All of lore.kernel.org
 help / color / mirror / Atom feed
From: Douglas Gilbert <dgilbert@interlog.com>
To: linux-scsi@vger.kernel.org
Cc: martin.petersen@oracle.com, jejb@linux.vnet.ibm.com, hare@suse.de
Subject: [PATCH v17 44/45] sg: add blk_poll support
Date: Wed,  7 Apr 2021 21:45:30 -0400	[thread overview]
Message-ID: <20210408014531.248890-45-dgilbert@interlog.com> (raw)
In-Reply-To: <20210408014531.248890-1-dgilbert@interlog.com>

The support is added via the new SGV4_FLAG_HIPRI command flag which
causes REQ_HIPRI to be set on the request. Before waiting on an
inflight request, it is checked to see if it has SGV4_FLAG_HIPRI,
and if so blk_poll() is called instead of the wait. In situations
where only the file descriptor is known (e.g. sg_poll() and
ioctl(SG_GET_NUM_WAITING)) all inflight requests associated with
the file descriptor that have SGV4_FLAG_HIPRI set, have blk_poll()
called on them.

It is important to know blk_execute_rq_nowait() has finished before
sending blk_poll() on that request. The SG_RS_INFLIGHT state is set
just before blk_execute_rq_nowait() is called so a new bit setting
SG_FRQ_ISSUED has been added that is set just after that calls
returns.

Note that the implementation of blk_poll() calls mq_poll() in the
LLD associated with the request. Then for any request found to be
ready, blk_poll() invokes the scsi_done() callback. When blk_poll()
returns > 0 , sg_rq_end_io() may have been called on the given
request. If so the given request will be in await_rcv state.

Signed-off-by: Douglas Gilbert <dgilbert@interlog.com>
---
 drivers/scsi/sg.c      | 108 ++++++++++++++++++++++++++++++++++++++---
 include/uapi/scsi/sg.h |   1 +
 2 files changed, 103 insertions(+), 6 deletions(-)

diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
index 0982cb031be9..19aafd8e23f1 100644
--- a/drivers/scsi/sg.c
+++ b/drivers/scsi/sg.c
@@ -116,12 +116,14 @@ enum sg_rq_state {	/* N.B. sg_rq_state_arr assumes SG_RS_AWAIT_RCV==2 */
 #define SG_FRQ_RECEIVING	7	/* guard against multiple receivers */
 #define SG_FRQ_FOR_MMAP		8	/* request needs PAGE_SIZE elements */
 #define SG_FRQ_COUNT_ACTIVE	9	/* sfp->submitted + waiting active */
+#define SG_FRQ_ISSUED		10	/* blk_execute_rq_nowait() finished */
 
 /* Bit positions (flags) for sg_fd::ffd_bm bitmask follow */
 #define SG_FFD_FORCE_PACKID	0	/* receive only given pack_id/tag */
 #define SG_FFD_CMD_Q		1	/* clear: only 1 active req per fd */
 #define SG_FFD_KEEP_ORPHAN	2	/* policy for this fd */
-#define SG_FFD_Q_AT_TAIL	3	/* set: queue reqs at tail of blk q */
+#define SG_FFD_HIPRI_SEEN	3	/* could have HIPRI requests active */
+#define SG_FFD_Q_AT_TAIL	4	/* set: queue reqs at tail of blk q */
 
