Merge tag 'misc-habanalabs-next-2021-06-22' of https://git.kernel.org…

…/pub/scm/linux/kernel/git/ogabbay/linux into char-misc-next

Oded writes:

This tag contains habanalabs driver changes for v5.14:

- Change communication protocol with f/w. The new protocl allows better
  backward compatibility between different f/w versions and is more
  stable.
- Send hard-reset cause to f/w after a hard-reset has happened.
- Move to indirection when generating interrupts to f/w.
- Better progress and error messages during the f/w load stage.
- Recognize that f/w is with enabled security according to device ID.
- Add validity check to event queue mechanism.
- Add new event from f/w that will indicate a daemon has been terminated
  inside the f/w.

- Move to TLB cache range invalidation in the device's MMU.
- Disable memory scrubbing by default for performance.

- Many fixes for sparse/smatch reported errors.
- Enable by default stop-on-err in the ASIC.
- Move to ASYNC device probing to speedup loading of driver in server
  with multiple devices.
- Fix to stop using disabled NIC ports when doing collective operation.
- Use standard error codes instead of positive values.
- Add support for resetting device after user has finished using it.
- Add debugfs option to avoid reset when a CS has got stuck.
- Add print of the last 8 CS pointers in case of error in QMANs.
- Add statistics on opening of the FD of a device.

* tag 'misc-habanalabs-next-2021-06-22' of https://git.kernel.org/pub/scm/linux/kernel/git/ogabbay/linux: (72 commits)
  habanalabs/gaudi: refactor hard-reset related code
  habanalabs/gaudi: add support for NIC DERR
  habanalabs: add validity check for signal cs
  habanalabs: get lower/upper 32 bits via masking
  habanalabs: allow reset upon device release
  debugfs: add skip_reset_on_timeout option
  habanalabs: fix typo
  habanalabs/gaudi: correct driver events numbering
  habanalabs: remove a rogue #ifdef
  habanalabs/gaudi: print last QM PQEs on error
  habanalabs/goya: add '__force' attribute to suppress false alarm
  habanalabs: added open_stats info ioctl
  habanalabs/gaudi: set the correct rc in case of err
  habanalabs/gaudi: update coresight configuration
  habanalabs: remove node from list before freeing the node
  habanalabs: set rc as 'valid' in case of intentional func exit
  habanalabs: zero complex structures using memset
  habanalabs: print more info when failing to pin user memory
  habanalabs: Fix an error handling path in 'hl_pci_probe()'
  habanalabs: print firmware versions
  ...

Documentation/ABI/testing/debugfs-driver-habanalabs

@@ -207,6 +207,14 @@ Contact:        [email protected]
 Description:    Sets the PCI power state. Valid values are "1" for D0 and "2"
                 for D3Hot
 
+What:           /sys/kernel/debug/habanalabs/hl<n>/skip_reset_on_timeout
+Date:           Jun 2021
+KernelVersion:  5.13
+Contact:        [email protected]
+Description:    Sets the skip reset on timeout option for the device. Value of
+                "0" means device will be reset in case some CS has timed out,
+                otherwise it will not be reset.
+
 What:           /sys/kernel/debug/habanalabs/hl<n>/stop_on_err
 Date:           Mar 2020
 KernelVersion:  5.6

drivers/misc/habanalabs/common/command_submission.c

@@ -556,6 +556,13 @@ static void cs_do_release(struct kref *ref)
 	else if (!cs->submitted)
 		cs->fence->error = -EBUSY;
 
+	if (unlikely(cs->skip_reset_on_timeout)) {
+		dev_err(hdev->dev,
+			"Command submission %llu completed after %llu (s)\n",
+			cs->sequence,
+			div_u64(jiffies - cs->submission_time_jiffies, HZ));
+	}
+
 	if (cs->timestamp)
 		cs->fence->timestamp = ktime_get();
 	complete_all(&cs->fence->completion);
@@ -571,6 +578,8 @@ static void cs_timedout(struct work_struct *work)
 	int rc;
 	struct hl_cs *cs = container_of(work, struct hl_cs,
 						 work_tdr.work);
+	bool skip_reset_on_timeout = cs->skip_reset_on_timeout;
+
 	rc = cs_get_unless_zero(cs);
 	if (!rc)
 		return;
@@ -581,7 +590,8 @@ static void cs_timedout(struct work_struct *work)
 	}
 
 	/* Mark the CS is timed out so we won't try to cancel its TDR */
-	cs->timedout = true;
+	if (likely(!skip_reset_on_timeout))
+		cs->timedout = true;
 
 	hdev = cs->ctx->hdev;
 
