Skip to contentSkip to conversation

remove two fields from struct request

PATCH In review Christoph Hellwig 6 participants 19 messages Nov 12, 2024 – Nov 26, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

Hi Jens,

this series removes two fields from struct request that just duplicate information in the bios. As-is it doesn't actually shrink the structure size but instead just creates a 4 byte hole, but that will probably become useful sooner or later.

Diffstat:

 block/blk-merge.c            |   26 ++++++++++++++------------
 block/blk-mq.c               |    5 +----
 drivers/scsi/sd.c            |    6 +++---
 include/linux/blk-mq.h       |    8 +++-----
 include/trace/events/block.h |    6 +++---
 5 files changed, 24 insertions(+), 27 deletions(-)
#1 lore raw

PATCH 1/2 block: remove the write_hint field from struct request

The write_hint is only used for read/write requests, which must have a bio attached to them. Just use the bio field instead.

---
 block/blk-merge.c      | 16 ++++++++++------
 block/blk-mq.c         |  2 --
 drivers/scsi/sd.c      |  6 +++---
 include/linux/blk-mq.h |  1 -
 4 files changed, 13 insertions(+), 12 deletions(-)
block/blk-merge.c +10 -6
index 7b0af8317c1c..2306014c108d 100644
@@ -867,9 +867,11 @@ static struct request *attempt_merge(struct request_queue *q,
 	if (rq_data_dir(req) != rq_data_dir(next))
 		return NULL;
 
-	/* Don't merge requests with different write hints. */
-	if (req->write_hint != next->write_hint)
-		return NULL;
+	if (req->bio && next->bio) {
+		/* Don't merge requests with different write hints. */
+		if (req->bio->bi_write_hint != next->bio->bi_write_hint)
+			return NULL;
+	}
 
 	if (req->ioprio != next->ioprio)
 		return NULL;
@@ -1001,9 +1003,11 @@ bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 	if (!bio_crypt_rq_ctx_compatible(rq, bio))
 		return false;
 
-	/* Don't merge requests with different write hints. */
-	if (rq->write_hint != bio->bi_write_hint)
-		return false;
+	if (rq->bio) {
+		/* Don't merge requests with different write hints. */
+		if (rq->bio->bi_write_hint != bio->bi_write_hint)
+			return false;
+	}
 
 	if (rq->ioprio != bio_prio(bio))
 		return false;
block/blk-mq.c +0 -2
index 5e240a4b6be0..65e6b86d341c 100644
@@ -2660,7 +2660,6 @@ static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
 		rq->cmd_flags |= REQ_FAILFAST_MASK;
 
 	rq->__sector = bio->bi_iter.bi_sector;
-	rq->write_hint = bio->bi_write_hint;
 	blk_rq_bio_prep(rq, bio, nr_segs);
 	if (bio_integrity(bio))
 		rq->nr_integrity_segments = blk_rq_count_integrity_sg(rq->q,
@@ -3308,7 +3307,6 @@ int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
 	}
 	rq->nr_phys_segments = rq_src->nr_phys_segments;
 	rq->ioprio = rq_src->ioprio;
-	rq->write_hint = rq_src->write_hint;
 
 	if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
 		goto free_and_out;
drivers/scsi/sd.c +3 -3
index ca4bc0ac76ad..8947dab132d7 100644
@@ -1190,8 +1190,8 @@ static u8 sd_group_number(struct scsi_cmnd *cmd)
 	if (!sdkp->rscs)
 		return 0;
 
-	return min3((u32)rq->write_hint, (u32)sdkp->permanent_stream_count,
-		    0x3fu);
+	return min3((u32)rq->bio->bi_write_hint,
+		    (u32)sdkp->permanent_stream_count, 0x3fu);
 }
 
 static blk_status_t sd_setup_rw32_cmnd(struct scsi_cmnd *cmd, bool write,
@@ -1389,7 +1389,7 @@ static blk_status_t sd_setup_read_write_cmnd(struct scsi_cmnd *cmd)
 		ret = sd_setup_rw16_cmnd(cmd, write, lba, nr_blocks,
 					 protect | fua, dld);
 	} else if ((nr_blocks > 0xff) || (lba > 0x1fffff) ||
-		   sdp->use_10_for_rw || protect || rq->write_hint) {
+		   sdp->use_10_for_rw || protect || rq->bio->bi_write_hint) {
 		ret = sd_setup_rw10_cmnd(cmd, write, lba, nr_blocks,
 					 protect | fua);
 	} else {
include/linux/blk-mq.h +0 -1
index 2035fad3131f..2804fe181d9d 100644
@@ -156,7 +156,6 @@ struct request {
 	struct blk_crypto_keyslot *crypt_keyslot;
 #endif
 
