| Message ID | 20180918114822.21926-1-david@redhat.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | mm: online/offline_pages called w.o. mem_hotplug_lock | expand |
On Tue, Sep 18, 2018 at 01:48:16PM +0200, David Hildenbrand wrote: > Reading through the code and studying how mem_hotplug_lock is to be used, > I noticed that there are two places where we can end up calling > device_online()/device_offline() - online_pages()/offline_pages() without > the mem_hotplug_lock. And there are other places where we call > device_online()/device_offline() without the device_hotplug_lock. > > While e.g. > echo "online" > /sys/devices/system/memory/memory9/state > is fine, e.g. > echo 1 > /sys/devices/system/memory/memory9/online > Will not take the mem_hotplug_lock. However the device_lock() and > device_hotplug_lock. > > E.g. via memory_probe_store(), we can end up calling > add_memory()->online_pages() without the device_hotplug_lock. So we can > have concurrent callers in online_pages(). We e.g. touch in online_pages() > basically unprotected zone->present_pages then. > > Looks like there is a longer history to that (see Patch #2 for details), > and fixing it to work the way it was intended is not really possible. We > would e.g. have to take the mem_hotplug_lock in device/base/core.c, which > sounds wrong. > > Summary: We had a lock inversion on mem_hotplug_lock and device_lock(). > More details can be found in patch 3 and patch 6. > > I propose the general rules (documentation added in patch 6): > > 1. add_memory/add_memory_resource() must only be called with > device_hotplug_lock. > 2. remove_memory() must only be called with device_hotplug_lock. This is > already documented and holds for all callers. > 3. device_online()/device_offline() must only be called with > device_hotplug_lock. This is already documented and true for now in core > code. Other callers (related to memory hotplug) have to be fixed up. > 4. mem_hotplug_lock is taken inside of add_memory/remove_memory/ > online_pages/offline_pages. > > To me, this looks way cleaner than what we have right now (and easier to > verify). And looking at the documentation of remove_memory, using > lock_device_hotplug also for add_memory() feels natural. > That seems reasonable, but also implies that device_online() would hold back add/remove memory, could you please also document what mode read/write the locks need to be held? For example can the device_hotplug_lock be held in read mode while add/remove memory via (mem_hotplug_lock) is held in write mode? Balbir Singh.
Am 19.09.18 um 03:22 schrieb Balbir Singh: > On Tue, Sep 18, 2018 at 01:48:16PM +0200, David Hildenbrand wrote: >> Reading through the code and studying how mem_hotplug_lock is to be used, >> I noticed that there are two places where we can end up calling >> device_online()/device_offline() - online_pages()/offline_pages() without >> the mem_hotplug_lock. And there are other places where we call >> device_online()/device_offline() without the device_hotplug_lock. >> >> While e.g. >> echo "online" > /sys/devices/system/memory/memory9/state >> is fine, e.g. >> echo 1 > /sys/devices/system/memory/memory9/online >> Will not take the mem_hotplug_lock. However the device_lock() and >> device_hotplug_lock. >> >> E.g. via memory_probe_store(), we can end up calling >> add_memory()->online_pages() without the device_hotplug_lock. So we can >> have concurrent callers in online_pages(). We e.g. touch in online_pages() >> basically unprotected zone->present_pages then. >> >> Looks like there is a longer history to that (see Patch #2 for details), >> and fixing it to work the way it was intended is not really possible. We >> would e.g. have to take the mem_hotplug_lock in device/base/core.c, which >> sounds wrong. >> >> Summary: We had a lock inversion on mem_hotplug_lock and device_lock(). >> More details can be found in patch 3 and patch 6. >> >> I propose the general rules (documentation added in patch 6): >> >> 1. add_memory/add_memory_resource() must only be called with >> device_hotplug_lock. >> 2. remove_memory() must only be called with device_hotplug_lock. This is >> already documented and holds for all callers. >> 3. device_online()/device_offline() must only be called with >> device_hotplug_lock. This is already documented and true for now in core >> code. Other callers (related to memory hotplug) have to be fixed up. >> 4. mem_hotplug_lock is taken inside of add_memory/remove_memory/ >> online_pages/offline_pages. >> >> To me, this looks way cleaner than what we have right now (and easier to >> verify). And looking at the documentation of remove_memory, using >> lock_device_hotplug also for add_memory() feels natural. >> > > That seems reasonable, but also implies that device_online() would