		skmsg->length += sizeof(struct skd_scsi_request);
		fmh->num_protocol_cmds_coalesced++;

		/*
		 * Update the active request counts.
		 * Capture the timeout timestamp.
		 */
		skreq->timeout_stamp = skdev->timeout_stamp;
		timo_slot = skreq->timeout_stamp & SKD_TIMEOUT_SLOT_MASK;
		skdev->timeout_slot[timo_slot]++;
		skdev->in_flight++;
		pr_debug("%s:%s:%d req=0x%x busy=%d\n",
			 skdev->name, __func__, __LINE__,
			 skreq->id, skdev->in_flight);

		/*
		 * If the FIT msg buffer is full send it.
		 */
		if (skmsg->length >= SKD_N_FITMSG_BYTES ||
		    fmh->num_protocol_cmds_coalesced >= skd_max_req_per_msg) {
			skd_send_fitmsg(skdev, skmsg);
			skmsg = NULL;
			fmh = NULL;
		}
	}

	/*
	 * Is a FIT msg in progress? If it is empty put the buffer back
	 * on the free list. If it is non-empty send what we got.
	 * This minimizes latency when there are fewer requests than
	 * what fits in a FIT msg.
	 */
	if (skmsg != NULL) {
		/* Bigger than just a FIT msg header? */
		if (skmsg->length > sizeof(struct fit_msg_hdr)) {
			pr_debug("%s:%s:%d sending msg=%p, len %d\n",
				 skdev->name, __func__, __LINE__,
				 skmsg, skmsg->length);
			skd_send_fitmsg(skdev, skmsg);
		} else {
			/*
			 * The FIT msg is empty. It means we got started
			 * on the msg, but the requests were rejected.
			 */
			skmsg->state = SKD_MSG_STATE_IDLE;
			skmsg->id += SKD_ID_INCR;
			skmsg->next = skdev->skmsg_free_list;
			skdev->skmsg_free_list = skmsg;
		}
		skmsg = NULL;
		fmh = NULL;
	}

	/*
	 * If req is non-NULL it means there is something to do but
	 * we are out of a resource.
	 */
	if (req)
		blk_stop_queue(skdev->queue);
}

static void skd_end_request(struct skd_device *skdev,
			    struct skd_request_context *skreq, int error)
{
	struct request *req = skreq->req;
	unsigned int io_flags = req->cmd_flags;

	if ((io_flags & REQ_DISCARD) &&
		(skreq->discard_page == 1)) {
		pr_debug("%s:%s:%d, free the page!",
			 skdev->name, __func__, __LINE__);
		__free_page(req->completion_data);
	}

	if (unlikely(error)) {
		struct request *req = skreq->req;
		char *cmd = (rq_data_dir(req) == READ) ? "read" : "write";
		u32 lba = (u32)blk_rq_pos(req);
		u32 count = blk_rq_sectors(req);

		pr_err("(%s): Error cmd=%s sect=%u count=%u id=0x%x\n",
		       skd_name(skdev), cmd, lba, count, skreq->id);
	} else
		pr_debug("%s:%s:%d id=0x%x error=%d\n",
			 skdev->name, __func__, __LINE__, skreq->id, error);

	__blk_end_request_all(skreq->req, error);
}

static int skd_preop_sg_list(struct skd_device *skdev,
			     struct skd_request_context *skreq)
{
	struct request *req = skreq->req;
	int writing = skreq->sg_data_dir == SKD_DATA_DIR_HOST_TO_CARD;
	int pci_dir = writing ? PCI_DMA_TODEVICE : PCI_DMA_FROMDEVICE;
	struct scatterlist *sg = &skreq->sg[0];
	int n_sg;
	int i;

	skreq->sg_byte_count = 0;

	/* SKD_ASSERT(skreq->sg_data_dir == SKD_DATA_DIR_HOST_TO_CARD ||
		   skreq->sg_data_dir == SKD_DATA_DIR_CARD_TO_HOST); */

	n_sg = blk_rq_map_sg(skdev->queue, req, sg);
	if (n_sg <= 0)
		return -EINVAL;

	/*
	 * Map scatterlist to PCI bus addresses.
	 * Note PCI might change the number of entries.
	 */
	n_sg = pci_map_sg(skdev->pdev, sg, n_sg, pci_dir);
	if (n_sg <= 0)
		return -EINVAL;