-	enum rw_hint write_hint;
 	unsigned short ioprio;
 
 	enum mq_rq_state state;
Signature · 1 line
2.45.2
#2 lore raw

Re: [PATCH 1/2] block: remove the write_hint field from struct request

block/blk-merge.c static struct request *attempt_merge(struct request_queue *q, · line 874

Quoted 8 lines · Christoph Hellwig
block/blk-merge.c +5 -3
-	/* Don't merge requests with different write hints. */
-	if (req->write_hint != next->write_hint)
-		return NULL;
+	if (req->bio && next->bio) {
+		/* Don't merge requests with different write hints. */
+		if (req->bio->bi_write_hint != next->bio->bi_write_hint)
+			return NULL;
+	}

The above two if-statements can be combined into a single if-statement.

block/blk-merge.c bool blk_rq_merge_ok(struct request *rq, struct bio *bio) · line 1010

Quoted 12 lines
block/blk-merge.c +5 -3
@@ -1001,9 +1003,11 @@ bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
  	if (!bio_crypt_rq_ctx_compatible(rq, bio))
  		return false;
 
-	/* Don't merge requests with different write hints. */
-	if (rq->write_hint != bio->bi_write_hint)
-		return false;
+	if (rq->bio) {
+		/* Don't merge requests with different write hints. */
+		if (rq->bio->bi_write_hint != bio->bi_write_hint)
+			return false;
+	}

Same comment here: the above two if-statements can also be combined into a single if-statement.

Otherwise this patch looks good to me.

Thanks,

Bart.

Reviewed-by
#3 lore raw

Re: [PATCH 1/2] block: remove the write_hint field from struct request

Quoted 27 lines · Bart Van Assche
Quoted 8 lines · Christoph Hellwig

-    /* Don't merge requests with different write hints. */ -    if (req->write_hint != next->write_hint)
-        return NULL;
+    if (req->bio && next->bio) {
+        /* Don't merge requests with different write hints. */ +        if (req->bio->bi_write_hint != next->bio->bi_write_hint) +            return NULL;
+    }

The above two if-statements can be combined into a single if-statement.

block/blk-merge.c bool blk_rq_merge_ok(struct request *rq, struct · line 1010

Quoted 12 lines
block/blk-merge.c +5 -3
@@ -1001,9 +1003,11 @@ bool blk_rq_merge_ok(struct request *rq, struct 
 bio *bio)
       if (!bio_crypt_rq_ctx_compatible(rq, bio))
           return false;
-    /* Don't merge requests with different write hints. */
-    if (rq->write_hint != bio->bi_write_hint)
-        return false;
+    if (rq->bio) {
+        /* Don't merge requests with different write hints. */
+        if (rq->bio->bi_write_hint != bio->bi_write_hint)
+            return false;
+    }

Same comment here: the above two if-statements can also be combined into a single if-statement.

Quoted 1 line

Otherwise this patch looks good to me.

(replying to my own email)

Since apparently the goal is to minimize diffs in the next patch,

#4 lore raw

PATCH 2/2 block: remove the ioprio field from struct request

In reply to Christoph Hellwig

The request ioprio is only initialized from the first attached bio, so requests without a bio already never set it. Directly use the bio field instead.