@@ -613,10 +623,12 @@ static void cs_timedout(struct work_struct *work)
 
 	cs_put(cs);
 
-	if (hdev->reset_on_lockup)
-		hl_device_reset(hdev, 0);
-	else
-		hdev->needs_reset = true;
+	if (likely(!skip_reset_on_timeout)) {
+		if (hdev->reset_on_lockup)
+			hl_device_reset(hdev, HL_RESET_TDR);
+		else
+			hdev->needs_reset = true;
+	}
 }
 
 static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
@@ -650,6 +662,10 @@ static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
 	cs->type = cs_type;
 	cs->timestamp = !!(flags & HL_CS_FLAGS_TIMESTAMP);
 	cs->timeout_jiffies = timeout;
+	cs->skip_reset_on_timeout =
+		hdev->skip_reset_on_timeout ||
+		!!(flags & HL_CS_FLAGS_SKIP_RESET_ON_TIMEOUT);
+	cs->submission_time_jiffies = jiffies;
 	INIT_LIST_HEAD(&cs->job_list);
 	INIT_DELAYED_WORK(&cs->work_tdr, cs_timedout);
 	kref_init(&cs->refcount);
@@ -1481,6 +1497,61 @@ static int hl_cs_ctx_switch(struct hl_fpriv *hpriv, union hl_cs_args *args,
 	return rc;
 }
 
+/*
+ * hl_cs_signal_sob_wraparound_handler: handle SOB value wrapaound case.
+ * if the SOB value reaches the max value move to the other SOB reserved
+ * to the queue.
+ * Note that this function must be called while hw_queues_lock is taken.
+ */
+int hl_cs_signal_sob_wraparound_handler(struct hl_device *hdev, u32 q_idx,
+			struct hl_hw_sob **hw_sob, u32 count)
+{
+	struct hl_sync_stream_properties *prop;
+	struct hl_hw_sob *sob = *hw_sob, *other_sob;
+	u8 other_sob_offset;
+
+	prop = &hdev->kernel_queues[q_idx].sync_stream_prop;
+
+	kref_get(&sob->kref);
+
+	/* check for wraparound */
+	if (prop->next_sob_val + count >= HL_MAX_SOB_VAL) {
+		/*
+		 * Decrement as we reached the max value.
+		 * The release function won't be called here as we've
+		 * just incremented the refcount right before calling this
+		 * function.
+		 */
+		kref_put(&sob->kref, hl_sob_reset_error);
+
+		/*
+		 * check the other sob value, if it still in use then fail
+		 * otherwise make the switch
+		 */
+		other_sob_offset = (prop->curr_sob_offset + 1) % HL_RSVD_SOBS;
+		other_sob = &prop->hw_sob[other_sob_offset];
+
+		if (kref_read(&other_sob->kref) != 1) {
+			dev_err(hdev->dev, "error: Cannot switch SOBs q_idx: %d\n",
+								q_idx);
+			return -EINVAL;
+		}
+
+		prop->next_sob_val = 1;
+
+		/* only two SOBs are currently in use */
+		prop->curr_sob_offset = other_sob_offset;
+		*hw_sob = other_sob;
+
+		dev_dbg(hdev->dev, "switched to SOB %d, q_idx: %d\n",
+				prop->curr_sob_offset, q_idx);
+	} else {
+		prop->next_sob_val += count;
+	}
+
+	return 0;
+}
+
 static int cs_ioctl_extract_signal_seq(struct hl_device *hdev,
 		struct hl_cs_chunk *chunk, u64 *signal_seq, struct hl_ctx *ctx)
 {

drivers/misc/habanalabs/common/context.c

@@ -12,7 +12,6 @@
 static void hl_ctx_fini(struct hl_ctx *ctx)
 {
 	struct hl_device *hdev = ctx->hdev;
-	u64 idle_mask[HL_BUSY_ENGINES_MASK_EXT_SIZE] = {0};
 	int i;
 
 	/* Release all allocated pending cb's, those cb's were never
@@ -57,14 +56,6 @@ static void hl_ctx_fini(struct hl_ctx *ctx)
 
