Skip to contentSkip to conversation

scsi: scsi_debug: fix zoned write validation

PATCH Superseded Niklas Cassel 3 participants 18 messages Sep 27 – Sep 28 SCSI

v9 of 11 Superseded Latest is v11 →

lore
Share & tools
lore
#0 lore raw

v9 cover scsi: scsi_debug: fix zoned write validation

This series fixes scsi_debug with regards to writes to a ZBC drive, and to WRITE ATOMIC (16). Each patch describes the problem that it fixes.

Two configurations are refused rather than emulated, and those two patches come first, as the rest of the series relies on what they
exclude: a zoned device whose lowest aligned LBA is not zero, which no host writing in units of the reported zone write granularity could ever write, and atomic writes on a zoned device, which no drive supports.

Little of this is reachable in a default configuration. Most of it needs physblk_exp, atomic_wr or logical block provisioning to be set, or an initiator that does not provide a data buffer matching the transfer length of the command.

The series is based on 7.4/scsi-staging rather than 7.3/scsi-fixes (as Damien suggested on v1) to avoid a build failure that would have happened if this series was simply merged with linux-next:

error: too many arguments to function 'mk_sense_buffer'

Tested on a zoned scsi_debug device with 512 byte logical blocks and a 4096 byte physical block, and on a device that is not zoned for the WRITE ATOMIC (16) patches. The notes below each patch describe what was observed before and after it.

Changes since v8:
- Patch 5 reports the residual of WRITE SCATTERED (16) on every exit once the parameter list has been fetched, including an error on a later range, as Sashiko pointed out. Dropped Damien's Reviewed-by, as the patch has changed since he gave it.

Niklas Cassel (11):
scsi: scsi_debug: Refuse a zoned device with a non-zero lowest aligned
LBA
scsi: scsi_debug: Make atomic writes and ZBC emulation mutually
exclusive
scsi: scsi_debug: Take the zone metadata lock before the data lock
scsi: scsi_debug: Evaluate scsi_debug_lbp() only once
scsi: scsi_debug: Report the residual of a write
scsi: scsi_debug: Enforce physical block alignment of zoned writes
scsi: scsi_debug: Do not write a partial physical block to a zoned
device
scsi: scsi_debug: Advance the write pointer over the data written
scsi: scsi_debug: Refuse a short WRITE ATOMIC (16) before writing it
scsi: scsi_debug: Map the region written by WRITE ATOMIC (16)
scsi: scsi_debug: Validate the access parameters of WRITE ATOMIC (16)

 drivers/scsi/scsi_debug.c | 120 +++++++++++++++++++++++++++++++-------
 1 file changed, 99 insertions(+), 21 deletions(-)

base-commit: f09d2c7485b32adb82336d0d748935c8237a649e

Signature · 1 line
2.55.0
#1 lore raw

PATCH v9 1/11 scsi: scsi_debug: Refuse a zoned device with a non-zero lowest aligned LBA

scsi_debug reports the lowest_aligned module parameter in the LOWEST ALIGNED LOGICAL BLOCK ADDRESS field, see SBC-6 r02 (T10/BSR INCITS 587), 5.20.2, but nothing else in the driver acts on it.

On a zoned device a non-zero value puts the start of every zone in the middle of a physical block. ZBC requires a write to a sequential write required zone to end on a physical block boundary, so a host writing in units of the reported zone write granularity could never write such a zone at all.

Refuse the combination rather than emulate a device that cannot be used.

---
Tested by loading the module with each combination:

  zbc=managed lowest_aligned=8    refused, -EINVAL
  zbc=aware   lowest_aligned=8    refused, -EINVAL
  zbc=managed lowest_aligned=0    loads
  lowest_aligned=8                loads, as the device is not zoned
---
 drivers/scsi/scsi_debug.c | 5 +++++
 1 file changed, 5 insertions(+)
drivers/scsi/scsi_debug.c +5 -0
index 6941809dfdb7..c664966e95cb 100644
@@ -8734,6 +8734,11 @@ static int __init scsi_debug_init(void)
 			sdebug_dev_size_mb = DEF_ZBC_DEV_SIZE_MB;
 	}
 
+	if (sdeb_zbc_in_use && sdebug_lowest_aligned) {
+		pr_err("lowest_aligned is not supported by a zoned device\n");
+		return -EINVAL;
+	}
+
 	if (sdebug_dev_size_mb == DEF_DEV_SIZE_PRE_INIT)
 		sdebug_dev_size_mb = DEF_DEV_SIZE_MB;
 	if (sdebug_dev_size_mb < 1)
Signature · 1 line
2.55.0
#2 lore raw

PATCH v9 2/11 scsi: scsi_debug: Make atomic writes and ZBC emulation mutually exclusive

In reply to Niklas Cassel

There are no SMR drives that support atomic writes, so emulating a device with both emulates something a user cannot encounter, and means maintaining the zone handling of a command that no zoned device implements.

Nothing in the standards forbids the combination, and for a host managed drive whose logical block size equals its physical block size there is no problem in it either: the vendor only has to give atomic writes an alignment that makes sense with regards to ZBC. This is a choice not to complicate scsi_debug.

Refuse the combination at initialisation. It can be changed back if such drives ever appear.

---
Tested by loading the module with each combination:

  zbc=managed atomic_wr=1    refused, -EINVAL
  zbc=aware   atomic_wr=1    refused, -EINVAL
  zbc=managed                loads
  atomic_wr=1                loads
---
 drivers/scsi/scsi_debug.c | 5 +++++
 1 file changed, 5 insertions(+)
drivers/scsi/scsi_debug.c +5 -0
index c664966e95cb..8e45a57ae406 100644
@@ -8739,6 +8739,11 @@ static int __init scsi_debug_init(void)
 		return -EINVAL;
 	}
 
+	if (sdeb_zbc_in_use && sdebug_atomic_wr) {
+		pr_err("atomic_wr is not supported by a zoned device\n");
+		return -EINVAL;
+	}
+
 	if (sdebug_dev_size_mb == DEF_DEV_SIZE_PRE_INIT)
 		sdebug_dev_size_mb = DEF_DEV_SIZE_MB;
 	if (sdebug_dev_size_mb < 1)
Signature · 1 line
2.55.0
#3 lore raw

PATCH v9 3/11 scsi: scsi_debug: Take the zone metadata lock before the data lock

In reply to Niklas Cassel

resp_comp_write() takes the data lock and then the zone metadata lock, while every other command that takes both takes the metadata lock first and reaches the data lock through do_device_access(). A COMPARE AND WRITE and any other write to the same store can therefore deadlock each other.