---
 block/blk-merge.c            | 10 ++++------
 block/blk-mq.c               |  3 +--
 include/linux/blk-mq.h       |  7 +++----
 include/trace/events/block.h |  6 +++---
 4 files changed, 11 insertions(+), 15 deletions(-)
block/blk-merge.c +4 -6
index 2306014c108d..df36f83f3738 100644
@@ -871,11 +871,10 @@ static struct request *attempt_merge(struct request_queue *q,
 		/* Don't merge requests with different write hints. */
 		if (req->bio->bi_write_hint != next->bio->bi_write_hint)
 			return NULL;
+		if (req->bio->bi_ioprio != next->bio->bi_ioprio)
+			return NULL;
 	}
 
-	if (req->ioprio != next->ioprio)
-		return NULL;
-
 	if (!blk_atomic_write_mergeable_rqs(req, next))
 		return NULL;
 
@@ -1007,11 +1006,10 @@ bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 		/* Don't merge requests with different write hints. */
 		if (rq->bio->bi_write_hint != bio->bi_write_hint)
 			return false;
+		if (rq->bio->bi_ioprio != bio->bi_ioprio)
+			return false;
 	}
 
-	if (rq->ioprio != bio_prio(bio))
-		return false;
-
 	if (blk_atomic_write_mergeable_rq_bio(rq, bio) == false)
 		return false;
 
block/blk-mq.c +1 -2
index 65e6b86d341c..3c6cadba75e3 100644
@@ -842,7 +842,7 @@ static void blk_print_req_error(struct request *req, blk_status_t status)
 		blk_op_str(req_op(req)),
 		(__force u32)(req->cmd_flags & ~REQ_OP_MASK),
 		req->nr_phys_segments,
-		IOPRIO_PRIO_CLASS(req->ioprio));
+		IOPRIO_PRIO_CLASS(req_get_ioprio(req)));
 }
 
 /*
@@ -3306,7 +3306,6 @@ int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
 		rq->special_vec = rq_src->special_vec;
 	}
 	rq->nr_phys_segments = rq_src->nr_phys_segments;
-	rq->ioprio = rq_src->ioprio;
 
 	if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
 		goto free_and_out;
include/linux/blk-mq.h +3 -4
index 2804fe181d9d..a28264442948 100644
@@ -156,8 +156,6 @@ struct request {
 	struct blk_crypto_keyslot *crypt_keyslot;
 #endif
 
-	unsigned short ioprio;
-
 	enum mq_rq_state state;
 	atomic_t ref;
 
@@ -221,7 +219,9 @@ static inline bool blk_rq_is_passthrough(struct request *rq)
 
 static inline unsigned short req_get_ioprio(struct request *req)
 {
-	return req->ioprio;
+	if (req->bio)
+		return req->bio->bi_ioprio;
+	return 0;
 }
 
 #define rq_data_dir(rq)		(op_is_write(req_op(rq)) ? WRITE : READ)
@@ -984,7 +984,6 @@ static inline void blk_rq_bio_prep(struct request *rq, struct bio *bio,
 	rq->nr_phys_segments = nr_segs;
 	rq->__data_len = bio->bi_iter.bi_size;
 	rq->bio = rq->biotail = bio;
-	rq->ioprio = bio_prio(bio);
 }
 
 void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,
include/trace/events/block.h +3 -3
index 1527d5d45e01..bd0ea07338eb 100644
@@ -99,7 +99,7 @@ TRACE_EVENT(block_rq_requeue,
 		__entry->dev	   = rq->q->disk ? disk_devt(rq->q->disk) : 0;
 		__entry->sector    = blk_rq_trace_sector(rq);
 		__entry->nr_sector = blk_rq_trace_nr_sectors(rq);
-		__entry->ioprio    = rq->ioprio;
+		__entry->ioprio    = req_get_ioprio(rq);
 
 		blk_fill_rwbs(__entry->rwbs, rq->cmd_flags);
 		__get_str(cmd)[0] = '\0';
@@ -136,7 +136,7 @@ DECLARE_EVENT_CLASS(block_rq_completion,
 		__entry->sector    = blk_rq_pos(rq);
 		__entry->nr_sector = nr_bytes >> 9;
 		__entry->error     = blk_status_to_errno(error);
-		__entry->ioprio    = rq->ioprio;
+		__entry->ioprio    = req_get_ioprio(rq);
 
 		blk_fill_rwbs(__entry->rwbs, rq->cmd_flags);
 		__get_str(cmd)[0] = '\0';
@@ -209,7 +209,7 @@ DECLARE_EVENT_CLASS(block_rq,
 		__entry->sector    = blk_rq_trace_sector(rq);
 		__entry->nr_sector = blk_rq_trace_nr_sectors(rq);
 		__entry->bytes     = blk_rq_bytes(rq);
-		__entry->ioprio	   = rq->ioprio;
+		__entry->ioprio	   = req_get_ioprio(rq);
 
 		blk_fill_rwbs(__entry->rwbs, rq->cmd_flags);
 		__get_str(cmd)[0] = '\0';
Signature · 1 line
2.45.2
#5 lore raw

Re: [PATCH 2/2] block: remove the ioprio field from struct request

block/blk-merge.c bool blk_rq_merge_ok(struct request *rq, struct bio *bio) · line 1009

Quoted 3 lines · Christoph Hellwig

if (rq->bio->bi_write_hint != bio->bi_write_hint)
return false;

block/blk-merge.c +1 -0
+		if (rq->bio->bi_ioprio != bio->bi_ioprio)

At this point bio_prio() looks to only be used in md code. Is it worth deleting that wrapper? I doubt its value.

And at many points bio->bi_ioprio is written without using bio_set_prio(), so I doubt the use of that one as well..

I do realize that all this is outside the scope of this series.

Quoted 6 lines
block/blk-merge.c +1 -3
+			return false;
  	}
 
-	if (rq->ioprio != bio_prio(bio))
-		return false;
-
#6 lore raw

Re: [PATCH 2/2] block: remove the ioprio field from struct request

Quoted 4 lines · John Garry

And at many points bio->bi_ioprio is written without using bio_set_prio(), so I doubt the use of that one as well..

I do realize that all this is outside the scope of this series.

Agreed to all of the above. If anyone is in for an easy cleanup, go for it!

#8 lore raw

Re: [PATCH 2/2] block: remove the ioprio field from struct request

In reply to Christoph Hellwig

Hi Christoph,

This patch causes a regression on E850-96 board. Specifically, there are two noticeable time lags when booting Debian rootfs:

1. When systemd reports this stage:

"Reached target getty.target - Login Prompts."

it hangs for 5 seconds or so, before going further.

2. When I'm logging in, the system hangs for 60 seconds or so before the shell prompt appears.

It only seems to reproduce the first time (during the boot). My attempt to re-start the mentioned systemd target or run "login" command again worked fine.

Reverting commit 6975c1a486a4 ("block: remove the ioprio field from struct request") fixes the issue for me. Do you have any ideas by chance what might be the reason? Or maybe you have any pointers on debugging it?

Thanks!

#9 lore raw

Re: [PATCH 2/2] block: remove the ioprio field from struct request

Quoted 4 lines · Sam Protsenko

Hi Christoph,

This patch causes a regression on E850-96 board. Specifically, there are two noticeable time lags when booting Debian rootfs:

What storage driver does this board use? Anything else interesting about the setup?

#10 lore raw

Re: [PATCH 2/2] block: remove the ioprio field from struct request

Quoted 10 lines · Christoph Hellwig
Quoted 4 lines · Sam Protsenko

Hi Christoph,

This patch causes a regression on E850-96 board. Specifically, there are two noticeable time lags when booting Debian rootfs:

What storage driver does this board use? Anything else interesting about the setup?

It's an Exynos based board with eMMC, so it uses DW MMC driver, with Exynos glue layer on top of it, so:

drivers/mmc/host/dw_mmc.c
drivers/mmc/host/dw_mmc-exynos.c

I'm using the regular ARM64 defconfig. Nothing fancy about this setup neither, the device tree with eMMC definition (mmc_0) is here:

    arch/arm64/boot/dts/exynos/exynos850-e850-96.dts

FWIW, I was able to narrow down the issue to dd_insert_request() function. With this hack the freeze is gone:

8<-------------------------------------------------------------------->8

block/mq-deadline.c +0 -0
index acdc28756d9d..83d272b66e71 100644
@@ -676,7 +676,7 @@ static void dd_insert_request(struct blk_mq_hw_ctx

*hctx, struct request *rq,

        struct request_queue *q = hctx->queue;
        struct deadline_data *dd = q->elevator->elevator_data;
        const enum dd_data_dir data_dir = rq_data_dir(rq);
-       u16 ioprio = req_get_ioprio(rq);
+       u16 ioprio = 0; /* the same as old req->ioprio */
        u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
        struct dd_per_prio *per_prio;
        enum dd_prio prio;

8<-------------------------------------------------------------------->8

Does it tell you anything about where the possible issue can be?

