Re: [PATCH] mm/memory_hotplug: Fix remove_memory() lockdep splat

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Fri, Jan 10, 2020 at 8:54 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> On 10.01.20 17:42, Dan Williams wrote:
> > On Fri, Jan 10, 2020 at 1:10 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> >>
> >> On 10.01.20 05:30, Dan Williams wrote:
> >>> The daxctl unit test for the dax_kmem driver currently triggers the
> >>> lockdep splat below. It results from the fact that
> >>> remove_memory_block_devices() is invoked under the mem_hotplug_lock()
> >>> causing lockdep entanglements with cpu_hotplug_lock().
> >>>
> >>> The mem_hotplug_lock() is not needed to synchronize the memory block
> >>> device sysfs interface vs the page online state, that is already handled
> >>> by lock_device_hotplug(). Specifically lock_device_hotplug()
> >>> is sufficient to allow try_remove_memory() to check the offline
> >>> state of the memblocks and be assured that subsequent online attempts
> >>> will be blocked. The device_online() path checks mem->section_count
> >>> before allowing any state manipulations and mem->section_count is
> >>> cleared in remove_memory_block_devices().
> >>>
> >>> The add_memory() path does create memblock devices under the lock, but
> >>> there is no lockdep report on that path, so it is left alone for now.
> >>>
> >>> This change is only possible thanks to the recent change that refactored
> >>> memory block device removal out of arch_remove_memory() (commit
> >>> 4c4b7f9ba948 mm/memory_hotplug: remove memory block devices before
> >>> arch_remove_memory()).
> >>>
> >>>     ======================================================
> >>>     WARNING: possible circular locking dependency detected
> >>>     5.5.0-rc3+ #230 Tainted: G           OE
> >>>     ------------------------------------------------------
> >>>     lt-daxctl/6459 is trying to acquire lock:
> >>>     ffff99c7f0003510 (kn->count#241){++++}, at: kernfs_remove_by_name_ns+0x41/0x80
> >>>
> >>>     but task is already holding lock:
> >>>     ffffffffa76a5450 (mem_hotplug_lock.rw_sem){++++}, at: percpu_down_write+0x20/0xe0
> >>>
> >>>     which lock already depends on the new lock.
> >>>
> >>>
> >>>     the existing dependency chain (in reverse order) is:
> >>>
> >>>     -> #2 (mem_hotplug_lock.rw_sem){++++}:
> >>>            __lock_acquire+0x39c/0x790
> >>>            lock_acquire+0xa2/0x1b0
> >>>            get_online_mems+0x3e/0xb0
> >>>            kmem_cache_create_usercopy+0x2e/0x260
> >>>            kmem_cache_create+0x12/0x20
> >>>            ptlock_cache_init+0x20/0x28
> >>>            start_kernel+0x243/0x547
> >>>            secondary_startup_64+0xb6/0xc0
> >>>
> >>>     -> #1 (cpu_hotplug_lock.rw_sem){++++}:
> >>>            __lock_acquire+0x39c/0x790
> >>>            lock_acquire+0xa2/0x1b0
> >>>            cpus_read_lock+0x3e/0xb0
> >>>            online_pages+0x37/0x300
> >>>            memory_subsys_online+0x17d/0x1c0
> >>>            device_online+0x60/0x80
> >>>            state_store+0x65/0xd0
> >>>            kernfs_fop_write+0xcf/0x1c0
> >>>            vfs_write+0xdb/0x1d0
> >>>            ksys_write+0x65/0xe0
> >>>            do_syscall_64+0x5c/0xa0
> >>>            entry_SYSCALL_64_after_hwframe+0x49/0xbe
> >>>
> >>>     -> #0 (kn->count#241){++++}:
> >>>            check_prev_add+0x98/0xa40
> >>>            validate_chain+0x576/0x860
> >>>            __lock_acquire+0x39c/0x790
> >>>            lock_acquire+0xa2/0x1b0
> >>>            __kernfs_remove+0x25f/0x2e0
> >>>            kernfs_remove_by_name_ns+0x41/0x80
> >>>            remove_files.isra.0+0x30/0x70
> >>>            sysfs_remove_group+0x3d/0x80
> >>>            sysfs_remove_groups+0x29/0x40
> >>>            device_remove_attrs+0x39/0x70
> >>>            device_del+0x16a/0x3f0
> >>>            device_unregister+0x16/0x60
> >>>            remove_memory_block_devices+0x82/0xb0
> >>>            try_remove_memory+0xb5/0x130
> >>>            remove_memory+0x26/0x40
> >>>            dev_dax_kmem_remove+0x44/0x6a [kmem]
> >>>            device_release_driver_internal+0xe4/0x1c0
> >>>            unbind_store+0xef/0x120
> >>>            kernfs_fop_write+0xcf/0x1c0
> >>>            vfs_write+0xdb/0x1d0
> >>>            ksys_write+0x65/0xe0
> >>>            do_syscall_64+0x5c/0xa0
> >>>            entry_SYSCALL_64_after_hwframe+0x49/0xbe
> >>>
> >>>     other info that might help us debug this:
> >>>
> >>>     Chain exists of:
> >>>       kn->count#241 --> cpu_hotplug_lock.rw_sem --> mem_hotplug_lock.rw_sem
> >>>
> >>>      Possible unsafe locking scenario:
> >>>
> >>>            CPU0                    CPU1
> >>>            ----                    ----
> >>>       lock(mem_hotplug_lock.rw_sem);
> >>>                                    lock(cpu_hotplug_lock.rw_sem);
> >>>                                    lock(mem_hotplug_lock.rw_sem);
> >>>       lock(kn->count#241);
> >>>
> >>>      *** DEADLOCK ***
> >>>
> >>> No fixes tag as this seems to have been a long standing issue that
> >>> likely predated the addition of kernfs lockdep annotations.
> >>>
> >>> Cc: <stable@xxxxxxxxxxxxxxx>
> >>> Cc: Vishal Verma <vishal.l.verma@xxxxxxxxx>
> >>> Cc: David Hildenbrand <david@xxxxxxxxxx>
> >>> Cc: Pavel Tatashin <pasha.tatashin@xxxxxxxxxx>
> >>> Cc: Michal Hocko <mhocko@xxxxxxxx>
> >>> Cc: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
> >>> Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> >>> ---
> >>>  mm/memory_hotplug.c |   12 +++++++++---
> >>>  1 file changed, 9 insertions(+), 3 deletions(-)
> >>>
> >>> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
> >>> index 55ac23ef11c1..a4e7dadded08 100644
> >>> --- a/mm/memory_hotplug.c
> >>> +++ b/mm/memory_hotplug.c
> >>> @@ -1763,8 +1763,6 @@ static int __ref try_remove_memory(int nid, u64 start, u64 size)
> >>>
> >>>       BUG_ON(check_hotplug_memory_range(start, size));
> >>>
> >>> -     mem_hotplug_begin();
> >>> -
> >>>       /*
> >>>        * All memory blocks must be offlined before removing memory.  Check
> >>>        * whether all memory blocks in question are offline and return error
> >>> @@ -1777,9 +1775,17 @@ static int __ref try_remove_memory(int nid, u64 start, u64 size)
> >>>       /* remove memmap entry */
> >>>       firmware_map_remove(start, start + size, "System RAM");
> >>>
> >>> -     /* remove memory block devices before removing memory */
> >>> +     /*
> >>> +      * Remove memory block devices before removing memory, and do
> >>> +      * not hold the mem_hotplug_lock() over kobject removal
> >>> +      * operations. lock_device_hotplug() keeps the
> >>> +      * check_memblock_offlined_cb result valid until the entire
> >>> +      * removal process is complete.
> >>> +      */
> >>
> >> Maybe shorten that to
> >>
> >> /*
> >>  * Remove memory block devices before removing memory. Protected
> >>  * by the device_hotplug_lock only.
> >>  */
> >
> > Why make someone dig for the reasons this lock is sufficient?
>
> I think 5 LOC of comment are too much for something that is documented
> e.g., in Documentation/core-api/memory-hotplug.rst ("Locking
> Internals"). But whatever you prefer.