 /* Bit positions (flags) for sg_device::fdev_bm bitmask follow */
 #define SG_FDEV_EXCLUDE		0	/* have fd open with O_EXCL */
@@ -210,6 +212,7 @@ struct sg_request {	/* active SCSI command or inactive request */
 	int sense_len;		/* actual sense buffer length (data-in) */
 	atomic_t rq_st;		/* request state, holds a enum sg_rq_state */
 	u8 cmd_opcode;		/* first byte of SCSI cdb */
+	blk_qc_t cookie;	/* ids 1 or more queues for blk_poll() */
 	u64 start_ns;		/* starting point of command duration calc */
 	unsigned long frq_bm[1];	/* see SG_FRQ_* defines above */
 	u8 *sense_bp;		/* mempool alloc-ed sense buffer, as needed */
@@ -299,6 +302,9 @@ static struct sg_device *sg_get_dev(int min_dev);
 static void sg_device_destroy(struct kref *kref);
 static struct sg_request *sg_mk_srp_sgat(struct sg_fd *sfp, bool first,
 					 int db_len);
+static int sg_sfp_blk_poll(struct sg_fd *sfp, int loop_count);
+static int sg_srp_q_blk_poll(struct sg_request *srp, struct request_queue *q,
+			     int loop_count);
 #if IS_ENABLED(CONFIG_SCSI_LOGGING) && IS_ENABLED(SG_DEBUG)
 static const char *sg_rq_st_str(enum sg_rq_state rq_st, bool long_str);
 #endif
@@ -1008,6 +1014,7 @@ sg_execute_cmd(struct sg_fd *sfp, struct sg_request *srp)
 {
 	bool at_head, is_v4h, sync;
 	struct sg_device *sdp = sfp->parentdp;
+	struct request *rqq = READ_ONCE(srp->rqq);
 
 	is_v4h = test_bit(SG_FRQ_IS_V4I, srp->frq_bm);
 	sync = test_bit(SG_FRQ_SYNC_INVOC, srp->frq_bm);
@@ -1031,7 +1038,12 @@ sg_execute_cmd(struct sg_fd *sfp, struct sg_request *srp)
 		atomic_inc(&sfp->submitted);
 		set_bit(SG_FRQ_COUNT_ACTIVE, srp->frq_bm);
 	}
-	blk_execute_rq_nowait(sdp->disk, READ_ONCE(srp->rqq), (int)at_head, sg_rq_end_io);
+	if (srp->rq_flags & SGV4_FLAG_HIPRI) {
+		rqq->cmd_flags |= REQ_HIPRI;
+		srp->cookie = request_to_qc_t(rqq->mq_hctx, rqq);
+	}
+	blk_execute_rq_nowait(sdp->disk, rqq, (int)at_head, sg_rq_end_io);
+	set_bit(SG_FRQ_ISSUED, srp->frq_bm);
 }
 
 /*
@@ -1693,6 +1705,13 @@ sg_wait_event_srp(struct file *filp, struct sg_fd *sfp, void __user *p,
 
 	if (atomic_read(&srp->rq_st) != SG_RS_INFLIGHT)
 		goto skip_wait;		/* and skip _acquire() */
+	if (srp->rq_flags & SGV4_FLAG_HIPRI) {
+		/* call blk_poll(), spinning till found */
+		res = sg_srp_q_blk_poll(srp, sdp->device->request_queue, -1);
+		if (res != -ENODATA && unlikely(res < 0))
+			return res;
+		goto skip_wait;
+	}
 	SG_LOG(3, sfp, "%s: about to wait_event...()\n", __func__);
 	/* usually will be woken up by sg_rq_end_io() callback */
 	res = wait_event_interruptible(sfp->read_wait,
@@ -2019,6 +2038,8 @@ sg_ioctl_common(struct file *filp, struct sg_device *sdp, struct sg_fd *sfp,
 		SG_LOG(3, sfp, "%s:    SG_GET_PACK_ID=%d\n", __func__, val);
 		return put_user(val, ip);
 	case SG_GET_NUM_WAITING:
+		if (test_bit(SG_FFD_HIPRI_SEEN, sfp->ffd_bm))
+			sg_sfp_blk_poll(sfp, 0);	/* LLD may have some ready */
 		val = atomic_read(&sfp->waiting);
 		if (val)
 			return put_user(val, ip);
@@ -2228,6 +2249,69 @@ sg_compat_ioctl(struct file *filp, unsigned int cmd_in, unsigned long arg)
 }
 #endif
 
