drivers: hwmon: occ: Cancel occ operations in remove()

From: "Edward A. James" <[email protected]>

Prevent hanging forever waiting for OCC ops to complete.

Signed-off-by Edward A. James <[email protected]>

Changes since v4:
* Add comments for the spinlock

---
 drivers/hwmon/occ/p9_sbe.c | 49 +++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 42 insertions(+), 7 deletions(-)

drivers/hwmon/occ/p9_sbe.c +16 -0

index c7e0d9c..ac18422 100644
@@ -14,37 +14,68 @@
 #include <linux/platform_device.h>
 #include <linux/occ.h>
 #include <linux/sched.h>
+#include <linux/spinlock.h>
 #include <linux/workqueue.h>
 
 struct p9_sbe_occ {
        struct occ occ;
        struct device *sbe;
+
+       /*
+        * Pointer to occ device client. We store this so that we can cancel
+        * the client operations in remove() if necessary. We only need one
+        * pointer since we do one OCC operation (open, write, read, close) at
+        * a time (access to p9_sbe_occ_send_cmd is locked in the common code
+        * with occ.lock).
+        */
+       struct occ_client *client;
+
+       /*
+        * This lock controls access to the client pointer and ensures atomic
+        * open, close and NULL assignment. This prevents simultaneous opening
+        * and closing of the client, or closing multiple times.
+        */

drivers/hwmon/occ/p9_sbe.c +20 -8

+       tmp_client = occ->client;
+       occ->client = NULL;
+       occ_drv_release(tmp_client);
+       spin_unlock_irq(&occ->lock);
+}
+
 static int p9_sbe_occ_send_cmd(struct occ *occ, u8 *cmd)
 {
        int rc, error;
-       struct occ_client *client;
        struct occ_response *resp = &occ->resp;
        struct p9_sbe_occ *p9_sbe_occ = to_p9_sbe_occ(occ);
 
-       client = occ_drv_open(p9_sbe_occ->sbe, 0);
-       if (!client) {
+       spin_lock_irq(&p9_sbe_occ->lock);
+       if (p9_sbe_occ->sbe)
+               p9_sbe_occ->client = occ_drv_open(p9_sbe_occ->sbe, 0);
+       spin_unlock_irq(&p9_sbe_occ->lock);
+
+       if (!p9_sbe_occ->client) {
                rc = -ENODEV;
                goto assign;
        }
 
-       rc = occ_drv_write(client, (const char *)&cmd[1], 7);
+       rc = occ_drv_write(p9_sbe_occ->client, (const char *)&cmd[1], 7);
        if (rc < 0)
                goto err;
 
-       rc = occ_drv_read(client, (char *)resp, sizeof(*resp));
+       rc = occ_drv_read(p9_sbe_occ->client, (char *)resp, sizeof(*resp));
        if (rc < 0)
                goto err;
 
-       occ_drv_release(client);
+       p9_sbe_occ_close_client(p9_sbe_occ);
 
        switch (resp->return_status) {
        case RESP_RETURN_CMD_IN_PRG:

@@ -72,7 +103,7 @@ static int p9_sbe_occ_send_cmd(struct occ *occ, u8 *cmd)
        goto done;
 
 err:
-       occ_drv_release(client);
+       p9_sbe_occ_close_client(p9_sbe_occ);
        dev_err(occ->bus_dev, "occ bus op failed rc:%d\n", rc);
 assign:
        error = rc;

@@ -132,6 +163,7 @@ static int p9_sbe_occ_probe(struct platform_device *pdev)
        p9_sbe_occ->sbe = pdev->dev.parent;
 
        occ = &p9_sbe_occ->occ;
+       spin_lock_init(&p9_sbe_occ->lock);
        occ->bus_dev = &pdev->dev;
        occ->groups[0] = &occ->group;
        occ->poll_cmd_data = 0x20;

@@ -152,7 +184,10 @@ static int p9_sbe_occ_probe(struct platform_device *pdev)
 static int p9_sbe_occ_remove(struct platform_device *pdev)
 {
        struct occ *occ = platform_get_drvdata(pdev);
+       struct p9_sbe_occ *p9_sbe_occ = to_p9_sbe_occ(occ);
 
+       p9_sbe_occ->sbe = NULL;
+       p9_sbe_occ_close_client(p9_sbe_occ);
        occ_remove_status_attrs(occ);
 
        atomic_dec(&occ_num_occs);
--

1.8.3.1

Quoted 44 lines · Eddie James

From: "Edward A. James" <[email protected]>

Prevent hanging forever waiting for OCC ops to complete.