	SKD_ASSERT(n_sg <= skdev->sgs_per_request);

	skreq->n_sg = n_sg;

	for (i = 0; i < n_sg; i++) {
		struct fit_sg_descriptor *sgd = &skreq->sksg_list[i];
		u32 cnt = sg_dma_len(&sg[i]);
		uint64_t dma_addr = sg_dma_address(&sg[i]);

		sgd->control = FIT_SGD_CONTROL_NOT_LAST;
		sgd->byte_count = cnt;
		skreq->sg_byte_count += cnt;
		sgd->host_side_addr = dma_addr;
		sgd->dev_side_addr = 0;
	}

	skreq->sksg_list[n_sg - 1].next_desc_ptr = 0LL;
	skreq->sksg_list[n_sg - 1].control = FIT_SGD_CONTROL_LAST;

	if (unlikely(skdev->dbg_level > 1)) {
		pr_debug("%s:%s:%d skreq=%x sksg_list=%p sksg_dma=%llx\n",
			 skdev->name, __func__, __LINE__,
			 skreq->id, skreq->sksg_list, skreq->sksg_dma_address);
		for (i = 0; i < n_sg; i++) {
			struct fit_sg_descriptor *sgd = &skreq->sksg_list[i];
			pr_debug("%s:%s:%d   sg[%d] count=%u ctrl=0x%x "
				 "addr=0x%llx next=0x%llx\n",
				 skdev->name, __func__, __LINE__,
				 i, sgd->byte_count, sgd->control,
				 sgd->host_side_addr, sgd->next_desc_ptr);
		}
	}

	return 0;
}

static void skd_postop_sg_list(struct skd_device *skdev,
			       struct skd_request_context *skreq)
{
	int writing = skreq->sg_data_dir == SKD_DATA_DIR_HOST_TO_CARD;
	int pci_dir = writing ? PCI_DMA_TODEVICE : PCI_DMA_FROMDEVICE;

	/*
	 * restore the next ptr for next IO request so we
	 * don't have to set it every time.
	 */
	skreq->sksg_list[skreq->n_sg - 1].next_desc_ptr =
		skreq->sksg_dma_address +
		((skreq->n_sg) * sizeof(struct fit_sg_descriptor));
	pci_unmap_sg(skdev->pdev, &skreq->sg[0], skreq->n_sg, pci_dir);
}

static void skd_request_fn_not_online(struct request_queue *q)
{
	struct skd_device *skdev = q->queuedata;
	int error;

	SKD_ASSERT(skdev->state != SKD_DRVR_STATE_ONLINE);

	skd_log_skdev(skdev, "req_not_online");
	switch (skdev->state) {
	case SKD_DRVR_STATE_PAUSING:
	case SKD_DRVR_STATE_PAUSED:
	case SKD_DRVR_STATE_STARTING:
	case SKD_DRVR_STATE_RESTARTING:
	case SKD_DRVR_STATE_WAIT_BOOT:
	/* In case of starting, we haven't started the queue,
	 * so we can't get here... but requests are
	 * possibly hanging out waiting for us because we
	 * reported the dev/skd0 already.  They'll wait
	 * forever if connect doesn't complete.
	 * What to do??? delay dev/skd0 ??
	 */
	case SKD_DRVR_STATE_BUSY:
	case SKD_DRVR_STATE_BUSY_IMMINENT:
	case SKD_DRVR_STATE_BUSY_ERASE:
	case SKD_DRVR_STATE_DRAINING_TIMEOUT:
		return;

	case SKD_DRVR_STATE_BUSY_SANITIZE:
	case SKD_DRVR_STATE_STOPPING:
	case SKD_DRVR_STATE_SYNCING:
	case SKD_DRVR_STATE_FAULT:
	case SKD_DRVR_STATE_DISAPPEARED:
	default:
		error = -EIO;
		break;
	}

	/* If we get here, terminate all pending block requeusts
	 * with EIO and any scsi pass thru with appropriate sense
	 */

	skd_fail_all_pending(skdev);
}

/*
 *****************************************************************************
 * TIMER
 *****************************************************************************
 */

static void skd_timer_tick_not_online(struct skd_device *skdev);

static void skd_timer_tick(ulong arg)
{
	struct skd_device *skdev = (struct skd_device *)arg;

	u32 timo_slot;
	u32 overdue_timestamp;
	unsigned long reqflags;
	u32 state;

	if (skdev->state == SKD_DRVR_STATE_FAULT)
		/* The driver has declared fault, and we want it to
		 * stay that way until driver is reloaded.
		 */
		return;

	spin_lock_irqsave(&skdev->lock, reqflags);

	state = SKD_READL(skdev, FIT_STATUS);
	state &= FIT_SR_DRIVE_STATE_MASK;
	if (state != skdev->drive_state)
		skd_isr_fwstate(skdev);

	if (skdev->state != SKD_DRVR_STATE_ONLINE) {
		skd_timer_tick_not_online(skdev);
		goto timer_func_out;
	}
	skdev->timeout_stamp++;
	timo_slot = skdev->timeout_stamp & SKD_TIMEOUT_SLOT_MASK;

