Skip to contentSkip to conversation

scsi: libsas: Fix disk not being scanned in after being removed

PATCH Superseded Xingui Yang 3 participants 4 messages Feb 27, 2024 – Mar 7, 2024

Subsystems (2)

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 scsi: libsas: Fix disk not being scanned in after being removed

As of commit d8649fc1c5e4 ("scsi: libsas: Do discovery on empty PHY to update PHY info"), do discovery will send a new SMP_DISCOVER and update phy->phy_change_count. We found that if the disk is reconnected and phy change_count changes at this time, the disk scanning process will not be triggered.

So update the PHY info with the last query results.

Fixes d8649fc1c5e4 ("scsi: libsas: Do discovery on empty PHY to update PHY info")

---
v1 -> v2:
Use sas_get_phy_discover() instead of sas_get_phy_attached_dev() in sas_rediscover_dev() and use disc_resp to update phy info.

---
 drivers/scsi/libsas/sas_expander.c | 37 ++++++++++++++++++++++--------
 1 file changed, 28 insertions(+), 9 deletions(-)
drivers/scsi/libsas/sas_expander.c +28 -9
index a2204674b680..a75dcce7a9ba 100644
@@ -1940,6 +1940,7 @@ static int sas_rediscover_dev(struct domain_device *dev, int phy_id,
 	struct expander_device *ex = &dev->ex_dev;
 	struct ex_phy *phy = &ex->ex_phy[phy_id];
 	enum sas_device_type type = SAS_PHY_UNUSED;
+	struct smp_disc_resp *disc_resp;
 	u8 sas_addr[SAS_ADDR_SIZE];
 	char msg[80] = "";
 	int res;
@@ -1951,33 +1952,47 @@ static int sas_rediscover_dev(struct domain_device *dev, int phy_id,
 		 SAS_ADDR(dev->sas_addr), phy_id, msg);
 
 	memset(sas_addr, 0, SAS_ADDR_SIZE);
-	res = sas_get_phy_attached_dev(dev, phy_id, sas_addr, &type);
+	disc_resp = alloc_smp_resp(DISCOVER_RESP_SIZE);
+	if (!disc_resp)
+		return -ENOMEM;
+
+	res = sas_get_phy_discover(dev, phy_id, disc_resp);
 	switch (res) {
 	case SMP_RESP_NO_PHY:
 		phy->phy_state = PHY_NOT_PRESENT;
 		sas_unregister_devs_sas_addr(dev, phy_id, last);
-		return res;
+		goto out;
 	case SMP_RESP_PHY_VACANT:
 		phy->phy_state = PHY_VACANT;
 		sas_unregister_devs_sas_addr(dev, phy_id, last);
-		return res;
+		goto out;
 	case SMP_RESP_FUNC_ACC:
 		break;
 	case -ECOMM:
 		break;
 	default:
-		return res;
+		goto out;
+	}
+
+	if (res == 0) {
+		struct discover_resp *dr = &disc_resp->disc;
+
+		memcpy(sas_addr, dr->attached_sas_addr, SAS_ADDR_SIZE);
+		type = to_dev_type(dr);
+		if (type == 0)
+			memset(sas_addr, 0, SAS_ADDR_SIZE);
 	}
 
 	if ((SAS_ADDR(sas_addr) == 0) || (res == -ECOMM)) {
 		phy->phy_state = PHY_EMPTY;
 		sas_unregister_devs_sas_addr(dev, phy_id, last);
 		/*
-		 * Even though the PHY is empty, for convenience we discover
-		 * the PHY to update the PHY info, like negotiated linkrate.
+		 * Even though the PHY is empty, for convenience we update
+		 * the PHY info, like negotiated linkrate.
 		 */
-		sas_ex_phy_discover(dev, phy_id);
-		return res;
+		if (res == 0)
+			sas_set_ex_phy(dev, phy_id, disc_resp);
+		goto out;
 	} else if (SAS_ADDR(sas_addr) == SAS_ADDR(phy->attached_sas_addr) &&
 		   dev_type_flutter(type, phy->attached_dev_type)) {
 		struct domain_device *ata_dev = sas_ex_to_ata(dev, phy_id);
@@ -1989,7 +2004,7 @@ static int sas_rediscover_dev(struct domain_device *dev, int phy_id,
 			action = ", needs recovery";
 		pr_debug("ex %016llx phy%02d broadcast flutter%s\n",
 			 SAS_ADDR(dev->sas_addr), phy_id, action);
-		return res;
+		goto out;
 	}
 
 	/* we always have to delete the old device when we went here */
@@ -1998,7 +2013,11 @@ static int sas_rediscover_dev(struct domain_device *dev, int phy_id,
 		SAS_ADDR(phy->attached_sas_addr));
 	sas_unregister_devs_sas_addr(dev, phy_id, last);
 
+	kfree(disc_resp);
 	return sas_discover_new(dev, phy_id);
+out:
+	kfree(disc_resp);
+	return res;
 }
 
 /**
Signature · 1 line
2.17.1
reviewer Reviewed-by on v2 1/1
#1 lore raw
Quoted 10 lines · Xingui Yang

As of commit d8649fc1c5e4 ("scsi: libsas: Do discovery on empty PHY to update PHY info"), do discovery will send a new SMP_DISCOVER and update phy->phy_change_count. We found that if the disk is reconnected and phy change_count changes at this time, the disk scanning process will not be triggered.

