On Thu, 20 Sep 2018 17:19:34 +0200
Eric Farman <farman@xxxxxxxxxxxxx> wrote:
> If I attach a vfio-ccw device to my guest, I get the following warning
> on the host when the host kernel is CONFIG_HARDENED_USERCOPY=y
Maybe I should try building with that on my systems as well :)
>
> [250757.595325] Bad or missing usercopy whitelist? Kernel memory overwrite attempt detected to SLUB object 'dma-kmalloc-512' (offset 64, size 124)!
> [250757.595365] WARNING: CPU: 2 PID: 10958 at mm/usercopy.c:81 usercopy_warn+0xac/0xd8
> [250757.595369] Modules linked in: kvm vhost_net vhost tap xt_CHECKSUM iptable_mangle ipt_MASQUERADE iptable_nat nf_nat_ipv4 nf_nat nf_conntrack_ipv4 nf_defrag_ipv4 xt_conntrack nf_conntrack libcrc32c devlink tun bridge stp llc ebtable_filter ebtables ip6table_filter ip6_tables sunrpc dm_multipath s390_trng crc32_vx_s390 ghash_s390 prng aes_s390 des_s390 des_generic sha512_s390 sha1_s390 eadm_sch tape_3590 tape tape_class qeth_l2 qeth ccwgroup vfio_ccw vfio_mdev zcrypt_cex4 mdev vfio_iommu_type1 zcrypt vfio sha256_s390 sha_common zfcp scsi_transport_fc qdio dasd_eckd_mod dasd_mod
> [250757.595424] CPU: 2 PID: 10958 Comm: CPU 2/KVM Not tainted 4.18.0-derp #2
> [250757.595426] Hardware name: IBM 3906 M05 780 (LPAR)
> ...snip regs...
> [250757.595523] Call Trace:
> [250757.595529] ([<0000000000349210>] usercopy_warn+0xa8/0xd8)
> [250757.595535] [<000000000032daaa>] __check_heap_object+0xfa/0x160
> [250757.595540] [<0000000000349396>] __check_object_size+0x156/0x1d0
> [250757.595547] [<000003ff80332d04>] vfio_ccw_mdev_write+0x74/0x148 [vfio_ccw]
> [250757.595552] [<000000000034ed12>] __vfs_write+0x3a/0x188
> [250757.595556] [<000000000034f040>] vfs_write+0xa8/0x1b8
> [250757.595559] [<000000000034f4e6>] ksys_pwrite64+0x86/0xc0
> [250757.595568] [<00000000008959a0>] system_call+0xdc/0x2b0
> [250757.595570] Last Breaking-Event-Address:
> [250757.595573] [<0000000000349210>] usercopy_warn+0xa8/0xd8
>
> While vfio_ccw_mdev_{write|read} validates that the input position/count
> does not run over the ccw_io_region struct, the usercopy code that does
> copy_{to|from}_user doesn't necessarily know this. It sees the variable
> length and gets worried that it's affecting a normal kmalloc'd struct,
> and generates the above warning.
>
> Adjust how the vfio_ccw_struct is alloc'd, with a whitelist for the
> ccw_io_region within it, to remove this warning. The boundary checking
> will continue to do its thing.
>
> Signed-off-by: Eric Farman <farman@xxxxxxxxxxxxx>
> ---
> drivers/s390/cio/vfio_ccw_drv.c | 21 ++++++++++++++++++---
> 1 file changed, 18 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/s390/cio/vfio_ccw_drv.c b/drivers/s390/cio/vfio_ccw_drv.c
> index 770fa9cfc310..8191adbf3490 100644
> --- a/drivers/s390/cio/vfio_ccw_drv.c
> +++ b/drivers/s390/cio/vfio_ccw_drv.c
> @@ -22,6 +22,10 @@
> #include "vfio_ccw_private.h"
>
> struct workqueue_struct *vfio_ccw_work_q;
> +struct kmem_cache *vfio_private_cache;
> +
> +#define IOREGION_OFFSET offsetof(struct vfio_ccw_private, io_region)
> +#define IOREGION_SIZE sizeof_field(struct vfio_ccw_private, io_region)
>
> /*
> * Helpers
> @@ -111,7 +115,7 @@ static int vfio_ccw_sch_probe(struct subchannel *sch)
> return -ENODEV;
> }
>
> - private = kzalloc(sizeof(*private), GFP_KERNEL | GFP_DMA);
> + private = kmem_cache_zalloc(vfio_private_cache, GFP_KERNEL | GFP_DMA);
> if (!private)
> return -ENOMEM;
> private->sch = sch;
> @@ -139,7 +143,7 @@ static int vfio_ccw_sch_probe(struct subchannel *sch)
> cio_disable_subchannel(sch);
> out_free:
> dev_set_drvdata(&sch->dev, NULL);
> - kfree(private);
> + kmem_cache_free(vfio_private_cache, private);
> return ret;
> }
>
> @@ -153,7 +157,7 @@ static int vfio_ccw_sch_remove(struct subchannel *sch)
>
> dev_set_drvdata(&sch->dev, NULL);
>
> - kfree(private);
> + kmem_cache_free(vfio_private_cache, private);
>
> return 0;
> }
> @@ -232,10 +236,20 @@ static int __init vfio_ccw_sch_init(void)
> if (!vfio_ccw_work_q)
> return -ENOMEM;
>
> + vfio_private_cache = kmem_cache_create_usercopy("vfio_ccw_private",
> + sizeof(struct vfio_ccw_private),
> + 0, SLAB_ACCOUNT, IOREGION_OFFSET,
> + IOREGION_SIZE, NULL);
That should work fine, but I'm currently (...) trying to add more
regions (for example, for halt/clear handling) and I'm wondering
whether we should change how we allocate our I/O regions, for example
using a dedicated region that is pointed to by the private structure.
Thoughts?
> + if (!vfio_private_cache) {
> + destroy_workqueue(vfio_ccw_work_q);
> + return -ENOMEM;
> + }
> +
> isc_register(VFIO_CCW_ISC);
> ret = css_driver_register(&vfio_ccw_sch_driver);
> if (ret) {
> isc_unregister(VFIO_CCW_ISC);
> + kmem_cache_destroy(vfio_private_cache);
> destroy_workqueue(vfio_ccw_work_q);
> }
>
> @@ -246,6 +260,7 @@ static void __exit vfio_ccw_sch_exit(void)
> {
> css_driver_unregister(&vfio_ccw_sch_driver);
> isc_unregister(VFIO_CCW_ISC);
> + kmem_cache_destroy(vfio_private_cache);
> destroy_workqueue(vfio_ccw_work_q);
> }
> module_init(vfio_ccw_sch_init);