Signed-off-by Edward A. James <[email protected]>

Changes since v4:
* Add comments for the spinlock

---
  drivers/hwmon/occ/p9_sbe.c | 49 +++++++++++++++++++++++++++++++++++++++-------
  1 file changed, 42 insertions(+), 7 deletions(-)

drivers/hwmon/occ/p9_sbe.c +17 -0

index c7e0d9c..ac18422 100644
@@ -14,37 +14,68 @@
  #include <linux/platform_device.h>
  #include <linux/occ.h>
  #include <linux/sched.h>
+#include <linux/spinlock.h>
  #include <linux/workqueue.h>
 
  struct p9_sbe_occ {
         struct occ occ;
         struct device *sbe;
+
+       /*
+        * Pointer to occ device client. We store this so that we can cancel
+        * the client operations in remove() if necessary. We only need one
+        * pointer since we do one OCC operation (open, write, read, close) at
+        * a time (access to p9_sbe_occ_send_cmd is locked in the common code
+        * with occ.lock).
+        */
+       struct occ_client *client;
+
+       /*
+        * This lock controls access to the client pointer and ensures atomic
+        * open, close and NULL assignment. This prevents simultaneous opening
+        * and closing of the client, or closing multiple times.
+        */
+       spinlock_t lock;
  };
 

drivers/hwmon/occ/p9_sbe.c line 47

Quoted 7 lines

drivers/hwmon/occ/p9_sbe.c +5 -0

  #define to_p9_sbe_occ(x)       container_of((x), struct p9_sbe_occ, occ)
 
+static void p9_sbe_occ_close_client(struct p9_sbe_occ *occ)
+{
+       struct occ_client *tmp_client;
+
+       spin_lock_irq(&occ->lock);

Why not spin_lock_irqsave/spin_unlock_irqsave? Same throughout this patch.

Cheers,

Joel

Quoted 71 lines

drivers/hwmon/occ/p9_sbe.c +20 -8

+       tmp_client = occ->client;
+       occ->client = NULL;
+       occ_drv_release(tmp_client);
+       spin_unlock_irq(&occ->lock);
+}
+
  static int p9_sbe_occ_send_cmd(struct occ *occ, u8 *cmd)
  {
         int rc, error;
-       struct occ_client *client;
         struct occ_response *resp = &occ->resp;
         struct p9_sbe_occ *p9_sbe_occ = to_p9_sbe_occ(occ);
 
-       client = occ_drv_open(p9_sbe_occ->sbe, 0);
-       if (!client) {
+       spin_lock_irq(&p9_sbe_occ->lock);
+       if (p9_sbe_occ->sbe)
+               p9_sbe_occ->client = occ_drv_open(p9_sbe_occ->sbe, 0);
+       spin_unlock_irq(&p9_sbe_occ->lock);
+
+       if (!p9_sbe_occ->client) {
                 rc = -ENODEV;
                 goto assign;
         }
 
-       rc = occ_drv_write(client, (const char *)&cmd[1], 7);
+       rc = occ_drv_write(p9_sbe_occ->client, (const char *)&cmd[1], 7);
         if (rc < 0)
                 goto err;
 
-       rc = occ_drv_read(client, (char *)resp, sizeof(*resp));
+       rc = occ_drv_read(p9_sbe_occ->client, (char *)resp, sizeof(*resp));
         if (rc < 0)
                 goto err;
 
-       occ_drv_release(client);
+       p9_sbe_occ_close_client(p9_sbe_occ);
 
         switch (resp->return_status) {
         case RESP_RETURN_CMD_IN_PRG:

@@ -72,7 +103,7 @@ static int p9_sbe_occ_send_cmd(struct occ *occ, u8 *cmd)
         goto done;
 
  err:
-       occ_drv_release(client);
+       p9_sbe_occ_close_client(p9_sbe_occ);
         dev_err(occ->bus_dev, "occ bus op failed rc:%d\n", rc);
  assign:
         error = rc;

@@ -132,6 +163,7 @@ static int p9_sbe_occ_probe(struct platform_device *pdev)
         p9_sbe_occ->sbe = pdev->dev.parent;
 
         occ = &p9_sbe_occ->occ;
+       spin_lock_init(&p9_sbe_occ->lock);
         occ->bus_dev = &pdev->dev;
         occ->groups[0] = &occ->group;
         occ->poll_cmd_data = 0x20;

@@ -152,7 +184,10 @@ static int p9_sbe_occ_probe(struct platform_device *pdev)
  static int p9_sbe_occ_remove(struct platform_device *pdev)
  {
         struct occ *occ = platform_get_drvdata(pdev);
+       struct p9_sbe_occ *p9_sbe_occ = to_p9_sbe_occ(occ);
 
+       p9_sbe_occ->sbe = NULL;
+       p9_sbe_occ_close_client(p9_sbe_occ);
         occ_remove_status_attrs(occ);
 
         atomic_dec(&occ_num_occs);
--

1.8.3.1

Quoted 7 lines

drivers/hwmon/occ/p9_sbe.c line 47

Quoted 5 lines

drivers/hwmon/occ/p9_sbe.c +5 -0

+static void p9_sbe_occ_close_client(struct p9_sbe_occ *occ)
+{
+       struct occ_client *tmp_client;
+
+       spin_lock_irq(&occ->lock);

Why not spin_lock_irqsave/spin_unlock_irqsave? Same throughout this patch.

Well spin_lock_irq/spin_unlock_irq is used through the occ and sbefifo drivers. Just following suit here. I can switch it if necessary.

Quoted 6 lines

Quoted 2 lines

Why not spin_lock_irqsave/spin_unlock_irqsave? Same throughout this patch.

Well spin_lock_irq/spin_unlock_irq is used through the occ and sbefifo drivers. Just following suit here. I can switch it if necessary.

If you understand the differences and used the non-save/restore versions intentionally, then we can keep the driver as it is.

In general, to be safe, more experienced kernel hackers than myself recommend the irqsave/irqrestore versions of the spin lock functions as they are safe to call from any context (IRQ or otherwise).

Cheers,

Joel