So update the PHY info with the last query results.

Fixes d8649fc1c5e4 ("scsi: libsas: Do discovery on empty PHY to update PHY info")
Quoted 7 lines

---
v1 -> v2:
Use sas_get_phy_discover() instead of sas_get_phy_attached_dev() in sas_rediscover_dev() and use disc_resp to update phy info.

---
  drivers/scsi/libsas/sas_expander.c | 37 ++++++++++++++++++++++--------
  1 file changed, 28 insertions(+), 9 deletions(-)

Looks good to me:

reviewer
#2 lore raw
In reply to Xingui Yang
Quoted 7 lines · Xingui Yang

As of commit d8649fc1c5e4 ("scsi: libsas: Do discovery on empty PHY to update PHY info"), do discovery will send a new SMP_DISCOVER and update phy->phy_change_count. We found that if the disk is reconnected and phy change_count changes at this time, the disk scanning process will not be triggered.

So update the PHY info with the last query results.

Please mention that sas_unregister_devs_sas_addr() should be called before sas_set_ex_phy() for this case.

Quoted 49 lines
Fixes d8649fc1c5e4 ("scsi: libsas: Do discovery on empty PHY to update PHY info")

---
v1 -> v2:
Use sas_get_phy_discover() instead of sas_get_phy_attached_dev() in sas_rediscover_dev() and use disc_resp to update phy info.

---
  drivers/scsi/libsas/sas_expander.c | 37 ++++++++++++++++++++++--------
  1 file changed, 28 insertions(+), 9 deletions(-)
drivers/scsi/libsas/sas_expander.c +8 -3
index a2204674b680..a75dcce7a9ba 100644
@@ -1940,6 +1940,7 @@ static int sas_rediscover_dev(struct domain_device *dev, int phy_id,
  	struct expander_device *ex = &dev->ex_dev;
  	struct ex_phy *phy = &ex->ex_phy[phy_id];
  	enum sas_device_type type = SAS_PHY_UNUSED;
+	struct smp_disc_resp *disc_resp;
  	u8 sas_addr[SAS_ADDR_SIZE];
  	char msg[80] = "";
  	int res;
@@ -1951,33 +1952,47 @@ static int sas_rediscover_dev(struct domain_device *dev, int phy_id,
  		 SAS_ADDR(dev->sas_addr), phy_id, msg);
 
  	memset(sas_addr, 0, SAS_ADDR_SIZE);
-	res = sas_get_phy_attached_dev(dev, phy_id, sas_addr, &type);
+	disc_resp = alloc_smp_resp(DISCOVER_RESP_SIZE);
+	if (!disc_resp)
+		return -ENOMEM;
+
+	res = sas_get_phy_discover(dev, phy_id, disc_resp);
  	switch (res) {
  	case SMP_RESP_NO_PHY:
  		phy->phy_state = PHY_NOT_PRESENT;
  		sas_unregister_devs_sas_addr(dev, phy_id, last);
-		return res;
+		goto out;
  	case SMP_RESP_PHY_VACANT:
  		phy->phy_state = PHY_VACANT;
  		sas_unregister_devs_sas_addr(dev, phy_id, last);
-		return res;
+		goto out;
  	case SMP_RESP_FUNC_ACC:
  		break;
  	case -ECOMM:
  		break;
  	default:

drivers/scsi/libsas/sas_expander.c static int sas_rediscover_dev(struct domain_device *dev, int phy_id, · line 1982

Quoted 10 lines
drivers/scsi/libsas/sas_expander.c +9 -1
-		return res;
+		goto out;
+	}
+
+	if (res == 0) {
+		struct discover_resp *dr = &disc_resp->disc;
+
+		memcpy(sas_addr, dr->attached_sas_addr, SAS_ADDR_SIZE);
+		type = to_dev_type(dr);
+		if (type == 0)

Please use SAS_PHY_UNUSED

drivers/scsi/libsas/sas_expander.c static int sas_rediscover_dev(struct domain_device *dev, int phy_id, · line 1984

Quoted 2 lines
drivers/scsi/libsas/sas_expander.c +1 -0
+			memset(sas_addr, 0, SAS_ADDR_SIZE);
  	}

Any chance we can factor out some of this code with sas_get_phy_attached_dev() (where it has been copied from)?