Take the locks in the same order as everywhere else.

Correct the comment above map_region() while here: the provisioning map is covered by the metadata lock rather than the data lock, which is why resp_unmap() takes only the metadata lock.

Assisted-by LLM
Fixes 84f3a3c01d70 ("scsi: scsi_debug: Atomic write support")

---
Tested with:

modprobe scsi_debug sector_size=512 dev_size_mb=128 lbpu=1 lbpws=1

COMPARE AND WRITE completes, and a READ(16) of the block that it wrote returns, so the reordered locks are still taken and released correctly.

The deadlock itself was not reproduced. This kernel is built without lockdep, and the window needs a COMPARE AND WRITE and another write to the same store to interleave between the two acquisitions. Found by review.

---
 drivers/scsi/scsi_debug.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/scsi/scsi_debug.c +3 -3
index 8e45a57ae406..18aefe83b7b6 100644
@@ -5575,8 +5575,8 @@ static int resp_comp_write(struct scsi_cmnd *scp,
 			    "indicated=%u, IO sent=%d bytes\n", my_name,
 			    dnum * lb_size, ret);
 
-	sdeb_data_write_lock(sip);
 	sdeb_meta_write_lock(sip);
+	sdeb_data_write_lock(sip);
 	if (!comp_write_worker(sip, lba, num, arr, false)) {
 		mk_sense_buffer(scp, MISCOMPARE,
 				MISCOMPARE_DURING_VERIFY_OPERATION);
@@ -5584,12 +5584,12 @@ static int resp_comp_write(struct scsi_cmnd *scp,
 		goto cleanup_unlock;
 	}
 
-	/* Cover sip->map_storep (which map_region()) sets with data lock */
+	/* Cover sip->map_storep (which map_region() sets) with the meta lock */
 	if (scsi_debug_lbp())
 		map_region(sip, lba, num);
 cleanup_unlock:
-	sdeb_meta_write_unlock(sip);
 	sdeb_data_write_unlock(sip);
+	sdeb_meta_write_unlock(sip);
 cleanup_free:
 	kfree(arr);
 	return retval;
Signature · 1 line
2.55.0
#4 lore raw

PATCH v9 4/11 scsi: scsi_debug: Evaluate scsi_debug_lbp() only once

In reply to Niklas Cassel

corrupt_lbas(), resp_write_dt0() and resp_write_same() call scsi_debug_lbp() to decide whether to take the zone metadata lock, and call it again to decide whether to touch the provisioning map that the lock protects.

The result is not constant: scsi_debug_lbp() is false while fake_rw is set, and fake_rw is writable at runtime, both as a module parameter and through its driver attribute. The second call can therefore touch the map after the first decided not to take the lock.

Call it once and use the result throughout.

---
Tested with:

modprobe scsi_debug sector_size=512 dev_size_mb=128 lbpu=1 lbpws=1

A WRITE(16) completes and GET LBA STATUS then reports the region as mapped, which covers resp_write_dt0(). A WRITE SAME and a WRITE SAME with the UNMAP bit set complete, and GET LBA STATUS reports the first region as mapped and the second as deallocated, which covers all three uses of the result in resp_write_same().

corrupt_lbas() is not covered. It is reached by writing to the corrupt file in debugfs, and that interface rejected the requests that were tried, for a reason that has nothing to do with this patch.

The window that the change closes was not reproduced. It needs fake_rw to be written between two calls, and was found by review.

---
 drivers/scsi/scsi_debug.c | 21 ++++++++++-----------
 1 file changed, 10 insertions(+), 11 deletions(-)
drivers/scsi/scsi_debug.c +10 -11
index 18aefe83b7b6..b64ae3ad300d 100644
@@ -4953,12 +4953,11 @@ static int corrupt_lbas(struct sdebug_dev_info *devip, u64 lba, u32 num,
 {
 	struct sdeb_store_info *sip = devip2sip(devip, false);
 	bool meta_data_locked = false;
+	bool lbp = scsi_debug_lbp();
 	u32 block, num_mapped, b, i;
 	int error = 0;
 
-	if (sdebug_dev_is_zoned(devip) ||
-	    sdebug_dix ||
-	    scsi_debug_lbp())  {
+	if (sdebug_dev_is_zoned(devip) || sdebug_dix || lbp) {
 		sdeb_meta_write_lock(sip);
 		meta_data_locked = true;
 	}
@@ -4975,8 +4974,7 @@ static int corrupt_lbas(struct sdebug_dev_info *devip, u64 lba, u32 num,
 		goto out_unlock;
 	}
 
-	if (scsi_debug_lbp() &&
-	    (!map_state(sip, lba, &num_mapped) || num > num_mapped)) {
+	if (lbp && (!map_state(sip, lba, &num_mapped) || num > num_mapped)) {
 		pr_err("can't modify unmapped logical blocks: %llu:%u",
 			lba, num);
 		error = -EINVAL;
@@ -5035,6 +5033,7 @@ static int resp_write_dt0(struct scsi_cmnd *scp, struct sdebug_dev_info *devip)
 	struct sdeb_store_info *sip = devip2sip(devip, true);
 	u8 *cmd = scp->cmnd;
 	bool meta_data_locked = false;
+	bool lbp = scsi_debug_lbp();
 
 	if (unlikely(sdebug_opts & SDEBUG_OPT_UNALIGNED_WRITE &&
 		     atomic_read(&sdeb_inject_pending))) {
@@ -5101,8 +5100,7 @@ static int resp_write_dt0(struct scsi_cmnd *scp, struct sdebug_dev_info *devip)
 	}
 
 	if (sdebug_dev_is_zoned(devip) ||
-	    (sdebug_dix && scsi_prot_sg_count(scp)) ||
-	    scsi_debug_lbp())  {
+	    (sdebug_dix && scsi_prot_sg_count(scp)) || lbp) {
 		sdeb_meta_write_lock(sip);
 		meta_data_locked = true;
 	}
@@ -5147,7 +5145,7 @@ static int resp_write_dt0(struct scsi_cmnd *scp, struct sdebug_dev_info *devip)
 	}
 
 	ret = do_device_access(sip, scp, 0, lba, num, group, true, false);
-	if (unlikely(scsi_debug_lbp()))
+	if (unlikely(lbp))
 		map_region(sip, lba, num);
 