 		/* Scrub both SRAM and DRAM */
 		hdev->asic_funcs->scrub_device_mem(hdev, 0, 0);
-
-		if ((!hdev->pldm) && (hdev->pdev) &&
-				(!hdev->asic_funcs->is_device_idle(hdev,
-					idle_mask,
-					HL_BUSY_ENGINES_MASK_EXT_SIZE, NULL)))
-			dev_notice(hdev->dev,
-					"device not idle after user context is closed (0x%llx, 0x%llx)\n",
-						idle_mask[0], idle_mask[1]);
 	} else {
 		dev_dbg(hdev->dev, "closing kernel context\n");
 		hdev->asic_funcs->ctx_fini(ctx);

drivers/misc/habanalabs/common/debugfs.c

@@ -1278,6 +1278,11 @@ void hl_debugfs_add_device(struct hl_device *hdev)
 				dev_entry->root,
 				&dev_entry->blob_desc);
 
+	debugfs_create_x8("skip_reset_on_timeout",
+				0644,
+				dev_entry->root,
+				&hdev->skip_reset_on_timeout);
+
 	for (i = 0, entry = dev_entry->entry_arr ; i < count ; i++, entry++) {
 		debugfs_create_file(hl_debugfs_list[i].name,
 					0444,

drivers/misc/habanalabs/common/device.c

@@ -51,6 +51,8 @@ bool hl_device_operational(struct hl_device *hdev,
 
 static void hpriv_release(struct kref *ref)
 {
+	u64 idle_mask[HL_BUSY_ENGINES_MASK_EXT_SIZE] = {0};
+	bool device_is_idle = true;
 	struct hl_fpriv *hpriv;
 	struct hl_device *hdev;
 
@@ -71,8 +73,20 @@ static void hpriv_release(struct kref *ref)
 
 	kfree(hpriv);
 
-	if (hdev->reset_upon_device_release)
-		hl_device_reset(hdev, 0);
+	if ((!hdev->pldm) && (hdev->pdev) &&
+			(!hdev->asic_funcs->is_device_idle(hdev,
+				idle_mask,
+				HL_BUSY_ENGINES_MASK_EXT_SIZE, NULL))) {
+		dev_err(hdev->dev,
+			"device not idle after user context is closed (0x%llx_%llx)\n",
+			idle_mask[1], idle_mask[0]);
+
+		device_is_idle = false;
+	}
+
+	if ((hdev->reset_if_device_not_idle && !device_is_idle)
+			|| hdev->reset_upon_device_release)
+		hl_device_reset(hdev, HL_RESET_DEVICE_RELEASE);
 }
 
 void hl_hpriv_get(struct hl_fpriv *hpriv)
@@ -118,6 +132,9 @@ static int hl_device_release(struct inode *inode, struct file *filp)
 		dev_warn(hdev->dev,
 			"Device is still in use because there are live CS and/or memory mappings\n");
 
+	hdev->last_open_session_duration_jif =
+		jiffies - hdev->last_successful_open_jif;
+
 	return 0;
 }
 
@@ -868,7 +885,7 @@ static void device_disable_open_processes(struct hl_device *hdev)
 int hl_device_reset(struct hl_device *hdev, u32 flags)
 {
 	u64 idle_mask[HL_BUSY_ENGINES_MASK_EXT_SIZE] = {0};
-	bool hard_reset, from_hard_reset_thread;
+	bool hard_reset, from_hard_reset_thread, hard_instead_soft = false;
 	int i, rc;
 
 	if (!hdev->init_done) {
@@ -880,11 +897,28 @@ int hl_device_reset(struct hl_device *hdev, u32 flags)
 	hard_reset = (flags & HL_RESET_HARD) != 0;
 	from_hard_reset_thread = (flags & HL_RESET_FROM_RESET_THREAD) != 0;
 
-	if ((!hard_reset) && (!hdev->supports_soft_reset)) {
-		dev_dbg(hdev->dev, "Doing hard-reset instead of soft-reset\n");
+	if (!hard_reset && !hdev->supports_soft_reset) {
+		hard_instead_soft = true;
+		hard_reset = true;
+	}
+
+	if (hdev->reset_upon_device_release &&
+			(flags & HL_RESET_DEVICE_RELEASE)) {
+		dev_dbg(hdev->dev,
+			"Perform %s-reset upon device release\n",
+			hard_reset ? "hard" : "soft");
+		goto do_reset;
+	}
+
+	if (!hard_reset && !hdev->allow_external_soft_reset) {
+		hard_instead_soft = true;
 		hard_reset = true;
 	}
 
+	if (hard_instead_soft)
+		dev_dbg(hdev->dev, "Doing hard-reset instead of soft-reset\n");
+
+do_reset:
 	/* Re-entry of reset thread */
 	if (from_hard_reset_thread && hdev->process_kill_trial_cnt)
 		goto kill_processes;
@@ -900,6 +934,19 @@ int hl_device_reset(struct hl_device *hdev, u32 flags)
 		if (rc)
 			return 0;
 
+		/*
+		 * 'reset cause' is being updated here, because getting here
+		 * means that it's the 1st time and the last time we're here
+		 * ('in_reset' makes sure of it). This makes sure that
+		 * 'reset_cause' will continue holding its 1st recorded reason!
+		 */
+		if (flags & HL_RESET_HEARTBEAT)
+			hdev->curr_reset_cause = HL_RESET_CAUSE_HEARTBEAT;
+		else if (flags & HL_RESET_TDR)
+			hdev->curr_reset_cause = HL_RESET_CAUSE_TDR;
+		else
+			hdev->curr_reset_cause = HL_RESET_CAUSE_UNKNOWN;
+
 		/*
 		 * if reset is due to heartbeat, device CPU is no responsive in
 		 * which case no point sending PCI disable message to it
@@ -943,9 +990,8 @@ int hl_device_reset(struct hl_device *hdev, u32 flags)
 		hdev->process_kill_trial_cnt = 0;
 
 		/*
-		 * Because the reset function can't run from interrupt or
-		 * from heartbeat work, we need to call the reset function
-		 * from a dedicated work
+		 * Because the reset function can't run from heartbeat work,
+		 * we need to call the reset function from a dedicated work.
 		 */
 		queue_delayed_work(hdev->device_reset_work.wq,
 			&hdev->device_reset_work.reset_work, 0);
@@ -1096,8 +1142,8 @@ int hl_device_reset(struct hl_device *hdev, u32 flags)
 	if (!hdev->asic_funcs->is_device_idle(hdev, idle_mask,
 			HL_BUSY_ENGINES_MASK_EXT_SIZE, NULL)) {
 		dev_err(hdev->dev,
-			"device is not idle (mask %#llx %#llx) after reset\n",
-			idle_mask[0], idle_mask[1]);
+			"device is not idle (mask 0x%llx_%llx) after reset\n",
+			idle_mask[1], idle_mask[0]);
 		rc = -EIO;
 		goto out_err;
 	}
@@ -1334,8 +1380,9 @@ int hl_device_init(struct hl_device *hdev, struct class *hclass)
 	}
 
 	/*
-	 * From this point, in case of an error, add char devices and create
-	 * sysfs nodes as part of the error flow, to allow debugging.
+	 * From this point, override rc (=0) in case of an error to allow
+	 * debugging (by adding char devices and create sysfs nodes as part of
+	 * the error flow).
 	 */
 	add_cdev_sysfs_on_err = true;
 
@@ -1369,7 +1416,7 @@ int hl_device_init(struct hl_device *hdev, struct class *hclass)
 
 	dev_info(hdev->dev, "Found %s device with %lluGB DRAM\n",
 		hdev->asic_name,
-		hdev->asic_prop.dram_size / 1024 / 1024 / 1024);
+		hdev->asic_prop.dram_size / SZ_1G);
 
 	rc = hl_vm_init(hdev);
 	if (rc) {
@@ -1475,13 +1522,19 @@ int hl_device_init(struct hl_device *hdev, struct class *hclass)
 void hl_device_fini(struct hl_device *hdev)
 {
 	ktime_t timeout;
+	u64 reset_sec;
 	int i, rc;
 
 	dev_info(hdev->dev, "Removing device\n");
 
 	hdev->device_fini_pending = 1;
 	flush_delayed_work(&hdev->device_reset_work.reset_work);
 
+	if (hdev->pldm)
+		reset_sec = HL_PLDM_HARD_RESET_MAX_TIMEOUT;
+	else
+		reset_sec = HL_HARD_RESET_MAX_TIMEOUT;
+
 	/*
 	 * This function is competing with the reset function, so try to
 	 * take the reset atomic and if we are already in middle of reset,
@@ -1490,8 +1543,7 @@ void hl_device_fini(struct hl_device *hdev)
 	 * ports, the hard reset could take between 10-30 seconds
 	 */
 
-	timeout = ktime_add_us(ktime_get(),
-				HL_HARD_RESET_MAX_TIMEOUT * 1000 * 1000);
+	timeout = ktime_add_us(ktime_get(), reset_sec * 1000 * 1000);
 	rc = atomic_cmpxchg(&hdev->in_reset, 0, 1);
 	while (rc) {
 		usleep_range(50, 200);