ACPI / scan: Acquire device_hotplug_lock in acpi_scan_init()
Subsystems (2)
PATCH 1/1 ACPI / scan: Acquire device_hotplug_lock in acpi_scan_init()
We end up calling __add_memory() without the device hotplug lock held. (I used a local patch to assert in __add_memory() that the
device_hotplug_lock is held - I might upstream that as well soon)
[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50
So perform the locking just like in acpi_device_hotplug().
--- drivers/acpi/scan.c | 3 +++ 1 file changed, 3 insertions(+)
drivers/acpi/scan.c +3 -0
@@ -2204,7 +2204,9 @@ int __init acpi_scan_init(void)acpi_gpe_apply_masked_gpes ();acpi_update_all_gpes (); +lock_device_hotplug ();mutex_lock (&acpi_scan_lock); +/* * Enumerate devices in the ACPI namespace. */
@@ -2232,6 +2234,7 @@ int __init acpi_scan_init(void) out:mutex_unlock (&acpi_scan_lock); +unlock_device_hotplug ();return result; }
Signature · 1 line
2.21.0
Quoted 20 lines · CEST David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50
Quoted 8 lines
So perform the locking just like in
acpi_device_hotplug().Cc "Rafael J. Wysocki" <[email protected]>Cc Len Brown <[email protected]Cc Andrew Morton <[email protected]>Cc Oscar Salvador <[email protected]>Cc Michal Hocko <[email protected]>Signed-off-by David Hildenbrand <[email protected]>
Quoted 27 lines
--- drivers/acpi/scan.c | 3 +++ 1 file changed, 3 insertions(+)drivers/acpi/scan.c +3
-0@@ -2204,7 +2204,9 @@ int __init acpi_scan_init(void)acpi_gpe_apply_masked_gpes ();acpi_update_all_gpes (); +lock_device_hotplug ();mutex_lock (&acpi_scan_lock); +/* * Enumerate devices in the ACPI namespace. */ @@ -2232,6 +2234,7 @@ int __init acpi_scan_init(void) out:mutex_unlock (&acpi_scan_lock); +unlock_device_hotplug ();return result; }
Quoted 20 lines · David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50
Quoted 8 lines
So perform the locking just like in
acpi_device_hotplug().Cc "Rafael J. Wysocki" <[email protected]>Cc Len Brown <[email protected]Cc Andrew Morton <[email protected]>Cc Oscar Salvador <[email protected]>Cc Michal Hocko <[email protected]>Signed-off-by David Hildenbrand <[email protected]>
Given that that call comes from a __init function, so while booting, I wonder how bad it is.
Anyway, let us be consistent:
Quoted 29 lines
--- drivers/acpi/scan.c | 3 +++ 1 file changed, 3 insertions(+)drivers/acpi/scan.c +3
-0@@ -2204,7 +2204,9 @@ int __init acpi_scan_init(void)acpi_gpe_apply_masked_gpes ();acpi_update_all_gpes (); +lock_device_hotplug ();mutex_lock (&acpi_scan_lock); +/* * Enumerate devices in the ACPI namespace. */ @@ -2232,6 +2234,7 @@ int __init acpi_scan_init(void) out:mutex_unlock (&acpi_scan_lock); +unlock_device_hotplug ();return result; }Signature · 1 line
2.21.0
Signature · 2 lines
Oscar Salvador SUSE L3
Rafael J. Wysocki
maintainerQuoted 31 lines · Oscar Salvador
Quoted 20 lines · David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50Quoted 8 lines
So perform the locking just like in
acpi_device_hotplug().Cc "Rafael J. Wysocki" <[email protected]>Cc Len Brown <[email protected]Cc Andrew Morton <[email protected]>Cc Oscar Salvador <[email protected]>Cc Michal Hocko <[email protected]>Signed-off-by David Hildenbrand <[email protected]>
Quoted 2 lines
Given that that call comes from a
__initfunction, so while booting, I wonder how bad it is.
Yes, it probably does not matter.
Quoted 1 line
Anyway, let us be consistent:
Right.
Quoted 32 lines · Rafael J. Wysocki
Quoted 31 lines · Oscar Salvador
Quoted 28 lines · David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50So perform the locking just like in
acpi_device_hotplug().Cc "Rafael J. Wysocki" <[email protected]>Cc Len Brown <[email protected]Cc Andrew Morton <[email protected]>Cc Oscar Salvador <[email protected]>Cc Michal Hocko <[email protected]>Signed-off-by David Hildenbrand <[email protected]>
Quoted 4 lines
Quoted 2 lines
Given that that call comes from a
__initfunction, so while booting, I wonder how bad it is.Yes, it probably does not matter.
It can at least confuse lockdep, but I agree that this is not stable material.