 	/* If ZBC zone then bump its write pointer */
@@ -5374,8 +5372,9 @@ static int resp_write_same(struct scsi_cmnd *scp, u64 lba, u32 num,
 	u8 *fs1p;
 	u8 *fsp;
 	bool meta_data_locked = false;
+	bool lbp = scsi_debug_lbp();
 
-	if (sdebug_dev_is_zoned(devip) || scsi_debug_lbp()) {
+	if (sdebug_dev_is_zoned(devip) || lbp) {
 		sdeb_meta_write_lock(sip);
 		meta_data_locked = true;
 	}
@@ -5384,7 +5383,7 @@ static int resp_write_same(struct scsi_cmnd *scp, u64 lba, u32 num,
 	if (ret)
 		goto out;
 
-	if (unmap && scsi_debug_lbp()) {
+	if (unmap && lbp) {
 		unmap_region(sip, lba, num);
 		goto out;
 	}
@@ -5414,7 +5413,7 @@ static int resp_write_same(struct scsi_cmnd *scp, u64 lba, u32 num,
 		block = do_div(lbaa, sdebug_store_sectors);
 		memmove(fsp + (block * lb_size), fs1p, lb_size);
 	}
-	if (scsi_debug_lbp())
+	if (lbp)
 		map_region(sip, lba, num);
 	/* If ZBC zone then bump its write pointer */
 	if (sdebug_dev_is_zoned(devip))
Signature · 1 line
2.55.0
#5 lore raw

PATCH v9 5/11 scsi: scsi_debug: Report the residual of a write

In reply to Niklas Cassel

A write whose data buffer is larger than its transfer length leaves a residual that the initiator is entitled to be told about. resp_read_dt0() and resp_write_tape() report it, but the write paths for a disk do not, so such a write completes with GOOD status and a residual of zero while a READ of the same length into the same buffer reports it correctly.

Report it in resp_write_dt0(), resp_write_scat() and resp_atomic_write().

resp_write_scat() differs twice over. Its data-out buffer also holds the parameter list header and the LBA range descriptors, so what the command consumed is sg_off rather than the bytes that the last range transferred; and sg_off counts what was asked for rather than what was transferred, so it can exceed the buffer and needs a guard. A command with no LBA range descriptors or a buffer transfer length of zero transfers nothing, so the whole buffer is residual. An error, real or injected, can end the command after some ranges have been written, so the residual is reported on every exit once the parameter list has been fetched.

Assisted-by LLM

---
Tested with:

  modprobe scsi_debug zbc=managed sector_size=512 physblk_exp=3 \
      zone_size_mb=8 dev_size_mb=128 zone_nr_conv=2

An eight logical block WRITE(16) with a 5000 byte buffer reports resid=904, having transferred 4096. A READ(16) of the same length into the same buffer reported that before this patch and the WRITE reported
0. A buffer of exactly 4096 bytes reports 0, as does a 512 byte buffer, which is consumed in its entirety. WRITE ATOMIC (16) behaves like the WRITE, on a device that is not zoned.

WRITE SCATTERED (16) with one LBA range descriptor of eight blocks and a 5632 byte buffer reports resid=1024, having consumed 512 bytes of parameter list and 4096 bytes of data. The same command with a 1024 byte buffer completes and reports resid=0: sg_off reaches 4608, past the end of the buffer, and without the guard the residual would have been reported as 4294963712.

Changes since v8: the residual is also reported when a later range fails. A WRITE SCATTERED (16) whose second range is out of range now reports resid=5120 of its 9728 byte buffer, where v8 reported 0.

---
 drivers/scsi/scsi_debug.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)
drivers/scsi/scsi_debug.c +16 -1
index b64ae3ad300d..0c66a3118ee8 100644
@@ -5162,6 +5162,8 @@ static int resp_write_dt0(struct scsi_cmnd *scp, struct sdebug_dev_info *devip)
 			    "%s: write: cdb indicated=%u, IO sent=%d bytes\n",
 			    my_name, num * sdebug_sector_size, ret);
 
+	scsi_set_resid(scp, scsi_bufflen(scp) - ret);
+
 	if (unlikely((sdebug_opts & SDEBUG_OPT_RECOV_DIF_DIX) &&
 		     atomic_read(&sdeb_inject_pending))) {
 		if (sdebug_opts & SDEBUG_OPT_RECOVERED_ERR) {
@@ -5233,8 +5235,10 @@ static int resp_write_scat(struct scsi_cmnd *scp,
 					    "Unprotected WR to DIF device\n");
 		}
 	}
-	if ((num_lrd == 0) || (bt_len == 0))
+	if (num_lrd == 0 || bt_len == 0) {
+		scsi_set_resid(scp, scsi_bufflen(scp));
 		return 0;       /* T10 says these do-nothings are not errors */
+	}
 	if (lbdof == 0) {
 		if (sdebug_verbose)
 			sdev_printk(KERN_INFO, scp->device,
@@ -5327,6 +5331,8 @@ static int resp_write_scat(struct scsi_cmnd *scp,
 
 		if (unlikely((sdebug_opts & SDEBUG_OPT_RECOV_DIF_DIX) &&
 			     atomic_read(&sdeb_inject_pending))) {
+			/* This range has been written */
+			sg_off += num_by;
 			if (sdebug_opts & SDEBUG_OPT_RECOVERED_ERR) {
 				mk_sense_buffer(scp, RECOVERED_ERROR,
 					FAILURE_PREDICTION_THRESHOLD_EXCEEDED);
@@ -5352,6 +5358,13 @@ static int resp_write_scat(struct scsi_cmnd *scp,
 	}
 	ret = 0;
 err_out_unlock:
+	/*
+	 * sg_off counts what the command asked for, which can exceed the
+	 * buffer: lbdof is not validated against it, and a range is counted
+	 * in full even if do_device_access() copied less.
+	 */
+	if (scsi_bufflen(scp) > sg_off)
+		scsi_set_resid(scp, scsi_bufflen(scp) - sg_off);
 	sdeb_meta_write_unlock(sip);
 err_out:
 	kfree(lrdp);
@@ -6206,6 +6219,8 @@ static int resp_atomic_write(struct scsi_cmnd *scp,
 		return DID_ERROR << 16;
 	if (unlikely(ret != len * sdebug_sector_size))
 		return DID_ERROR << 16;
+
+	scsi_set_resid(scp, scsi_bufflen(scp) - ret);
 	return 0;
 }
 