+/*
+ * If the sg_request object is not inflight, return -ENODATA. This function
+ * returns 1 if the given object was in inflight state and is in await_rcv
+ * state after blk_poll() returns 1 or more. If blk_poll() fails, then that
+ * (negative) value is returned. Otherwise returns 0. Note that blk_poll()
+ * may complete unrelated requests that share the same q and cookie.
+ */
+static int
+sg_srp_q_blk_poll(struct sg_request *srp, struct request_queue *q, int loop_count)
+{
+	int k, n, num;
+
+	num = (loop_count < 1) ? 1 : loop_count;
+	for (k = 0; k < num; ++k) {
+		if (atomic_read(&srp->rq_st) != SG_RS_INFLIGHT)
+			return -ENODATA;
+		n = blk_poll(q, srp->cookie, loop_count < 0 /* spin if negative */);
+		if (n > 0)
+			return atomic_read(&srp->rq_st) == SG_RS_AWAIT_RCV;
+		if (n < 0)
+			return n;
+	}
+	return 0;
+}
+
+/*
+ * Check all requests on this sfp that are both inflight and HIPRI. That check involves calling
+ * blk_poll(spin<-false) loop_count times. If loop_count is 0 then call blk_poll once.
+ * If loop_count is negative then call blk_poll(spin <- true)) once for each request.
+ * Returns number found (could be 0) or a negated errno value.
+ */
+static int
+sg_sfp_blk_poll(struct sg_fd *sfp, int loop_count)
+{
+	int res = 0;
+	int n;
+	unsigned long idx, iflags;
+	struct sg_request *srp;
+	struct scsi_device *sdev = sfp->parentdp->device;
+	struct request_queue *q = sdev ? sdev->request_queue : NULL;
+	struct xarray *xafp = &sfp->srp_arr;
+
+	if (!q)
+		return -EINVAL;
+	xa_lock_irqsave(xafp, iflags);
+	xa_for_each(xafp, idx, srp) {
+		if ((srp->rq_flags & SGV4_FLAG_HIPRI) &&
+		    atomic_read(&srp->rq_st) == SG_RS_INFLIGHT &&
+		    test_bit(SG_FRQ_ISSUED, srp->frq_bm)) {
+			xa_unlock_irqrestore(xafp, iflags);
+			n = sg_srp_q_blk_poll(srp, q, loop_count);
+			if (n == -ENODATA)
+				n = 0;
+			if (unlikely(n < 0))
+				return n;
+			xa_lock_irqsave(xafp, iflags);
+			res += n;
+		}
+	}
+	xa_unlock_irqrestore(xafp, iflags);
+	return res;
+}
+
 /*
  * Implements the poll(2) system call for this driver. Returns various EPOLL*
  * flags OR-ed together.
@@ -2239,6 +2323,8 @@ sg_poll(struct file *filp, poll_table * wait)
 	__poll_t p_res = 0;
 	struct sg_fd *sfp = filp->private_data;
 
+	if (test_bit(SG_FFD_HIPRI_SEEN, sfp->ffd_bm))
+		sg_sfp_blk_poll(sfp, 0);	/* LLD may have some ready to push up */
 	num = atomic_read(&sfp->waiting);
 	if (num < 1) {
 		poll_wait(filp, &sfp->read_wait, wait);
@@ -2523,6 +2609,7 @@ sg_rq_end_io(struct request *rqq, blk_status_t status)
 		}
 	}
 	xa_lock_irqsave(&sfp->srp_arr, iflags);
+	__set_bit(SG_FRQ_ISSUED, srp->frq_bm);
 	sg_rq_chg_state_force_ulck(srp, rqq_state);
 	WRITE_ONCE(srp->rqq, NULL);
 	if (test_bit(SG_FRQ_COUNT_ACTIVE, srp->frq_bm)) {
@@ -2548,7 +2635,8 @@ sg_rq_end_io(struct request *rqq, blk_status_t status)
 