	/*
	 * All requests that happened during the previous use of
	 * this slot should be done by now. The previous use was
	 * over 7 seconds ago.
	 */
	if (skdev->timeout_slot[timo_slot] == 0)
		goto timer_func_out;

	/* Something is overdue */
	overdue_timestamp = skdev->timeout_stamp - SKD_N_TIMEOUT_SLOT;

	pr_debug("%s:%s:%d found %d timeouts, draining busy=%d\n",
		 skdev->name, __func__, __LINE__,
		 skdev->timeout_slot[timo_slot], skdev->in_flight);
	pr_err("(%s): Overdue IOs (%d), busy %d\n",
	       skd_name(skdev), skdev->timeout_slot[timo_slot],
	       skdev->in_flight);

	skdev->timer_countdown = SKD_DRAINING_TIMO;
	skdev->state = SKD_DRVR_STATE_DRAINING_TIMEOUT;
	skdev->timo_slot = timo_slot;
	blk_stop_queue(skdev->queue);

timer_func_out:
	mod_timer(&skdev->timer, (jiffies + HZ));

	spin_unlock_irqrestore(&skdev->lock, reqflags);
}

static void skd_timer_tick_not_online(struct skd_device *skdev)
{
	switch (skdev->state) {
	case SKD_DRVR_STATE_IDLE:
	case SKD_DRVR_STATE_LOAD:
		break;
	case SKD_DRVR_STATE_BUSY_SANITIZE:
		pr_debug("%s:%s:%d drive busy sanitize[%x], driver[%x]\n",
			 skdev->name, __func__, __LINE__,
			 skdev->drive_state, skdev->state);
		/* If we've been in sanitize for 3 seconds, we figure we're not
		 * going to get anymore completions, so recover requests now
		 */
		if (skdev->timer_countdown > 0) {
			skdev->timer_countdown--;
			return;
		}
		skd_recover_requests(skdev, 0);
		break;

	case SKD_DRVR_STATE_BUSY:
	case SKD_DRVR_STATE_BUSY_IMMINENT:
	case SKD_DRVR_STATE_BUSY_ERASE:
		pr_debug("%s:%s:%d busy[%x], countdown=%d\n",
			 skdev->name, __func__, __LINE__,
			 skdev->state, skdev->timer_countdown);
		if (skdev->timer_countdown > 0) {
			skdev->timer_countdown--;
			return;
		}
		pr_debug("%s:%s:%d busy[%x], timedout=%d, restarting device.",
			 skdev->name, __func__, __LINE__,
			 skdev->state, skdev->timer_countdown);
		skd_restart_device(skdev);
		break;

	case SKD_DRVR_STATE_WAIT_BOOT:
	case SKD_DRVR_STATE_STARTING:
		if (skdev->timer_countdown > 0) {
			skdev->timer_countdown--;
			return;
		}
		/* For now, we fault the drive.  Could attempt resets to
		 * revcover at some point. */
		skdev->state = SKD_DRVR_STATE_FAULT;

		pr_err("(%s): DriveFault Connect Timeout (%x)\n",
		       skd_name(skdev), skdev->drive_state);

		/*start the queue so we can respond with error to requests */
		/* wakeup anyone waiting for startup complete */
		blk_start_queue(skdev->queue);
		skdev->gendisk_on = -1;
		wake_up_interruptible(&skdev->waitq);
		break;

	case SKD_DRVR_STATE_ONLINE:
		/* shouldn't get here. */
		break;

	case SKD_DRVR_STATE_PAUSING:
	case SKD_DRVR_STATE_PAUSED:
		break;

	case SKD_DRVR_STATE_DRAINING_TIMEOUT:
		pr_debug("%s:%s:%d "
			 "draining busy [%d] tick[%d] qdb[%d] tmls[%d]\n",
			 skdev->name, __func__, __LINE__,
			 skdev->timo_slot,
			 skdev->timer_countdown,
			 skdev->in_flight,
			 skdev->timeout_slot[skdev->timo_slot]);
		/* if the slot has cleared we can let the I/O continue */
		if (skdev->timeout_slot[skdev->timo_slot] == 0) {
			pr_debug("%s:%s:%d Slot drained, starting queue.\n",
				 skdev->name, __func__, __LINE__);
			skdev->state = SKD_DRVR_STATE_ONLINE;
			blk_start_queue(skdev->queue);
			return;
		}
		if (skdev->timer_countdown > 0) {
			skdev->timer_countdown--;
			return;
		}
		skd_restart_device(skdev);
		break;

