Re: [PATCH net] net/dibs: Fix UAF of dmb_clientid_arr after dibs_dev_del()

This is an AI-generated review of your patch. The human sending this
email has considered the AI review valid, or at least plausible.
---
net/dibs: Fix UAF of dmb_clientid_arr after dibs_dev_del()

This moves the kfree() of dibs->dmb_clientid_arr out of dibs_dev_del() and
into dibs_dev_release(), so the array survives until the last device
reference is dropped.  The error paths in dibs_dev_add() are converted from
gotos to direct returns and no longer free the array.

[Severity: High]
This isn't a bug introduced by this patch, but does the new ownership rule
here leave a double free of dibs->dmb_clientid_arr in the loopback probe
path?

After this change dibs_dev_release() is the sole owner of the buffer, while
dibs_lo_dev_probe() in drivers/dibs/dibs_loopback.c still frees it itself
when dibs_dev_add() fails:

drivers/dibs/dibs_loopback.c:dibs_lo_dev_probe() {
	...
	ret = dibs_dev_add(dibs);
	if (ret)
		goto err_reg;
	lo_dev = ldev;
	return 0;

err_reg:
	kfree(dibs->dmb_clientid_arr);
	/* pairs with dibs_dev_alloc() */
	put_device(&dibs->dev);
	kfree(ldev);
	...
}

dibs_dev_alloc() leaves the refcount at 1 via device_initialize(), and the
get_device()/put_device() pair inside device_add() nets to zero, so the
put_device() above drops the last reference and runs dibs_dev_release(),
which frees the same pointer a second time.

In the baseline this was already a double free (dibs_dev_add() freed the
array on its free_client_arr path and dibs_lo_dev_probe() freed it again),
so the defect predates the patch, but since this patch is the one that
redefines who owns dmb_clientid_arr, should the kfree() at err_reg in
dibs_lo_dev_probe() be dropped in the same change?

For comparison, drivers/s390/net/ism_drv.c:ism_probe() already relies only
on put_device() and has no explicit kfree, so the two callers of the same
exported API currently disagree about ownership:

drivers/s390/net/ism_drv.c:ism_probe() {
	...
err_ism:
	ism_dev_exit(ism);
err_dibs:
	/* pairs with dibs_dev_alloc() */
	put_device(&dibs->dev);
	...
}

Since this is tagged for stable, would it be worth also stating the new
caller contract in the changelog, i.e. that callers must only call
put_device() when dibs_dev_add() fails?

[Severity: High]
These are the two failure returns that reach the loopback err_reg path
described above: device_add() returning an error (-ENOMEM, or a
kobject_add()/duplicate-name failure) and sysfs_create_group() returning
-ENOMEM.  In both cases dmb_clientid_arr is still allocated and not NULLed,
so the caller-side kfree() and dibs_dev_release() both free it.  The
kzalloc-failure case is fine because the pointer is still NULL.

[Severity: Medium]
This is a pre-existing issue and not introduced by the patch, but the
changelog reasons about the placement of the allocation:

    Note that allocating in dibs_dev_add() instead of dibs_dev_alloc() is ok
    for now, because no dmbs can be registered before dibs_dev_add().

Does that reasoning also cover dibs->lock, which is initialized in the same
place?

drivers/dibs/dibs_main.c:dibs_dev_add() {
	max_dmbs = dibs->ops->max_dmbs();
	spin_lock_init(&dibs->lock);
	dibs->dmb_clientid_arr = kzalloc(max_dmbs, GFP_KERNEL);
	...
}

ism_probe() enables interrupts before that, via ism_dev_init() doing
request_irq(pci_irq_vector(pdev, 0), ism_handle_irq, ...) plus
register_sba()/register_ieq(), and only calls dibs_dev_add() afterwards.  In
that window dibs comes from kzalloc, so dibs->lock has never been
initialized and dmb_clientid_arr is NULL, yet the handler takes the lock and
indexes the array before any check:

drivers/s390/net/ism_drv.c:ism_handle_irq() {
	spin_lock(&dibs->lock);
	ism->sba->s = 0;
	...
		client_id = dibs->dmb_clientid_arr[bit];
		if (unlikely(client_id == NO_DIBS_CLIENT ||
			     !dibs->subs[client_id]))
			continue;
	...
}

The comment in ism_probe() says interrupts and events are ignored before
dibs_dev_add() adds the device to any clients, but that check happens on
dibs->subs[], after the lock acquisition and after the dmb_clientid_arr
dereference.  With CONFIG_DEBUG_SPINLOCK or lockdep this window would give
"spinlock bad magic" or "trying to register non-static key" from hardirq
context, and spin_lock_init() in dibs_dev_add() re-zeroing the lock while
the handler may hold it would void mutual exclusion.

Would moving spin_lock_init() and the allocation into dibs_dev_alloc() close
this window?

[ ... ]
-- 
pw-bot: cr