QDMA: release cdev region on device cleanup to stop major-number leak

[andrew-bolin]

Problem

qdma_cdev_device_cleanup() clears xcb->cdev_major but never calls
unregister_chrdev_region(). The 4096-minor character-device region that each
board allocates in qdma_cdev_device_init() is therefore leaked on every device
remove (for example a PCIe remove/rescan). The region is only ever reclaimed by
the module-wide qdma_cdev_cleanup() at module unload.

As a result, repeated remove/rescan cycles steadily consume the kernel's pool of
dynamic character-device major numbers. Once the pool is exhausted,
alloc_chrdev_region() returns -EBUSY and probe fails:

CHRDEV "qdma-pf" dynamic allocation region is full
qdma_pf:qdma_cdev_device_init: unable to allocate cdev region -16.
qdma-pf: probe of 0000:XX:00.1 failed with error -16

This was observed on with V80 FPGAs (QDMA PF on XX:00.1): after a number of
remove/rescan cycles, /proc/devices showed ~126 qdma-pf majors held, filling
the entire 384–511 dynamic range, and all subsequent probes failed with -16.
The only stock recovery is a full rmmod / modprobe, which is undesirable in
production.

Fix

Reference-count each board's major number. Multiple PFs on the same bus already
share one major via the dedup path in qdma_cdev_device_init(), so the count is
incremented there and initialised to 1 for a freshly allocated board.
qdma_cdev_device_cleanup() now decrements the count and, when the last PF on
the board goes away, releases the region with unregister_chrdev_region() +
list_del() + kfree(). The module-wide qdma_cdev_cleanup() remains as a
safety net.

With this change, remove/rescan no longer leaks majors and the -16 exhaustion
no longer occurs.

Testing

Tested on a host with six V80s (QDMA 2024.1.0, kernel
5.15.0-179-generic), counting grep -c qdma-pf /proc/devices across PCIe
remove/rescan cycles on one (idle) card:

• Before (stock module): baseline 6 majors → 7 → 8, climbing by one per
  remove/rescan cycle while the number of bound PFs stayed at 6. In production
  this had accumulated ~126 leaked majors, filling the 384–511 dynamic range
  and failing all subsequent probes with -16.
• After (this patch): baseline 6 majors, holding steady at 6 across four
  remove/rescan cycles, with all 6 PFs re-binding each time. Normal
  load/unload/queue setup unaffected. Builds clean under -Werror.