hold > back add/remove memory, could you please also document what mode > read/write the locks need to be held? For example can the device_hotplug_lock > be held in read mode while add/remove memory via (mem_hotplug_lock) is held > in write mode? device_hotplug_lock is an ordinary mutex. So no option there. Only mem_hotplug_lock is a per CPU RW mutex. And as of now it only exists to not require get_online_mems()/put_online_mems() to take the device_hotplug_lock. Which is perfectly valid, because these users only care about memory (not any other devices) not suddenly vanish. And that RW lock makes things fast. Any modifications (online/offline/add/remove) require the mem_hotplug_lock in write. I can add some more details to documentation in patch #6. "... we should always hold the mem_hotplug_lock (via mem_hotplug_begin/mem_hotplug_done) in write mode to serialize memory hotplug" ..." "In addition, mem_hotplug_lock (in contrast to device_hotplug_lock) in read mode allows for a quite efficient get_online_mems/put_online_mems implementation, so code accessing memory can protect from that memory vanishing." Would that work for you? Thanks! > > Balbir Singh. > >
On Wed, Sep 19, 2018 at 09:35:07AM +0200, David Hildenbrand wrote: > Am 19.09.18 um 03:22 schrieb Balbir Singh: > > On Tue, Sep 18, 2018 at 01:48:16PM +0200, David Hildenbrand wrote: > >> Reading through the code and studying how mem_hotplug_lock is to be used, > >> I noticed that there are two places where we can end up calling > >> device_online()/device_offline() - online_pages()/offline_pages() without > >> the mem_hotplug_lock. And there are other places where we call > >> device_online()/device_offline() without the device_hotplug_lock. > >> > >> While e.g. > >> echo "online" > /sys/devices/system/memory/memory9/state > >> is fine, e.g. > >> echo 1 > /sys/devices/system/memory/memory9/online > >> Will not take the mem_hotplug_lock. However the device_lock() and > >> device_hotplug_lock. > >> > >> E.g. via memory_probe_store(), we can end up calling > >> add_memory()->online_pages() without the device_hotplug_lock. So we can > >> have concurrent callers in online_pages(). We e.g. touch in online_pages() > >> basically unprotected zone->present_pages then. > >> > >> Looks like there is a longer history to that (see Patch #2 for details), > >> and fixing it to work the way it was intended is not really possible. We > >> would e.g. have to take the mem_hotplug_lock in device/base/core.c, which > >> sounds wrong. > >> > >> Summary: We had a lock inversion on mem_hotplug_lock and device_lock(). > >> More details can be found in patch 3 and patch 6. > >> > >> I propose the general rules (documentation added in patch 6): > >> > >> 1. add_memory/add_memory_resource() must only be called with > >> device_hotplug_lock. > >> 2. remove_memory() must only be called with device_hotplug_lock. This is > >> already documented and holds for all callers. > >> 3. device_online()/device_offline() must only be called with > >> device_hotplug_lock. This is already documented and true for now in core > >> code. Other callers (related to memory hotplug) have to be fixed up. > >> 4. mem_hotplug_lock is taken inside of add_memory/remove_memory/ > >> online_pages/offline_pages. > >> > >> To me, this looks way cleaner than what we have right now (and easier to > >> verify). And looking at the documentation of remove_memory, using > >> lock_device_hotplug also for add_memory() feels natural. > >> > > > > That seems reasonable, but also implies that device_online() would hold > > back add/remove memory, could you please also document what mode > > read/write the locks need to be held? For example can the device_hotplug_lock > > be held in read mode while add/remove memory via (mem_hotplug_lock) is held > > in write mode? > > device_hotplug_lock is an ordinary mutex. So no option there. > > Only mem_hotplug_lock is a per CPU RW mutex. And as of now it only > exists to not require get_online_mems()/put_online_mems() to take the > device_hotplug_lock. Which is perfectly valid, because these users only > care about memory (not any other devices) not suddenly vanish. And that > RW lock makes things fast. > > Any modifications (online/offline/add/remove) require the > mem_hotplug_lock in write. > > I can add some more details to documentation in patch #6. > > "... we should always hold the mem_hotplug_lock (via > mem_hotplug_begin/mem_hotplug_done) in write mode to serialize memory > hotplug" ..." > > "In addition, mem_hotplug_lock (in contrast to device_hotplug_lock) in > read mode allows for a quite efficient get_online_mems/put_online_mems > implementation, so code accessing memory can protect from that memory > vanishing." > > Would that work for you? Yes, Thanks Balbir Singh.