	case SKD_DRVR_STATE_RESTARTING:
		if (skdev->timer_countdown > 0) {
			skdev->timer_countdown--;
			return;
		}
		/* For now, we fault the drive. Could attempt resets to
		 * revcover at some point. */
		skdev->state = SKD_DRVR_STATE_FAULT;
		pr_err("(%s): DriveFault Reconnect Timeout (%x)\n",
		       skd_name(skdev), skdev->drive_state);

		/*
		 * Recovering does two things:
		 * 1. completes IO with error
		 * 2. reclaims dma resources
		 * When is it safe to recover requests?
		 * - if the drive state is faulted
		 * - if the state is still soft reset after out timeout
		 * - if the drive registers are dead (state = FF)
		 * If it is "unsafe", we still need to recover, so we will
		 * disable pci bus mastering and disable our interrupts.
		 */

		if ((skdev->drive_state == FIT_SR_DRIVE_SOFT_RESET) ||
		    (skdev->drive_state == FIT_SR_DRIVE_FAULT) ||
		    (skdev->drive_state == FIT_SR_DRIVE_STATE_MASK))
			/* It never came out of soft reset. Try to
			 * recover the requests and then let them
			 * fail. This is to mitigate hung processes. */
			skd_recover_requests(skdev, 0);
		else {
			pr_err("(%s): Disable BusMaster (%x)\n",
			       skd_name(skdev), skdev->drive_state);
			pci_disable_device(skdev->pdev);
			skd_disable_interrupts(skdev);
			skd_recover_requests(skdev, 0);
		}

		/*start the queue so we can respond with error to requests */
		/* wakeup anyone waiting for startup complete */
		blk_start_queue(skdev->queue);
		skdev->gendisk_on = -1;
		wake_up_interruptible(&skdev->waitq);
		break;

	case SKD_DRVR_STATE_RESUMING:
	case SKD_DRVR_STATE_STOPPING:
	case SKD_DRVR_STATE_SYNCING:
	case SKD_DRVR_STATE_FAULT:
	case SKD_DRVR_STATE_DISAPPEARED:
	default:
		break;
	}
}

static int skd_start_timer(struct skd_device *skdev)
{
	int rc;

	init_timer(&skdev->timer);
	setup_timer(&skdev->timer, skd_timer_tick, (ulong)skdev);

	rc = mod_timer(&skdev->timer, (jiffies + HZ));
	if (rc)
		pr_err("%s: failed to start timer %d\n",
		       __func__, rc);
	return rc;
}

static void skd_kill_timer(struct skd_device *skdev)
{
	del_timer_sync(&skdev->timer);
}

/*
 *****************************************************************************
 * IOCTL
 *****************************************************************************
 */
static int skd_ioctl_sg_io(struct skd_device *skdev,
			   fmode_t mode, void __user *argp);
static int skd_sg_io_get_and_check_args(struct skd_device *skdev,
					struct skd_sg_io *sksgio);
static int skd_sg_io_obtain_skspcl(struct skd_device *skdev,
				   struct skd_sg_io *sksgio);
static int skd_sg_io_prep_buffering(struct skd_device *skdev,
				    struct skd_sg_io *sksgio);
static int skd_sg_io_copy_buffer(struct skd_device *skdev,
				 struct skd_sg_io *sksgio, int dxfer_dir);
static int skd_sg_io_send_fitmsg(struct skd_device *skdev,
				 struct skd_sg_io *sksgio);
static int skd_sg_io_await(struct skd_device *skdev, struct skd_sg_io *sksgio);
static int skd_sg_io_release_skspcl(struct skd_device *skdev,
				    struct skd_sg_io *sksgio);
static int skd_sg_io_put_status(struct skd_device *skdev,
				struct skd_sg_io *sksgio);

static void skd_complete_special(struct skd_device *skdev,
				 volatile struct fit_completion_entry_v1
				 *skcomp,
				 volatile struct fit_comp_error_info *skerr,
				 struct skd_special_context *skspcl);

static int skd_bdev_ioctl(struct block_device *bdev, fmode_t mode,
			  uint cmd_in, ulong arg)
{
	int rc = 0;
	struct gendisk *disk = bdev->bd_disk;
	struct skd_device *skdev = disk->private_data;
	void __user *p = (void