Quoted 19 lines
drivers/scsi/libsas/sas_expander.c +5 -4
  	if ((SAS_ADDR(sas_addr) == 0) || (res == -ECOMM)) {
  		phy->phy_state = PHY_EMPTY;
  		sas_unregister_devs_sas_addr(dev, phy_id, last);
  		/*
-		 * Even though the PHY is empty, for convenience we discover
-		 * the PHY to update the PHY info, like negotiated linkrate.
+		 * Even though the PHY is empty, for convenience we update
+		 * the PHY info, like negotiated linkrate.
  		 */
-		sas_ex_phy_discover(dev, phy_id);
-		return res;
+		if (res == 0)
+			sas_set_ex_phy(dev, phy_id, disc_resp);
+		goto out;
  	} else if (SAS_ADDR(sas_addr) == SAS_ADDR(phy->attached_sas_addr) &&
  		   dev_type_flutter(type, phy->attached_dev_type)) {
  		struct domain_device *ata_dev = sas_ex_to_ata(dev, phy_id);

drivers/scsi/libsas/sas_expander.c static int sas_rediscover_dev(struct domain_device *dev, int phy_id, · line 2007

Quoted 5 lines
drivers/scsi/libsas/sas_expander.c +1 -1
@@ -1989,7 +2004,7 @@ static int sas_rediscover_dev(struct domain_device *dev, int phy_id,
  			action = ", needs recovery";
  		pr_debug("ex %016llx phy%02d broadcast flutter%s\n",
  			 SAS_ADDR(dev->sas_addr), phy_id, action);
-		return res;
+		goto out;

is there any way to get rid of all the goto's? Ideally we would have just one location to call kfree(disc_resp)

drivers/scsi/libsas/sas_expander.c static int sas_rediscover_dev(struct domain_device *dev, int phy_id, · line 2018

Quoted 10 lines
drivers/scsi/libsas/sas_expander.c +2 -0
  	}
 
  	/* we always have to delete the old device when we went here */
@@ -1998,7 +2013,11 @@ static int sas_rediscover_dev(struct domain_device *dev, int phy_id,
  		SAS_ADDR(phy->attached_sas_addr));
  	sas_unregister_devs_sas_addr(dev, phy_id, last);
 
+	kfree(disc_resp);
  	return sas_discover_new(dev, phy_id);
+out:

Calling this simply "out" is odd, as we exit in many places

drivers/scsi/libsas/sas_expander.c static int sas_rediscover_dev(struct domain_device *dev, int phy_id, · line 2023

Quoted 5 lines
drivers/scsi/libsas/sas_expander.c +2 -0
+	kfree(disc_resp);
+	return res;
  }
 
  /**

As an aside, could libsas - and your changes here - be simpler if we changed smp_execute_task() like this:

static int smp_execute_task(struct domain_device *dev, void *req /* can be on the stack */, int req_size,
void *resp /* can be on the stack */, int resp_size)

{
	struct scatterlist req_sg;
	struct scatterlist resp_sg;
	int ret;
	void *_req = kmemdup(req, req_size, GFP_KERNEL);
	void *_resp = alloc_smp_resp(resp_size);
	if (!_req || !resp)
		return -ENOMEM;

	sg_init_one(&req_sg, _req, req_size);
	sg_init_one(&resp_sg, _resp, resp_size);
	ret = smp_execute_task_sg(dev, &req_sg, &resp_sg);
	memcpy(resp, _resp, resp_size);
	kfree(_req);
	kfree(_resp);
	return ret;
}

We need to use alloc_smp_resp() and alloc_smp_req() as we can't allocate these memories on the stack for calling sg_init_one(). But if we changed smp_execute_task() to memcpy from/to data on the stack, it might make callers simpler. I'm not sure.

Thanks,
John

reviewer
#3 lore raw
Quoted 22 lines · John Garry

As an aside, could libsas - and your changes here - be simpler if we changed smp_execute_task() like this:

static int smp_execute_task(struct domain_device *dev, void *req /* can be on the stack */, int req_size,
                void *resp /* can be on the stack */, int resp_size)
{
    struct scatterlist req_sg;
    struct scatterlist resp_sg;
    int ret;
    void *_req = kmemdup(req, req_size, GFP_KERNEL);
    void *_resp = alloc_smp_resp(resp_size);
    if (!_req || !resp)
        return -ENOMEM;

    sg_init_one(&req_sg, _req, req_size); sg_init_one(&resp_sg, _resp, resp_size); ret = smp_execute_task_sg(dev, &req_sg, &resp_sg); memcpy(resp, _resp, resp_size); kfree(_req);
    kfree(_resp);
    return ret;
}

Quoted 1 line

We need to use alloc_smp_resp() and alloc_smp_req() as we can't allocate these memories on the stack for calling sg_init_one(). But if we changed smp_execute_task() to memcpy from/to data on the stack, it might make callers simpler. I'm not sure.

Maybe simpler. I have not check all the length of these buffers, but there is still a risk of stack overflow if the buffer on stack is too large.

Thanks,
Jason