Thanks!

#11 lore raw

Re: [PATCH 2/2] block: remove the ioprio field from struct request

Quoted 41 lines · Sam Protsenko
Quoted 10 lines · Christoph Hellwig
Quoted 4 lines · Sam Protsenko

Hi Christoph,

This patch causes a regression on E850-96 board. Specifically, there are two noticeable time lags when booting Debian rootfs:

What storage driver does this board use? Anything else interesting about the setup?

It's an Exynos based board with eMMC, so it uses DW MMC driver, with Exynos glue layer on top of it, so:

drivers/mmc/host/dw_mmc.c
drivers/mmc/host/dw_mmc-exynos.c

I'm using the regular ARM64 defconfig. Nothing fancy about this setup neither, the device tree with eMMC definition (mmc_0) is here:

     arch/arm64/boot/dts/exynos/exynos850-e850-96.dts

FWIW, I was able to narrow down the issue to dd_insert_request() function. With this hack the freeze is gone:

8<-------------------------------------------------------------------->8

block/mq-deadline.c +1 -1
index acdc28756d9d..83d272b66e71 100644
@@ -676,7 +676,7 @@ static void dd_insert_request(struct blk_mq_hw_ctx
 *hctx, struct request *rq,
         struct request_queue *q = hctx->queue;
         struct deadline_data *dd = q->elevator->elevator_data;
         const enum dd_data_dir data_dir = rq_data_dir(rq);
-       u16 ioprio = req_get_ioprio(rq);
+       u16 ioprio = 0; /* the same as old req->ioprio */
         u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
         struct dd_per_prio *per_prio;
         enum dd_prio prio;
Quoted 3 lines

8<-------------------------------------------------------------------->8

Does it tell you anything about where the possible issue can be?

It seems like eMMC devices do not tolerate I/O prioritization. How about disabling I/O prioritization for eMMC setups? Is the ioprio cgroup controller perhaps activated by the user space software that is running on this setup?

Thanks,

Bart.

#12 lore raw

Re: [PATCH 2/2] block: remove the ioprio field from struct request

Hi Bart,

Quoted 45 lines · Bart Van Assche
Quoted 43 lines · Sam Protsenko
Quoted 10 lines · Christoph Hellwig
Quoted 4 lines · Sam Protsenko

Hi Christoph,

This patch causes a regression on E850-96 board. Specifically, there are two noticeable time lags when booting Debian rootfs:

What storage driver does this board use? Anything else interesting about the setup?

It's an Exynos based board with eMMC, so it uses DW MMC driver, with Exynos glue layer on top of it, so:

drivers/mmc/host/dw_mmc.c
drivers/mmc/host/dw_mmc-exynos.c

I'm using the regular ARM64 defconfig. Nothing fancy about this setup neither, the device tree with eMMC definition (mmc_0) is here:

     arch/arm64/boot/dts/exynos/exynos850-e850-96.dts

FWIW, I was able to narrow down the issue to dd_insert_request() function. With this hack the freeze is gone:

8<-------------------------------------------------------------------->8

block/mq-deadline.c +1 -1
index acdc28756d9d..83d272b66e71 100644
@@ -676,7 +676,7 @@ static void dd_insert_request(struct blk_mq_hw_ctx
 *hctx, struct request *rq,
         struct request_queue *q = hctx->queue;
         struct deadline_data *dd = q->elevator->elevator_data;
         const enum dd_data_dir data_dir = rq_data_dir(rq);
-       u16 ioprio = req_get_ioprio(rq);
+       u16 ioprio = 0; /* the same as old req->ioprio */
         u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
         struct dd_per_prio *per_prio;
         enum dd_prio prio;

8<-------------------------------------------------------------------->8

Quoted 7 lines
Quoted 1 line

Does it tell you anything about where the possible issue can be?