 	if (likely(rqq_state == SG_RS_AWAIT_RCV)) {
 		/* Wake any sg_read()/ioctl(SG_IORECEIVE) awaiting this req */
-		wake_up_interruptible(&sfp->read_wait);
+		if (!(srp->rq_flags & SGV4_FLAG_HIPRI))
+			wake_up_interruptible(&sfp->read_wait);
 		kill_fasync(&sfp->async_qp, SIGPOLL, POLL_IN);
 		kref_put(&sfp->f_ref, sg_remove_sfp);
 	} else {        /* clean up orphaned request that aren't being kept */
@@ -2991,6 +3079,8 @@ sg_start_req(struct sg_request *srp, struct sg_comm_wr_t *cwrp, int dxfer_dir)
 	/* current sg_request protected by SG_RS_BUSY state */
 	scsi_rp = scsi_req(rqq);
 	WRITE_ONCE(srp->rqq, rqq);
+	if (rq_flags & SGV4_FLAG_HIPRI)
+		set_bit(SG_FFD_HIPRI_SEEN, sfp->ffd_bm);
 
 	if (cwrp->cmd_len > BLK_MAX_CDB)
 		scsi_rp->cmd = long_cmdp;	/* transfer ownership */
@@ -3100,7 +3190,10 @@ sg_finish_scsi_blk_rq(struct sg_request *srp)
 	SG_LOG(4, sfp, "%s: srp=0x%pK%s\n", __func__, srp,
 	       (srp->parentfp->rsv_srp == srp) ? " rsv" : "");
 	if (test_and_clear_bit(SG_FRQ_COUNT_ACTIVE, srp->frq_bm)) {
-		atomic_dec(&sfp->submitted);
+		bool now_zero = atomic_dec_and_test(&sfp->submitted);
+
+		if (now_zero)
+			clear_bit(SG_FFD_HIPRI_SEEN, sfp->ffd_bm);
 		atomic_dec(&sfp->waiting);
 	}
 
@@ -3297,6 +3390,8 @@ sg_find_srp_by_id(struct sg_fd *sfp, int pack_id)
 	struct sg_request *srp = NULL;
 	struct xarray *xafp = &sfp->srp_arr;
 
+	if (test_bit(SG_FFD_HIPRI_SEEN, sfp->ffd_bm))
+		sg_sfp_blk_poll(sfp, 0);	/* LLD may have some ready to push up */
 	if (num_waiting < 1) {
 		num_waiting = atomic_read_acquire(&sfp->waiting);
 		if (num_waiting < 1)
@@ -4101,8 +4196,9 @@ sg_proc_debug_sreq(struct sg_request *srp, int to, char *obp, int len)
 	else if (dur < U32_MAX)	/* in-flight or busy (so ongoing) */
 		n += scnprintf(obp + n, len - n, " t_o/elap=%us/%ums",
 			       to / 1000, dur);
-	n += scnprintf(obp + n, len - n, " sgat=%d op=0x%02x\n",
-		       srp->sgat_h.num_sgat, srp->cmd_opcode);
+	cp = (srp->rq_flags & SGV4_FLAG_HIPRI) ? "hipri " : "";
+	n += scnprintf(obp + n, len - n, " sgat=%d %sop=0x%02x\n",
+		       srp->sgat_h.num_sgat, cp, srp->cmd_opcode);
 	return n;
 }
 
diff --git a/include/uapi/scsi/sg.h b/include/uapi/scsi/sg.h
index 6162a5d5995c..11b58b279241 100644
--- a/include/uapi/scsi/sg.h
+++ b/include/uapi/scsi/sg.h
@@ -110,6 +110,7 @@ typedef struct sg_io_hdr {
 #define SGV4_FLAG_Q_AT_TAIL SG_FLAG_Q_AT_TAIL
 #define SGV4_FLAG_Q_AT_HEAD SG_FLAG_Q_AT_HEAD
 #define SGV4_FLAG_IMMED 0x400 /* for polling with SG_IOR, ignored in SG_IOS */
+#define SGV4_FLAG_HIPRI 0x800 /* request will use blk_poll to complete */
 