Signature · 1 line
2.55.0
#6 lore raw

Re: [PATCH v9 05/11] scsi: scsi_debug: Report the residual of a write

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Medium] scsi: scsi_debug: incorrect residual for WRITE SCATTERED due to 32-bit overflow --

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] scsi: scsi_debug: incorrect residual for WRITE SCATTERED due to 32-bit overflow --

commit 8bc6d8e408815c5b6d6afe8b3e8dbdaabb79d3d5
Author: Niklas Cassel <[email protected]>
scsi: scsi_debug: Report the residual of a write

This commit adds residual reporting for writes in scsi_debug, ensuring that writes with data buffers larger than their transfer length report the residual correctly, similar to reads. It updates resp_write_dt0(), resp_write_scat(), and resp_atomic_write() to calculate and set the residual based on what the command consumed.

Quoted 4 lines
drivers/scsi/scsi_debug.c +0 -0
index b64ae3ad300d8..0c66a3118ee89 100644

[ ... ]

drivers/scsi/scsi_debug.c static int resp_write_scat(struct scsi_cmnd *scp, · line 5367

Quoted 11 lines
drivers/scsi/scsi_debug.c +7 -0
@@ -5352,6 +5358,13 @@ static int resp_write_scat(struct scsi_cmnd *scp,
 	}
 	ret = 0;
 err_out_unlock:
+	/*
+	 * sg_off counts what the command asked for, which can exceed the
+	 * buffer: lbdof is not validated against it, and a range is counted
+	 * in full even if do_device_access() copied less.
+	 */
+	if (scsi_bufflen(scp) > sg_off)
+		scsi_set_resid(scp, scsi_bufflen(scp) - sg_off);

[Severity: Medium]
Will this calculation produce an incorrect residual if the command includes a very large LBA range?

In resp_write_scat(), the variables num_by, num, and lb_size are all 32-bit integers, and num_by is calculated as:

    num_by = num * lb_size;

If num is sufficiently large (e.g., > 8.38M blocks), does this 32-bit multiplication overflow and wrap around to a small value?

If num_by wraps, it appears the sg_off accumulator will also wrap below the actual buffer length, causing this new guard if (scsi_bufflen(scp) > sg_off) to erroneously evaluate to true. Could this result in a bogus residual being set for a write that actually fully consumed the buffer, potentially confusing the initiator or midlayer into failing or retrying completed I/Os?

Quoted 3 lines
drivers/scsi/scsi_debug.c +0 -0
 	sdeb_meta_write_unlock(sip);
 err_out:
 	kfree(lrdp);
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=5
#7 lore raw

Re: [PATCH v9 05/11] scsi: scsi_debug: Report the residual of a write

Quoted 11 lines · On Sun, Sep 27, 2026 at 05:40:39AM +0000, [email protected] wrote:

[Severity: Medium]
Will this calculation produce an incorrect residual if the command includes a very large LBA range?

In resp_write_scat(), the variables num_by, num, and lb_size are all 32-bit integers, and num_by is calculated as:

    num_by = num * lb_size;

If num is sufficiently large (e.g., > 8.38M blocks), does this 32-bit multiplication overflow and wrap around to a small value?

num_by and sg_off have been u32 since commit 481b5e5c7949 ("scsi: scsi_debug:
add resp_write_scat function").

sg_off (which can already overflow in mainline) is already passed to do_device_access(), so this is not a new problem...

But sure... I can introduce a prep patch which changes both to u64.

Kind regards,
Niklas

#8 lore raw

Re: [PATCH v9 05/11] scsi: scsi_debug: Report the residual of a write

In reply to Niklas Cassel
Quoted 5 lines · Niklas Cassel

A write whose data buffer is larger than its transfer length leaves a residual that the initiator is entitled to be told about. resp_read_dt0() and resp_write_tape() report it, but the write paths for a disk do not, so such a write completes with GOOD status and a residual of zero while a READ of the same length into the same buffer reports it correctly.

I am not convinced this is correct. resid (residual count) is supposed to indicate the amount of data that was *not* transferred. So if the buffer size is larger that the command cdb transfer count, all data can be transferred and resid should be 0. Which would mean that for this case, the write processing is correct but the read side is not.

The SBC and SPC specs are very silent about this though, I do not think that the resid is actually standardized. And I do not see a clear definition for it in the documentation/code comments.

Quoted 100 lines

Report it in resp_write_dt0(), resp_write_scat() and resp_atomic_write().

resp_write_scat() differs twice over. Its data-out buffer also holds the parameter list header and the LBA range descriptors, so what the command consumed is sg_off rather than the bytes that the last range transferred; and sg_off counts what was asked for rather than what was transferred, so it can exceed the buffer and needs a guard. A command with no LBA range descriptors or a buffer transfer length of zero transfers nothing, so the whole buffer is residual. An error, real or injected, can end the command after some ranges have been written, so the residual is reported on every exit once the parameter list has been fetched.

Assisted-by LLM

---
Tested with:

  modprobe scsi_debug zbc=managed sector_size=512 physblk_exp=3 \
      zone_size_mb=8 dev_size_mb=128 zone_nr_conv=2

An eight logical block WRITE(16) with a 5000 byte buffer reports resid=904, having transferred 4096. A READ(16) of the same length into the same buffer reported that before this patch and the WRITE reported
0. A buffer of exactly 4096 bytes reports 0, as does a 512 byte buffer, which is consumed in its entirety. WRITE ATOMIC (16) behaves like the WRITE, on a device that is not zoned.

WRITE SCATTERED (16) with one LBA range descriptor of eight blocks and a 5632 byte buffer reports resid=1024, having consumed 512 bytes of parameter list and 4096 bytes of data. The same command with a 1024 byte buffer completes and reports resid=0: sg_off reaches 4608, past the end of the buffer, and without the guard the residual would have been reported as 4294963712.

Changes since v8: the residual is also reported when a later range fails. A WRITE SCATTERED (16) whose second range is out of range now reports resid=5120 of its 9728 byte buffer, where v8 reported 0.

---
 drivers/scsi/scsi_debug.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)
drivers/scsi/scsi_debug.c +16 -1
index b64ae3ad300d..0c66a3118ee8 100644
@@ -5162,6 +5162,8 @@ static int resp_write_dt0(struct scsi_cmnd *scp, struct sdebug_dev_info *devip)
 			    "%s: write: cdb indicated=%u, IO sent=%d bytes\n",
 			    my_name, num * sdebug_sector_size, ret);
 
+	scsi_set_resid(scp, scsi_bufflen(scp) - ret);
+
 	if (unlikely((sdebug_opts & SDEBUG_OPT_RECOV_DIF_DIX) &&
 		     atomic_read(&sdeb_inject_pending))) {
 		if (sdebug_opts & SDEBUG_OPT_RECOVERED_ERR) {
@@ -5233,8 +5235,10 @@ static int resp_write_scat(struct scsi_cmnd *scp,
 					    "Unprotected WR to DIF device\n");
 		}
 	}
-	if ((num_lrd == 0) || (bt_len == 0))
+	if (num_lrd == 0 || bt_len == 0) {
+		scsi_set_resid(scp, scsi_bufflen(scp));
 		return 0;       /* T10 says these do-nothings are not errors */
+	}
 	if (lbdof == 0) {
 		if (sdebug_verbose)
 			sdev_printk(KERN_INFO, scp->device,
@@ -5327,6 +5331,8 @@ static int resp_write_scat(struct scsi_cmnd *scp,
 
 		if (unlikely((sdebug_opts & SDEBUG_OPT_RECOV_DIF_DIX) &&
 			     atomic_read(&sdeb_inject_pending))) {
+			/* This range has been written */
+			sg_off += num_by;
 			if (sdebug_opts & SDEBUG_OPT_RECOVERED_ERR) {
 				mk_sense_buffer(scp, RECOVERED_ERROR,
 					FAILURE_PREDICTION_THRESHOLD_EXCEEDED);
@@ -5352,6 +5358,13 @@ static int resp_write_scat(struct scsi_cmnd *scp,
 	}
 	ret = 0;
 err_out_unlock:
+	/*
+	 * sg_off counts what the command asked for, which can exceed the
+	 * buffer: lbdof is not validated against it, and a range is counted
+	 * in full even if do_device_access() copied less.
+	 */
+	if (scsi_bufflen(scp) > sg_off)
+		scsi_set_resid(scp, scsi_bufflen(scp) - sg_off);
 	sdeb_meta_write_unlock(sip);
 err_out:
 	kfree(lrdp);
@@ -6206,6 +6219,8 @@ static int resp_atomic_write(struct scsi_cmnd *scp,
 		return DID_ERROR << 16;
 	if (unlikely(ret != len * sdebug_sector_size))
 		return DID_ERROR << 16;
+
+	scsi_set_resid(scp, scsi_bufflen(scp) - ret);
 	return 0;
 }
 
Signature · 2 lines
Damien Le Moal
Western Digital Research
#9 lore raw

Re: [PATCH v9 05/11] scsi: scsi_debug: Report the residual of a write

Quoted 13 lines · Damien Le Moal
Quoted 5 lines · Niklas Cassel

A write whose data buffer is larger than its transfer length leaves a residual that the initiator is entitled to be told about. resp_read_dt0() and resp_write_tape() report it, but the write paths for a disk do not, so such a write completes with GOOD status and a residual of zero while a READ of the same length into the same buffer reports it correctly.

I am not convinced this is correct. resid (residual count) is supposed to indicate the amount of data that was *not* transferred. So if the buffer size is larger that the command cdb transfer count, all data can be transferred and resid should be 0. Which would mean that for this case, the write processing is correct but the read side is not.

Quoted 3 lines

The SBC and SPC specs are very silent about this though, I do not think that the resid is actually standardized. And I do not see a clear definition for it in the documentation/code comments.

The definition is in Documentation/scsi/scsi_mid_low_api.rst, under resid_len: "an LLD should set this unsigned integer to the requested transfer length (i.e. 'request_bufflen') less the number of bytes that are actually transferred".

Also in include/scsi/sg.h: "resid; /* [o] dxfer_len - actual_transferred */".

Both measure it against the buffer, not the CDB.

That is also what iSCSI reports: RFC 7143 11.4.5.1 measures the residual against the Expected Data Transfer Length, which the initiator sets from the buffer.

And it is what scsi_debug has always done for reads, in fill_from_dev_buffer(). This patch just does the same for writes.

Let me clarify the commit message in v10 to actually reference Documentation/scsi/scsi_mid_low_api.rst.

Kind regards,
Niklas

#10 lore raw

Re: [PATCH v9 05/11] scsi: scsi_debug: Report the residual of a write

Quoted 25 lines · Niklas Cassel
Quoted 13 lines · Damien Le Moal
Quoted 5 lines · Niklas Cassel

A write whose data buffer is larger than its transfer length leaves a residual that the initiator is entitled to be told about. resp_read_dt0() and resp_write_tape() report it, but the write paths for a disk do not, so such a write completes with GOOD status and a residual of zero while a READ of the same length into the same buffer reports it correctly.

I am not convinced this is correct. resid (residual count) is supposed to indicate the amount of data that was *not* transferred. So if the buffer size is larger that the command cdb transfer count, all data can be transferred and resid should be 0. Which would mean that for this case, the write processing is correct but the read side is not.

Quoted 3 lines

The SBC and SPC specs are very silent about this though, I do not think that the resid is actually standardized. And I do not see a clear definition for it in the documentation/code comments.

The definition is in Documentation/scsi/scsi_mid_low_api.rst, under resid_len: "an LLD should set this unsigned integer to the requested transfer length (i.e. 'request_bufflen') less the number of bytes that are actually transferred".

Also in include/scsi/sg.h: "resid; /* [o] dxfer_len - actual_transferred */".

Quoted 8 lines

Both measure it against the buffer, not the CDB.

That is also what iSCSI reports: RFC 7143 11.4.5.1 measures the residual against the Expected Data Transfer Length, which the initiator sets from the buffer.

And it is what scsi_debug has always done for reads, in fill_from_dev_buffer(). This patch just does the same for writes.

Hmm. OK. But then that really should be set automatically by the sd/st/sr drivers and sg driver. scsi_debug emulates a device, which does not give the residual. So scsi_debug has no business setting that command residual at all, for any command.

Quoted 8 lines

Let me clarify the commit message in v10 to actually reference Documentation/scsi/scsi_mid_low_api.rst.

Kind regards,
Niklas

Signature · 2 lines
Damien Le Moal
Western Digital Research
#11 lore raw

Re: [PATCH v9 05/11] scsi: scsi_debug: Report the residual of a write

Quoted 32 lines · Damien Le Moal
Quoted 31 lines · Niklas Cassel
Quoted 13 lines · Damien Le Moal
Quoted 5 lines · Niklas Cassel

A write whose data buffer is larger than its transfer length leaves a residual that the initiator is entitled to be told about. resp_read_dt0() and resp_write_tape() report it, but the write paths for a disk do not, so such a write completes with GOOD status and a residual of zero while a READ of the same length into the same buffer reports it correctly.

I am not convinced this is correct. resid (residual count) is supposed to indicate the amount of data that was *not* transferred. So if the buffer size is larger that the command cdb transfer count, all data can be transferred and resid should be 0. Which would mean that for this case, the write processing is correct but the read side is not.

Quoted 3 lines

The SBC and SPC specs are very silent about this though, I do not think that the resid is actually standardized. And I do not see a clear definition for it in the documentation/code comments.

The definition is in Documentation/scsi/scsi_mid_low_api.rst, under resid_len: "an LLD should set this unsigned integer to the requested transfer length (i.e. 'request_bufflen') less the number of bytes that are actually transferred".

Also in include/scsi/sg.h: "resid; /* [o] dxfer_len - actual_transferred */".

Both measure it against the buffer, not the CDB.

That is also what iSCSI reports: RFC 7143 11.4.5.1 measures the residual against the Expected Data Transfer Length, which the initiator sets from the buffer.

Quoted 7 lines
Quoted 2 lines

And it is what scsi_debug has always done for reads, in fill_from_dev_buffer(). This patch just does the same for writes.

Hmm. OK. But then that really should be set automatically by the sd/st/sr drivers and sg driver. scsi_debug emulates a device, which does not give the residual. So scsi_debug has no business setting that command residual at all, for any command.

Please ignore this rambling. scsi_debug emulates a scsi host (i.e. an adapter) with logical units (devices) attached to it. And it is perfectly fine for a host driver to set the residual. Your patch is fine and I think I need more coffee :)

Signature · 2 lines
Damien Le Moal
Western Digital Research
#12 lore raw

PATCH v9 6/11 scsi: scsi_debug: Enforce physical block alignment of zoned writes

In reply to Niklas Cassel

ZBC-3 r06 (T10/BSR INCITS 579), 4.5.3.3.2, requires a write to a sequential write required zone to be terminated with ILLEGAL REQUEST / UNALIGNED WRITE COMMAND both when it does not start at the write pointer and when it does not end on a physical block boundary. That is why sd_zbc_read_zones() sets the zone_write_granularity queue limit to the physical block size of a host-managed device.

check_zbc_access_params() implements the first condition but not the second. So with zbc=managed sector_size=512 physblk_exp=3, a write of a single logical block at the write pointer is accepted and advances the write pointer by one logical block, leaving it off the granularity that was advertised: nothing can write at it any more, and the zone can only be used again after being reset.

Check the ending LBA as well, with the same sense data, as the standard gives both conditions the same. An entire medium write same command, which the standard excludes, needs no special case: the preceding check already terminates it with WRITE BOUNDARY VIOLATION. Sequential write preferred zones are left alone, as writes to them need not be sequential, and with the default physblk_exp=0 the check is a no-op.

Assisted-by LLM
Fixes f0d1cf9378bd ("scsi: scsi_debug: Add ZBC zone commands")

---
Tested with:

  modprobe scsi_debug zbc=managed sector_size=512 physblk_exp=3 \
      zone_size_mb=8 dev_size_mb=128 zone_nr_conv=2

Before this patch, a one logical block WRITE(16) at the write pointer of an empty sequential write required zone completes with GOOD status and leaves the write pointer at LBA 0x18001, which is not a multiple of the 4096 byte zone_write_granularity reported for the device. After it, the same command is terminated with ILLEGAL REQUEST / UNALIGNED WRITE COMMAND and the zone is left EMPTY, while an aligned eight block write is still accepted.

---
 drivers/scsi/scsi_debug.c | 10 ++++++++++
 1 file changed, 10 insertions(+)
drivers/scsi/scsi_debug.c +10 -0
index 0c66a3118ee8..82f93c63923c 100644
@@ -3980,6 +3980,16 @@ static int check_zbc_access_params(struct scsi_cmnd *scp,
 					UNALIGNED_WRITE_COMMAND);
 			return check_condition_result;
 		}
+		/*
+		 * Writes must end on a physical block boundary, that is, the
+		 * transfer length must be a multiple of the physical block
+		 * size.
+		 */
+		if (!IS_ALIGNED(lba + num, 1U << sdebug_physblk_exp)) {
+			mk_sense_buffer(scp, ILLEGAL_REQUEST,
+					UNALIGNED_WRITE_COMMAND);
+			return check_condition_result;
+		}
 	}
 
 	/* Handle implicit open of closed and empty zones */
Signature · 1 line
2.55.0
#13 lore raw

PATCH v9 7/11 scsi: scsi_debug: Do not write a partial physical block to a zoned device

In reply to Niklas Cassel

do_device_access() copies one logical block at a time and stops at the first short copy, so a data-out buffer smaller than the transfer length of the command can leave part of a physical block written. An initiator can arrange that with SG_IO.

That is acceptable everywhere but a sequential write required zone, where ZBC-3 r06 (T10/BSR INCITS 579), 4.5.3.3.2, requires a write to end on a physical block boundary.

For such a zone, stop at the last whole physical block that the buffer holds. The bytes left over are not written, and are reported to the initiator as part of the residual.