It seems like eMMC devices do not tolerate I/O prioritization. How about disabling I/O prioritization for eMMC setups? Is the ioprio cgroup controller perhaps activated by the user space software that is running on this setup?

Can you please elaborate on why eMMC devices might not play well with prios? Do they lack some particular hardware support, like required commands? Also, I've noticed (probably) the same issue reported recently [1] for USB SSD drives. so it'd nice to have some references showing it's actually the case specifically for eMMC.

Do you know if we have any config options to disable I/O prioritization? Not sure how exactly we can do that specifically for eMMC devices too. I'm not an expert in block layer, would appreciate some help with this one.

For ioprio cgroup: CONFIG_BLK_CGROUP_IOPRIO option is disabled in my case (ARM64 defconfig), so I don't think it has to do with ioprio cgroup?

[1] https://lore.kernel.org/all/CAP-bSRbCo7=wfUBZ8H7c3Q-7XSG+SB=R4MHHNNGPvBoinsVSZg@mail.gmail.com/

Thanks!

Quoted 3 lines

Thanks,

Bart.

#13 lore raw

Re: [PATCH 2/2] block: remove the ioprio field from struct request

In reply to Sam Protsenko
Quoted 10 lines · Sam Protsenko

It's an Exynos based board with eMMC, so it uses DW MMC driver, with Exynos glue layer on top of it, so:

drivers/mmc/host/dw_mmc.c
drivers/mmc/host/dw_mmc-exynos.c

I'm using the regular ARM64 defconfig. Nothing fancy about this setup neither, the device tree with eMMC definition (mmc_0) is here:

    arch/arm64/boot/dts/exynos/exynos850-e850-96.dts

Thanks. eMMC itself never looks at the ioprio field.

Quoted 2 lines

FWIW, I was able to narrow down the issue to dd_insert_request() function. With this hack the freeze is gone:

Sounds like it isn't the driver that matters here, but the scheduler.

Quoted 16 lines

8<-------------------------------------------------------------------->8

block/mq-deadline.c +1 -1
index acdc28756d9d..83d272b66e71 100644
@@ -676,7 +676,7 @@ static void dd_insert_request(struct blk_mq_hw_ctx
 *hctx, struct request *rq,
        struct request_queue *q = hctx->queue;
        struct deadline_data *dd = q->elevator->elevator_data;
        const enum dd_data_dir data_dir = rq_data_dir(rq);
-       u16 ioprio = req_get_ioprio(rq);
+       u16 ioprio = 0; /* the same as old req->ioprio */
        u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
        struct dd_per_prio *per_prio;
        enum dd_prio prio;
Quoted 3 lines

8<-------------------------------------------------------------------->8

Does it tell you anything about where the possible issue can be?

Can you dump the ioprities you see here with and without the reverted patch?

#14 lore raw

Re: [PATCH 2/2] block: remove the ioprio field from struct request

Hi Christoph,

Quoted 38 lines · Christoph Hellwig
Quoted 10 lines · Sam Protsenko

It's an Exynos based board with eMMC, so it uses DW MMC driver, with Exynos glue layer on top of it, so:

drivers/mmc/host/dw_mmc.c
drivers/mmc/host/dw_mmc-exynos.c

I'm using the regular ARM64 defconfig. Nothing fancy about this setup neither, the device tree with eMMC definition (mmc_0) is here:

    arch/arm64/boot/dts/exynos/exynos850-e850-96.dts

Thanks. eMMC itself never looks at the ioprio field.

Quoted 2 lines

FWIW, I was able to narrow down the issue to dd_insert_request() function. With this hack the freeze is gone:

Sounds like it isn't the driver that matters here, but the scheduler.

Quoted 18 lines

8<-------------------------------------------------------------------->8

block/mq-deadline.c +1 -1
index acdc28756d9d..83d272b66e71 100644
@@ -676,7 +676,7 @@ static void dd_insert_request(struct blk_mq_hw_ctx
 *hctx, struct request *rq,
        struct request_queue *q = hctx->queue;
        struct deadline_data *dd = q->elevator->elevator_data;
        const enum dd_data_dir data_dir = rq_data_dir(rq);
-       u16 ioprio = req_get_ioprio(rq);
+       u16 ioprio = 0; /* the same as old req->ioprio */
        u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
        struct dd_per_prio *per_prio;
        enum dd_prio prio;

8<-------------------------------------------------------------------->8

Quoted 5 lines
Quoted 1 line

Does it tell you anything about where the possible issue can be?

Can you dump the ioprities you see here with and without the reverted patch?

Collected the logs for you:
- with patch reverted (ioprio is always 0): [1]
- with patch present: [2]

This code was added for printing the traces:

8<---------------------------------------------------------------------->8
static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
blk_insert_t flags, struct list_head *free)
{
+#define IOPRIO_N_LIMIT 100

+       static int ioprio_prev = 0, ioprio_n = 0;
        struct request_queue *q = hctx->queue;
        struct deadline_data *dd = q->elevator->elevator_data;
        const enum dd_data_dir data_dir = rq_data_dir(rq);
        u16 ioprio = req_get_ioprio(rq);
        u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
        struct dd_per_prio *per_prio;
        enum dd_prio prio;

+ ioprio_n++;
+ if (ioprio != ioprio_prev || ioprio_n == IOPRIO_N_LIMIT) {
+ pr_err("### %-20d : %d times\n", ioprio_prev, ioprio_n);
+ ioprio_n = 0;
+ }
+ ioprio_prev = ioprio;
+

        lockdep_assert_held(&dd->lock);

8<---------------------------------------------------------------------->8

Specifically I'd pay attention to the next two places in [2], where the delays were introduced:

1. Starting getty service (5 second delay):

8<---------------------------------------------------------------------->8

[   14.875199] ### 24580                : 1 times
...
[  OK  ] Started [email protected] - Getty on tty1.
[  OK  ] Started serial-getty@ttySAâice - Serial Getty on ttySAC0.
[  OK  ] Reached target getty.target - Login Prompts.
[   19.425354] ### 0                    : 100 times

...
8<---------------------------------------------------------------------->8

2. Login (more than 60 seconds delay):

8<---------------------------------------------------------------------->8 runner-vwmj3eza-project-40964107-concurrent-0 login: root
...

[   22.827432] ### 0                    : 100 times
...
[  100.100402] ### 24580                : 1 times

# 8<---------------------------------------------------------------------->8

I guess the results look similar to the logs from the neighboring thread [3]. Not sure if those tools (getty service and login tool) are running ioprio_set() internally, or it's just I/O scheduler acting up, but the freeze is happening consistently in those two places. Please let me know if I can help you debug that further. Not a block layer expert by any means, but that looks like a regression, at least on E850-96 board. Wonder if it's possible to reproduce that on other platforms.

Thanks!

[1] https://termbin.com/aus7
[2] https://termbin.com/za3t
[3] https://lore.kernel.org/all/CAP-bSRab1C-_aaATfrgWjt9w0fcYUCQCG7u+TCb1FSPSd6CEaA@mail.gmail.com/

#15 lore raw

Re: [PATCH 2/2] block: remove the ioprio field from struct request

Hi Sam,

please try the patch below:

block/mq-deadline.c +4 -9
index acdc28756d9d..91b3789f710e 100644
@@ -685,10 +685,9 @@ static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
 
 	prio = ioprio_class_to_prio[ioprio_class];
 	per_prio = &dd->per_prio[prio];
-	if (!rq->elv.priv[0]) {
+	if (!rq->elv.priv[0])
 		per_prio->stats.inserted++;
-		rq->elv.priv[0] = (void *)(uintptr_t)1;
-	}
+	rq->elv.priv[0] = per_prio;
 
 	if (blk_mq_sched_try_insert_merge(q, rq, free))
 		return;
@@ -753,18 +752,14 @@ static void dd_prepare_request(struct request *rq)
  */
 static void dd_finish_request(struct request *rq)
 {
-	struct request_queue *q = rq->q;
-	struct deadline_data *dd = q->elevator->elevator_data;
-	const u8 ioprio_class = dd_rq_ioclass(rq);
-	const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
-	struct dd_per_prio *per_prio = &dd->per_prio[prio];
+	struct dd_per_prio *per_prio = rq->elv.priv[0];
 
 	/*
 	 * The block layer core may call dd_finish_request() without having
 	 * called dd_insert_requests(). Skip requests that bypassed I/O
 	 * scheduling. See also blk_mq_request_bypass_insert().
 	 */
-	if (rq->elv.priv[0])
+	if (per_prio)
 		atomic_inc(&per_prio->stats.completed);
 }
 
#16 lore raw

Re: [PATCH 2/2] block: remove the ioprio field from struct request

Quoted 5 lines · Christoph Hellwig

Hi Sam,

please try the patch below:

I can confirm the patch fixes the issue on my side. Please add me to
Cc: if you're going to send the fix. And I'm always available to run more tests on E850-96 board if you need it.

Thanks!

Quoted 38 lines
block/mq-deadline.c +4 -9
index acdc28756d9d..91b3789f710e 100644
@@ -685,10 +685,9 @@ static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
 
        prio = ioprio_class_to_prio[ioprio_class];
        per_prio = &dd->per_prio[prio];
-       if (!rq->elv.priv[0]) {
+       if (!rq->elv.priv[0])
                per_prio->stats.inserted++;
-               rq->elv.priv[0] = (void *)(uintptr_t)1;
-       }
+       rq->elv.priv[0] = per_prio;
 
        if (blk_mq_sched_try_insert_merge(q, rq, free))
                return;
@@ -753,18 +752,14 @@ static void dd_prepare_request(struct request *rq)
  */
 static void dd_finish_request(struct request *rq)
 {
-       struct request_queue *q = rq->q;
-       struct deadline_data *dd = q->elevator->elevator_data;
-       const u8 ioprio_class = dd_rq_ioclass(rq);
-       const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
-       struct dd_per_prio *per_prio = &dd->per_prio[prio];
+       struct dd_per_prio *per_prio = rq->elv.priv[0];
 
        /*
         * The block layer core may call dd_finish_request() without having
         * called dd_insert_requests(). Skip requests that bypassed I/O
         * scheduling. See also blk_mq_request_bypass_insert().
         */
-       if (rq->elv.priv[0])
+       if (per_prio)
                atomic_inc(&per_prio->stats.completed);
 }
 
maintainer Applied
#17 lore raw
In reply to Christoph Hellwig
Quoted 6 lines · Christoph Hellwig

this series removes two fields from struct request that just duplicate information in the bios. As-is it doesn't actually shrink the structure size but instead just creates a 4 byte hole, but that will probably become useful sooner or later.

Diffstat:

Quoted 8 lines
 block/blk-merge.c            |   26 ++++++++++++++------------
 block/blk-mq.c               |    5 +----
 drivers/scsi/sd.c            |    6 +++---
 include/linux/blk-mq.h       |    8 +++-----
 include/trace/events/block.h |    6 +++---
 5 files changed, 24 insertions(+), 27 deletions(-)

[...]

Applied, thanks!

[1/2] block: remove the write_hint field from struct request

      commit: 61952bb73486fff0f5550bccdf4062d9dd0fb163
[2/2] block: remove the ioprio field from struct request
      commit: 6975c1a486a40446b5bc77a89d9c520f8296fd08

Best regards,

Signature · 1 line
Jens Axboe
Reviewed-by
#18 lore raw
In reply to Christoph Hellwig
Quoted 7 lines · Christoph Hellwig

Hi Jens,

this series removes two fields from struct request that just duplicate information in the bios. As-is it doesn't actually shrink the structure size but instead just creates a 4 byte hole, but that will probably become useful sooner or later.

Quoted 7 lines

Diffstat:

block/blk-merge.c            |   26 ++++++++++++++------------
block/blk-mq.c               |    5 +----
drivers/scsi/sd.c            |    6 +++---
include/linux/blk-mq.h       |    8 +++-----
include/trace/events/block.h |    6 +++---
5 files changed, 24 insertions(+), 27 deletions(-)