 /* Output (potentially OR-ed together) in v3::info or v4::info field */
 #define SG_INFO_OK_MASK 0x1
-- 
2.25.1


  parent reply	other threads:[~2021-04-08  1:46 UTC|newest]

Thread overview: 53+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-04-08  1:44 [PATCH v17 00/45] sg: add v4 interface Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 01/45] sg: move functions around Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 02/45] sg: remove typedefs, type+formatting cleanup Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 03/45] sg: sg_log and is_enabled Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 04/45] sg: rework sg_poll(), minor changes Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 05/45] sg: bitops in sg_device Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 06/45] sg: make open count an atomic Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 07/45] sg: move header to uapi section Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 08/45] sg: speed sg_poll and sg_get_num_waiting Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 09/45] sg: sg_allow_if_err_recovery and renames Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 10/45] sg: improve naming Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 11/45] sg: change rwlock to spinlock Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 12/45] sg: ioctl handling Douglas Gilbert
2021-04-08  1:44 ` [PATCH v17 13/45] sg: split sg_read Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 14/45] sg: sg_common_write add structure for arguments Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 15/45] sg: rework sg_vma_fault Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 16/45] sg: rework sg_mmap Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 17/45] sg: replace sg_allow_access Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 18/45] sg: rework scatter gather handling Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 19/45] sg: introduce request state machine Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 20/45] sg: sg_find_srp_by_id Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 21/45] sg: sg_fill_request_element Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 22/45] sg: printk change %p to %pK Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 23/45] sg: xarray for fds in device Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 24/45] sg: xarray for reqs in fd Douglas Gilbert
2021-04-08  8:05   ` Hannes Reinecke
2021-04-08  1:45 ` [PATCH v17 25/45] sg: replace rq array with xarray Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 26/45] sg: sense buffer rework Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 27/45] sg: add sg v4 interface support Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 28/45] sg: rework debug info Douglas Gilbert
2021-04-08  8:06   ` Hannes Reinecke
2021-04-08  1:45 ` [PATCH v17 29/45] sg: add 8 byte SCSI LUN to sg_scsi_id Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 30/45] sg: expand sg_comm_wr_t Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 31/45] sg: add sg_iosubmit_v3 and sg_ioreceive_v3 ioctls Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 32/45] sg: add some __must_hold macros Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 33/45] sg: move procfs objects to avoid forward decls Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 34/45] sg: protect multiple receivers Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 35/45] sg: first debugfs support Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 36/45] sg: rework mmap support Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 37/45] sg: defang allow_dio Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 38/45] sg: warn v3 write system call users Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 39/45] sg: add mmap_sz tracking Douglas Gilbert
2021-04-08  8:07   ` Hannes Reinecke
2021-04-08  1:45 ` [PATCH v17 40/45] sg: remove rcv_done request state Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 41/45] sg: track lowest inactive and await indexes Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 42/45] sg: remove unit attention check for device changed Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 43/45] sg: no_dxfer: move to/from kernel buffers Douglas Gilbert
2021-04-08  8:07   ` Hannes Reinecke
2021-04-08  1:45 ` Douglas Gilbert [this message]
2021-04-08  8:14   ` [PATCH v17 44/45] sg: add blk_poll support Hannes Reinecke
2021-04-08 16:28     ` Douglas Gilbert
2021-04-09  6:00       ` Douglas Gilbert
2021-04-08  1:45 ` [PATCH v17 45/45] sg: bump version to 4.0.12 Douglas Gilbert

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20210408014531.248890-45-dgilbert@interlog.com \
    --to=dgilbert@interlog.com \
    --cc=hare@suse.de \
    --cc=jejb@linux.vnet.ibm.com \
    --cc=linux-scsi@vger.kernel.org \
    --cc=martin.petersen@oracle.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.