Re: [PATCH 1/2] qla2xxx: Fix a recently introduced memory leak

From: Bart Van Assche <Bart.VanAssche@sandisk.com>
To: "hch@lst.de" <hch@lst.de>
Cc: "linux-scsi@vger.kernel.org" <linux-scsi@vger.kernel.org>
Subject: Re: [PATCH 1/2] qla2xxx: Fix a recently introduced memory leak
Date: Sun, 29 Jan 2017 05:17:38 +0000	[thread overview]
Message-ID: <1485667034.15853.1.camel@sandisk.com> (raw)
In-Reply-To: <20170126143614.GA19849@lst.de>

On Thu, 2017-01-26 at 15:36 +0100, hch@lst.de wrote:
> On Wed, Jan 25, 2017 at 03:47:20PM +0000, Bart Van Assche wrote:
> > =============================================================================
> > BUG kmalloc-16 (Not tainted): Redzone overwritten
> > -----------------------------------------------------------------------------
> > 
> > Disabling lock debugging due to kernel taint
> > INFO: 0xffff880030bacc78-0xffff880030bacc7f. First byte 0xf instead of 0xcc
> > INFO: Allocated in irq_create_affinity_masks+0x5f/0x260 age=0 cpu=3 pid=812
> > 	___slab_alloc.constprop.79+0x482/0x4f0
> > 	__slab_alloc.isra.75.constprop.78+0x55/0xa0
> > 	__kmalloc+0x27c/0x310
> > 	irq_create_affinity_masks+0x5f/0x260
> 
> This is the normal affinity mask allocation.

(reduced CC-list again)

Hello Christoph,

It seems like irq_create_affinity_masks() wrote past the bounds of the masks array
it allocated. After I had added the following debug code in irq_create_affinity_masks():

        WARN_ON_ONCE(affv <= 0);
        pr_err("%s: affd = { .pre = %d, .post = %d }, nvecs = %d\n",
                __func__, affd->pre_vectors, affd->post_vectors, nvecs);

The following output appeared:

WARNING: CPU: 0 PID: 814 at kernel/irq/affinity.c:69 irq_create_affinity_masks+0x2cd/0x2f0
Call Trace:
 dump_stack+0x85/0xc2
 __warn+0xcb/0xf0
 warn_slowpath_null+0x1d/0x20
 irq_create_affinity_masks+0x2cd/0x2f0
 __pci_enable_msix+0x314/0x4c0
 pci_alloc_irq_vectors_affinity+0xb7/0x140
 qla2x00_request_irqs+0xa6/0x6d0 [qla2xxx]
 qla2x00_probe_one+0xc2e/0x25f0 [qla2xxx]
 pci_device_probe+0x8a/0xf0
 driver_probe_device+0x1f5/0x450
 __driver_attach+0xe3/0xf0
 bus_for_each_dev+0x66/0xa0
 driver_attach+0x1e/0x20
 bus_add_driver+0x200/0x270
 driver_register+0x60/0xe0
 __pci_register_driver+0x5d/0x60
 qla2x00_module_init+0x1c9/0x217 [qla2xxx]
 do_one_initcall+0x44/0x180
 do_init_module+0x5f/0x1f9
 load_module+0x2582/0x2a00
 SYSC_finit_module+0xbc/0xf0
 SyS_finit_module+0xe/0x10
 entry_SYSCALL_64_fastpath+0x23/0xc6

irq_create_affinity_masks: affd = { .pre = 2, .post = 0 }, nvecs = 2

affd comes from the qla2xxx driver: struct irq_affinity desc = { .pre_vectors =
QLA_BASE_VECTORS }. Shouldn't irq_calc_affinity_vectors() guarantee that it
returns a value that is strictly greater than affd->pre_vectors + affd->post_vectors
instead of greater than or equal to affd->pre_vectors + affd->post_vectors?

Thanks,

Bart.