Re: [PATCH] mpt2sas: mpt3sas: Fix memory corruption during initialization

[James Bottomley]

On Fri, 2015-04-10 at 00:14 -0700, Calvin Owens wrote:
> While _scsih_probe_sas() is initializing devices, the hardware can trigger a
> MPI2_EVENT_SAS_TOPO_RC_TARG_NOT_RESPONDING interrupt.
> 
> The handler for TARG_NOT_RESPONDING calls _scsih_device_remove_by_handle(),
> which deletes the device in question from either ioc->sas_device_list or
> ioc->sas_device_init_list. Since _scsih_probe_sas() uses no exclusion when
> iterating over ioc->sas_device_init_list, this results in a use-after-free
> in _scsih_probe_sas(), and also corrupts the list:
> 
>   mpt2sas1: removing handle(0x0020), sas_addr(0x5f80f418573360e0)
>   mpt2sas1: log_info(0x31111000): originator(PL), code(0x11), sub_code(0x1000)
>   ------------[ cut here ]------------
>   WARNING: at lib/list_debug.c:56 __list_del_entry+0xc3/0xd0()
>   list_del corruption, ffff88240012fa00->prev is LIST_POISON2 (dead000000200200)
>   <snip>
>   Workqueue: events work_for_cpu_fn
>    ffffffff810c4f17 ffff881214825b38 0000000000000009 ffff881214825ae8
>    ffffffff8169b61e ffff881214825b28 ffffffff8104a990 0000000000000002
>    ffff88240012f900 ffff88240012fa00 ffff881217595af8 0000000000000282
>   Call Trace:
>    [<ffffffff810c4f17>] ? print_modules+0xd7/0x120
>    [<ffffffff8169b61e>] dump_stack+0x19/0x1b
>    [<ffffffff8104a990>] warn_slowpath_common+0x70/0xa0
>    [<ffffffff8104aa76>] warn_slowpath_fmt+0x46/0x50
>    [<ffffffff816a2234>] ? _raw_spin_lock_irqsave+0x84/0xa0
>    [<ffffffffa0010e8e>] ? _scsih_probe_sas+0x8e/0x110 [mpt2sas]
>    [<ffffffff8132a5a3>] __list_del_entry+0xc3/0xd0
>    [<ffffffffa0010e99>] _scsih_probe_sas+0x99/0x110 [mpt2sas]
>    [<ffffffffa0011d5f>] _scsih_scan_finished+0x19f/0x2c0 [mpt2sas]
>    [<ffffffff81429d67>] do_scsi_scan_host+0x77/0xa0
>    [<ffffffff81429f20>] scsi_scan_host+0x190/0x1c0
>    [<ffffffffa0011402>] _scsih_probe+0x452/0x640 [mpt2sas]
>    [<ffffffff813444eb>] local_pci_probe+0x4b/0x80
>    [<ffffffff8106b848>] work_for_cpu_fn+0x18/0x30
>    [<ffffffff81070012>] process_one_work+0x212/0x6e0
>    [<ffffffff8106ffa6>] ? process_one_work+0x1a6/0x6e0
>    [<ffffffff8108ed1f>] ? local_clock+0x4f/0x60
>    [<ffffffff8107050c>] process_scheduled_works+0x2c/0x40
>    [<ffffffff81070ae2>] worker_thread+0x262/0x370
>    [<ffffffff81070880>] ? rescuer_thread+0x360/0x360
>    [<ffffffff81078f3b>] kthread+0xdb/0xe0
>    [<ffffffff810b5e8d>] ? trace_hardirqs_on+0xd/0x10
>    [<ffffffff81078e60>] ? kthread_create_on_node+0x140/0x140
>    [<ffffffff816ac01c>] ret_from_fork+0x7c/0xb0
>    [<ffffffff81078e60>] ? kthread_create_on_node+0x140/0x140
>   ---[ end trace 41352a0bd2d0d61b ]---
> 
> This either results in an immediate panic, or corrupts random memory and
> causes nasty problems later in the box's uptime.
> 
> This patch splices the discovered devices out of the global list while
> holding the lock, since _scsih_probe_sas() always removes them from that
> global list anyway (either deleting them if initialization fails, or
> moving them onto ioc->sas_device_list if it succeeds). The interrupt that
> caused this bug will no longer cause the device to be removed during
> initialization, since it won't exist on the global lists, but
> _scsih_probe_sas() will remove it anyway when it fails to initialize.

Hopefully the avago team will curate this, but just in case they don't,
the correct list to make sure it gets the attention of storage people
should be

linux-scsi@xxxxxxxxxxxxxxx

James