Reading through the code and studying how mem_hotplug_lock is to be used, I noticed that there are two places where we can end up calling device_online()/device_offline() - online_pages()/offline_pages() without the mem_hotplug_lock. And there are other places where we call device_online()/device_offline() without the device_hotplug_lock. While e.g. echo "online" > /sys/devices/system/memory/memory9/state is fine, e.g. echo 1 > /sys/devices/system/memory/memory9/online Will not take the mem_hotplug_lock. However the device_lock() and device_hotplug_lock. E.g. via memory_probe_store(), we can end up calling add_memory()->online_pages() without the device_hotplug_lock. So we can have concurrent callers in online_pages(). We e.g. touch in online_pages() basically unprotected zone->present_pages then. Looks like there is a longer history to that (see Patch #2 for details), and fixing it to work the way it was intended is not really possible. We would e.g. have to take the mem_hotplug_lock in device/base/core.c, which sounds wrong. Summary: We had a lock inversion on mem_hotplug_lock and device_lock(). More details can be found in patch 3 and patch 6. I propose the general rules (documentation added in patch 6): 1. add_memory/add_memory_resource() must only be called with device_hotplug_lock. 2. remove_memory() must only be called with device_hotplug_lock. This is already documented and holds for all callers. 3. device_online()/device_offline() must only be called with device_hotplug_lock. This is already documented and true for now in core code. Other callers (related to memory hotplug) have to be fixed up. 4. mem_hotplug_lock is taken inside of add_memory/remove_memory/ online_pages/offline_pages. To me, this looks way cleaner than what we have right now (and easier to verify). And looking at the documentation of remove_memory, using lock_device_hotplug also for add_memory() feels natural. RFCv2 -> v1: - Dropped an unnecessary _ref from remove_memory() in patch #1 - Minor patch description fixes. - Added rb's RFC -> RFCv2: - Don't export device_hotplug_lock, provide proper remove_memory/add_memory wrappers. - Split up the patches a bit. - Try to improve powernv memtrace locking - Add some documentation for locking that matches my knowledge David Hildenbrand (6): mm/memory_hotplug: make remove_memory() take the device_hotplug_lock mm/memory_hotplug: make add_memory() take the device_hotplug_lock mm/memory_hotplug: fix online/offline_pages called w.o. mem_hotplug_lock powerpc/powernv: hold device_hotplug_lock when calling device_online() powerpc/powernv: hold device_hotplug_lock in memtrace_offline_pages() memory-hotplug.txt: Add some details about locking internals Documentation/memory-hotplug.txt | 39 +++++++++++- arch/powerpc/platforms/powernv/memtrace.c | 14 +++-- .../platforms/pseries/hotplug-memory.c | 8 +-- drivers/acpi/acpi_memhotplug.c | 4 +- drivers/base/memory.c | 22 +++---- drivers/xen/balloon.c | 3 + include/linux/memory_hotplug.h | 4 +- mm/memory_hotplug.c | 59 +++++++++++++++---- 8 files changed, 115 insertions(+), 38 deletions(-)