On Mon, Jan 13, 2020 at 09:43:23PM +0800, Zhiqiang Liu wrote: > In brd_init func, rd_nr num of brd_device are firstly allocated > and add in brd_devices, then brd_devices are traversed to add each > brd_device by calling add_disk func. When allocating brd_device, > the disk->first_minor is set to i * max_part, if rd_nr * max_part > is larger than MINORMASK, two different brd_device may have the same > devt, then only one of them can be successfully added. It is just because disk->first_minor is >= 0x100000, then same dev_t can be allocated in blk_alloc_devt(). MKDEV(disk->major, disk->first_minor + part->partno) But block layer does support extended dynamic devt allocation, and brd sets flag of GENHD_FL_EXT_DEVT too. So I think the correct fix is to fallback to extended dynamic allocation when running out of consecutive minor space. How about the following approach? And of course, ext devt allocation may fail too, but that is another generic un-solved issue: error handling isn't done for adding disk. diff --git a/drivers/block/brd.c b/drivers/block/brd.c index a8730cc4db10..9aa7ce7c9abf 100644 --- a/drivers/block/brd.c +++ b/drivers/block/brd.c @@ -398,7 +398,16 @@ static struct brd_device *brd_alloc(int i) if (!disk) goto out_free_queue; disk->major = RAMDISK_MAJOR; - disk->first_minor = i * max_part; + + /* + * Clear .minors when running out of consecutive minor space since + * GENHD_FL_EXT_DEVT is set, and we can allocate from extended devt + */ + if ((i * disk->minors) & ~MINORMASK) + disk->minors = 0; + else + disk->first_minor = i * disk->minors; + disk->fops = &brd_fops; disk->private_data = brd; disk->flags = GENHD_FL_EXT_DEVT; Thanks, Ming