---
Tested with:

  modprobe scsi_debug zbc=managed sector_size=512 physblk_exp=3 \
      zone_size_mb=8 dev_size_mb=128 zone_nr_conv=2

issuing a sixteen logical block WRITE(16) through SG_IO with a buffer of twelve blocks. To a sequential write required zone it now transfers eight blocks, one whole physical block, and reports the remaining 2048 bytes as the residual; before this patch it transferred twelve. To a conventional zone it still transfers all twelve and reports no residual.

With physblk_exp=0 an eight block write with a buffer of one block still transfers one block, wherever it is addressed.

---
 drivers/scsi/scsi_debug.c | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)
drivers/scsi/scsi_debug.c +21 -0
index 82f93c63923c..904a3e5952e4 100644
@@ -4273,6 +4273,8 @@ static int do_device_access(struct sdeb_store_info *sip, struct scsi_cmnd *scp,
 	u64 block;
 	enum dma_data_direction dir;
 	struct scsi_data_buffer *sdb = &scp->sdb;
+	struct scsi_device *sdp = scp->device;
+	struct sdebug_dev_info *devip = (struct sdebug_dev_info *)sdp->hostdata;
 	u8 *fsp;
 	int i, total = 0;
 
@@ -4300,6 +4302,25 @@ static int do_device_access(struct sdeb_store_info *sip, struct scsi_cmnd *scp,
 
 	fsp = sip->storep;
 
+	/*
+	 * A write to a sequential write required zone has to end on a physical
+	 * block boundary, so if the data-out buffer does not hold all of the
+	 * data that the command asks for, write up to the last whole physical
+	 * block that it does hold. What is left over is reported as part of
+	 * the residual.
+	 */
+	if (do_write && sdebug_dev_is_zoned(devip)) {
+		struct sdeb_zone_state *zsp = zbc_zone(devip, lba);
+
+		if (zsp->z_type == ZBC_ZTYPE_SWR) {
+			u32 avail = (sdb->length - sg_skip)
+				>> ilog2(sdebug_sector_size);
+
+			if (avail < num)
+				num = round_down(avail, 1U << sdebug_physblk_exp);
+		}
+	}
+
 	block = do_div(lba, sdebug_store_sectors);
 
 	/* Only allow 1x atomic write or multiple non-atomic writes at any given time */
Signature · 1 line
2.55.0
#14 lore raw

PATCH v9 8/11 scsi: scsi_debug: Advance the write pointer over the data written

In reply to Niklas Cassel

resp_write_dt0() and resp_write_scat() advance the write pointer of a sequential write required zone by the transfer length of the command without looking at what do_device_access() returned, which is -1 when the data direction of the command does not match the operation, and a short byte count when the data-out buffer is smaller than the transfer length. An initiator can produce both with SG_IO.

The zone then describes more data than is on the medium, and a write at the position where the data really ends is terminated with UNALIGNED WRITE COMMAND, so the zone has to be reset before it can be written to again.

Advance the write pointer over the data that was written instead. As do_device_access() does not write a partial physical block to such a zone, that leaves the write pointer where a write can end.

Assisted-by LLM
Fixes f0d1cf9378bd ("scsi: scsi_debug: Add ZBC zone commands")

---
Tested with:

  modprobe scsi_debug zbc=managed sector_size=512 physblk_exp=3 \
      zone_size_mb=8 dev_size_mb=128 zone_nr_conv=2

issuing WRITE(16) through SG_IO with a data-out buffer shorter than the transfer length of the command. A sixteen block write with a buffer of twelve blocks leaves the write pointer of an empty sequential write required zone eight blocks on, which is where the data it wrote ends, and a following eight block write continues from there. An eight block write with a buffer of one block writes nothing and leaves the write pointer where it was.

Before this patch both advanced the write pointer by the full transfer length, and a write at the end of the data that had actually been written was then terminated with UNALIGNED WRITE COMMAND.

---
 drivers/scsi/scsi_debug.c | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)
drivers/scsi/scsi_debug.c +7 -6
index 904a3e5952e4..d18f0d71aab6 100644
@@ -5179,9 +5179,9 @@ static int resp_write_dt0(struct scsi_cmnd *scp, struct sdebug_dev_info *devip)
 	if (unlikely(lbp))
 		map_region(sip, lba, num);
 
-	/* If ZBC zone then bump its write pointer */
-	if (sdebug_dev_is_zoned(devip))
-		zbc_inc_wp(devip, lba, num);
+	/* If ZBC zone then bump its write pointer over the data written */
+	if (sdebug_dev_is_zoned(devip) && ret > 0)
+		zbc_inc_wp(devip, lba, ret >> ilog2(sdebug_sector_size));
 	if (meta_data_locked)
 		sdeb_meta_write_unlock(sip);
 
@@ -5347,9 +5347,10 @@ static int resp_write_scat(struct scsi_cmnd *scp,
 		 * writes behaviour as possible.
 		 */
 		ret = do_device_access(sip, scp, sg_off, lba, num, group, true, true);
-		/* If ZBC zone then bump its write pointer */
-		if (sdebug_dev_is_zoned(devip))
-			zbc_inc_wp(devip, lba, num);
+		/* If ZBC zone then bump its write pointer over the data written */
+		if (sdebug_dev_is_zoned(devip) && ret > 0)
+			zbc_inc_wp(devip, lba,
+				   ret >> ilog2(sdebug_sector_size));
 		if (unlikely(scsi_debug_lbp()))
 			map_region(sip, lba, num);
 		if (unlikely(-1 == ret)) {
Signature · 1 line
2.55.0
#15 lore raw

PATCH v9 9/11 scsi: scsi_debug: Refuse a short WRITE ATOMIC (16) before writing it

In reply to Niklas Cassel

A write whose data-out buffer is shorter than its transfer length is short. An ordinary write transfers what the buffer holds and reports the rest as a residual, but an atomic write cannot be short: SBC-6 r02 (T10/BSR INCITS 587), 4.28.1, requires each atomic write operation to write either all of its data or none of it, and 4.28.2 requires one that cannot complete to leave the LBAs that it specifies unaltered.

resp_atomic_write() does fail a short WRITE ATOMIC (16) with DID_ERROR, but only after do_device_access() has returned, and do_device_access() copies one logical block at a time, so by then the blocks that the buffer did hold have been written. An eight block WRITE ATOMIC (16) with a buffer of four fails having overwritten the first four.

Check the length of the buffer before writing anything, and fail the command with the same DID_ERROR as before.

Assisted-by LLM
Fixes 84f3a3c01d70 ("scsi: scsi_debug: Atomic write support")

---
Tested with:

  modprobe scsi_debug sector_size=512 physblk_exp=3 dev_size_mb=128 \
      atomic_wr=1 lbpu=1

issuing a WRITE ATOMIC (16) of eight blocks through SG_IO with a buffer of four blocks of 0xaa. It fails with DID_ERROR before and after this patch, but before it the first four blocks read back as 0xaa afterwards, and after it all eight read back as they were. A WRITE ATOMIC (16) with a full buffer writes all eight blocks, as before.

---
 drivers/scsi/scsi_debug.c | 4 ++++
 1 file changed, 4 insertions(+)
drivers/scsi/scsi_debug.c +4 -0
index d18f0d71aab6..1753b7b087b3 100644
@@ -6246,6 +6246,10 @@ static int resp_atomic_write(struct scsi_cmnd *scp,
 		}
 	}
 
+	/* Short atomic writes are not allowed. */
+	if (scsi_bufflen(scp) < len * sdebug_sector_size)
+		return DID_ERROR << 16;
+
 	ret = do_device_access(sip, scp, 0, lba, len, 0, true, true);
 	if (unlikely(ret == -1))
 		return DID_ERROR << 16;
Signature · 1 line
2.55.0
#16 lore raw

PATCH v9 10/11 scsi: scsi_debug: Map the region written by WRITE ATOMIC (16)

In reply to Niklas Cassel

When logical block provisioning is enabled, every command that writes user data calls map_region() so that GET LBA STATUS reports the region as mapped. resp_atomic_write() does not, so GET LBA STATUS keeps reporting a region as deallocated after a WRITE ATOMIC (16) has written it.

Call map_region() the way resp_write_dt0() does, holding the zone metadata write lock across the access. That lock is what serialises the map against resp_unmap(), which does nothing unless logical block provisioning is enabled.

Assisted-by LLM
Fixes 84f3a3c01d70 ("scsi: scsi_debug: Atomic write support")

---
Tested with:

  modprobe scsi_debug sector_size=512 physblk_exp=3 dev_size_mb=128 \
      atomic_wr=1 lbpu=1

GET LBA STATUS reports a region that has never been written as deallocated, and reports it as mapped after a WRITE ATOMIC (16) of eight blocks. Only an ordinary WRITE did so before this patch.

---
 drivers/scsi/scsi_debug.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
drivers/scsi/scsi_debug.c +14 -0
index 1753b7b087b3..11f2565fb676 100644
@@ -4876,6 +4876,11 @@ static unsigned int map_state(struct sdeb_store_info *sip, sector_t lba,
 	return mapped;
 }
 
+/*
+ * Callers may map the whole range of a write without checking how much of
+ * it was written. SBC-6 4.7.4.6.2 allows a deallocated LBA to become mapped
+ * at any time, and only requires an LBA that was written to be mapped.
+ */
 static void map_region(struct sdeb_store_info *sip, sector_t lba,
 		       unsigned int len)
 {
@@ -6202,6 +6207,7 @@ static int resp_atomic_write(struct scsi_cmnd *scp,
 	u8 *cmd = scp->cmnd;
 	u16 boundary, len;
 	u64 lba, lba_tmp;
+	bool lbp = scsi_debug_lbp();
 	int ret;
 
 	if (!scsi_debug_atomic_write()) {
@@ -6250,7 +6256,15 @@ static int resp_atomic_write(struct scsi_cmnd *scp,
 	if (scsi_bufflen(scp) < len * sdebug_sector_size)
 		return DID_ERROR << 16;
 
+	if (lbp)
+		sdeb_meta_write_lock(sip);
+
 	ret = do_device_access(sip, scp, 0, lba, len, 0, true, true);
+	if (lbp) {
+		map_region(sip, lba, len);
+		sdeb_meta_write_unlock(sip);
+	}
+
 	if (unlikely(ret == -1))
 		return DID_ERROR << 16;
 	if (unlikely(ret != len * sdebug_sector_size))
Signature · 1 line
2.55.0
#17 lore raw

PATCH v9 11/11 scsi: scsi_debug: Validate the access parameters of WRITE ATOMIC (16)

In reply to Niklas Cassel

resp_atomic_write() validates the fields that are specific to an atomic write, but never the range that the command addresses. Every other command that writes user data calls check_device_access_params() first, which rejects a transfer that ends beyond the capacity of the device, one whose length exceeds the size of the store, and any write to a write protected device.

Call check_device_access_params(). The zone checks that it ends with are unreachable, as atomic writes and ZBC emulation are mutually exclusive.

Assisted-by LLM
Fixes 84f3a3c01d70 ("scsi: scsi_debug: Atomic write support")

---
Tested with:

  modprobe scsi_debug sector_size=512 physblk_exp=3 dev_size_mb=128 \
      atomic_wr=1

The device has a capacity of 262144 logical blocks. A WRITE ATOMIC (16) of eight blocks at LBA 262140 is now terminated with ILLEGAL REQUEST / LOGICAL BLOCK ADDRESS OUT OF RANGE; before this patch it completed with GOOD status, having written eight blocks at LBA 0 instead.

With the wp module parameter set to 1, a WRITE ATOMIC (16) is now terminated with DATA PROTECT / LOGICAL UNIT SOFTWARE WRITE PROTECTED, which is what an ordinary WRITE(16) has always returned; before this patch it completed with GOOD status and wrote the data.

---
 drivers/scsi/scsi_debug.c | 4 ++++
 1 file changed, 4 insertions(+)
drivers/scsi/scsi_debug.c +4 -0
index 11f2565fb676..e4ad689da45e 100644
@@ -6252,6 +6252,10 @@ static int resp_atomic_write(struct scsi_cmnd *scp,
 		}
 	}
 
+	ret = check_device_access_params(scp, lba, len, true);
+	if (ret)
+		return ret;
+
 	/* Short atomic writes are not allowed. */
 	if (scsi_bufflen(scp) < len * sdebug_sector_size)
 		return DID_ERROR << 16;
Signature · 1 line
2.55.0