Sure, lets beef up that doc to clarify this case and refer to it.

>
> >
> >>
> >> AFAIK, the device hotplug lock is sufficient here. The memory hotplug
> >> lock / cpu hotplug lock is only needed when calling into arch code
> >> (especially for PPC). We hold both locks when onlining/offlining memory.
> >>
> >>>       remove_memory_block_devices(start, size);
> >>>
> >>> +     mem_hotplug_begin();
> >>> +
> >>>       arch_remove_memory(nid, start, size, NULL);
> >>>       memblock_free(start, size);
> >>>       memblock_remove(start, size);
> >>>
> >>
> >> I'd suggest to do the same in the adding part right away (if easily
> >> possible) to make it clearer.
> >
> > Let's let this fix percolate upstream for a bit to make sure there was
> > no protection the mem_hotplug_begin() was inadvertently providing.
>
> Yeah, why not.
>
> >
> >> I properly documented the semantics of
> >> add_memory_block_devices()/remove_memory_block_devices() already (that
> >> they need the device hotplug lock).
> >
> > I see that, but I prefer lockdep_assert_held() in the code rather than
> > comments. I'll send a patch to fix that up.
>
> That won't work as early boot code from ACPI won't hold it while it adds
> memory. And we decided (especially Michal :) ) to keep it like that.

So then the comment is actively misleading for that case. I would
expect an explicit _unlocked path for that case with a comment about
why it's special. Is there already a comment to that effect somewhere?




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux