cxl keeps a driver use count, which is used with the hash memory model on p8 to know when to upgrade local TLBIs to global and to trigger callbacks to manage the MMU for PSL8.

If a process opens a context and closes without attaching or fails the attachment, the driver use count is never decremented. As a consequence, TLB invalidations remain global, even if there are no active cxl contexts.

We should increment the driver use count when the process is attaching to the cxl adapter, and not on open. It's not needed before the adapter starts using the context and the use count is decremented on the detach path, so it makes more sense.

It affects only the user api. The kernel api is already doing The Right Thing.

Signed-off-by Frederic Barrat <[email protected]>

Fixes 7bb5d91a4dda ("cxl: Rework context lifetimes")

Acked-by Andrew Donnellan <[email protected]>

Signed-off-by Michael Ellerman <[email protected]>

[ajd: backport to stable v4.4 tree]

Signed-off-by Andrew Donnellan <[email protected]>

Signed-off-by Greg Kroah-Hartman <[email protected]>

---
 drivers/misc/cxl/api.c  |    4 ++++
 drivers/misc/cxl/file.c |    8 +++++++-
 2 files changed, 11 insertions(+), 1 deletion(-)

drivers/misc/cxl/api.c +4 -0

@@ -176,6 +176,10 @@ int cxl_start_context(struct cxl_context
 		kernel = false;
 	}
 
+	/*
+	 * Increment driver use count. Enables global TLBIs for hash
+	 * and callbacks to handle the segment table
+	 */
 	cxl_ctx_get();
 
 	if ((rc = cxl_attach_process(ctx, kernel, wed , 0))) {

drivers/misc/cxl/file.c +7 -1

@@ -94,7 +94,6 @@ static int __afu_open(struct inode *inod
 
 	pr_devel("afu_open pe: %i\n", ctx->pe);
 	file->private_data = ctx;
-	cxl_ctx_get();
 
 	/* indicate success */
 	rc = 0;

@@ -205,11 +204,18 @@ static long afu_ioctl_start_work(struct
 	ctx->pid = get_task_pid(current, PIDTYPE_PID);
 	ctx->glpid = get_task_pid(current->group_leader, PIDTYPE_PID);
 
+	/*
+	 * Increment driver use count. Enables global TLBIs for hash
+	 * and callbacks to handle the segment table
+	 */
+	cxl_ctx_get();
+
 	trace_cxl_attach(ctx, work.work_element_descriptor, work.num_interrupts, amr);
 
 	if ((rc = cxl_attach_process(ctx, false, work.work_element_descriptor,
 				     amr))) {
 		afu_release_irqs(ctx, ctx);
+		cxl_ctx_put();
 		goto out;
 	}

Patches currently in stable-queue which might be from [email protected] are

queue-4.4/cxl-fix-driver-use-count.patch