Quoted 5 lines
Quoted 2 lines
Anyway, let us be consistent:
Right.
Signature · 3 lines
Thanks, David / dhildenb
Quoted 20 lines · David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50
Quoted 1 line
So perform the locking just like in
acpi_device_hotplug().
While playing with the device_hotplug_lock, can we actually document what it is protecting please? I have a bad feeling that we are adding this lock just because some other code path does rather than with a good idea why it is needed. This patch just confirms that. What exactly does the lock protect from here in an early boot stage.
Quoted 34 lines
Cc "Rafael J. Wysocki" <[email protected]>Cc Len Brown <[email protected]Cc Andrew Morton <[email protected]>Cc Oscar Salvador <[email protected]>Cc Michal Hocko <[email protected]>Signed-off-by David Hildenbrand <[email protected]>--- drivers/acpi/scan.c | 3 +++ 1 file changed, 3 insertions(+)drivers/acpi/scan.c +3
-0@@ -2204,7 +2204,9 @@ int __init acpi_scan_init(void)acpi_gpe_apply_masked_gpes ();acpi_update_all_gpes (); +lock_device_hotplug ();mutex_lock (&acpi_scan_lock); +/* * Enumerate devices in the ACPI namespace. */ @@ -2232,6 +2234,7 @@ int __init acpi_scan_init(void) out:mutex_unlock (&acpi_scan_lock); +unlock_device_hotplug ();return result; }Signature · 1 line
2.21.0
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 21 lines · Michal Hocko
Quoted 20 lines · David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50
Quoted 7 lines
Quoted 1 line
So perform the locking just like in
acpi_device_hotplug().While playing with the
device_hotplug_lock, can we actually document what it is protecting please? I have a bad feeling that we are adding this lock just because some other code path does rather than with a good idea why it is needed. This patch just confirms that. What exactly does the lock protect from here in an early boot stage.
We have plenty of documentation already
mm/memory_hotplug.c
git grep -C5 device_hotplug mm/memory_hotplug.c
Also see
Documentation/core-api/memory-hotplug.rst
Regarding the early stage: primarily lockdep as I mentioned.
Signature · 3 lines
Thanks, David / dhildenb
Quoted 32 lines · David Hildenbrand
Quoted 21 lines · Michal Hocko
Quoted 20 lines · David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50Quoted 7 lines
Quoted 1 line
So perform the locking just like in
acpi_device_hotplug().While playing with the
device_hotplug_lock, can we actually document what it is protecting please? I have a bad feeling that we are adding this lock just because some other code path does rather than with a good idea why it is needed. This patch just confirms that. What exactly does the lock protect from here in an early boot stage.We have plenty of documentation already
Quoted 7 lines
mm/memory_hotplug.cgit grep -C5
device_hotplugmm/memory_hotplug.cAlso see
Documentation/core-api/memory-hotplug.rst
OK, fair enough. I was more pointing to a documentation right there where the lock is declared because that is the place where people usually check for documentation. The core-api documentation looks quite nice. And based on that doc it seems that this patch is actually not needed because neither the online/offline or cpu hotplug should be possible that early unless I am missing something.
Quoted 1 line
Regarding the early stage: primarily lockdep as I mentioned.
Could you add a lockdep splat that would be fixed by this patch to the changelog for reference?
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 39 lines · Michal Hocko
Quoted 38 lines · David Hildenbrand
Quoted 21 lines · Michal Hocko
Quoted 20 lines · David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50Quoted 7 lines
Quoted 1 line
So perform the locking just like in
acpi_device_hotplug().While playing with the
device_hotplug_lock, can we actually document what it is protecting please? I have a bad feeling that we are adding this lock just because some other code path does rather than with a good idea why it is needed. This patch just confirms that. What exactly does the lock protect from here in an early boot stage.We have plenty of documentation already
mm/memory_hotplug.cgit grep -C5
device_hotplugmm/memory_hotplug.cAlso see
Quoted 8 lines
Quoted 1 line
Documentation/core-api/memory-hotplug.rstOK, fair enough. I was more pointing to a documentation right there where the lock is declared because that is the place where people usually check for documentation. The core-api documentation looks quite nice. And based on that doc it seems that this patch is actually not needed because neither the online/offline or cpu hotplug should be possible that early unless I am missing something.
I really prefer to stick to locking rules as outlined on the interfaces if it doesn't hurt. Why it is not needed is not clear.
Quoted 6 lines
Quoted 2 lines
Regarding the early stage: primarily lockdep as I mentioned.
Could you add a lockdep splat that would be fixed by this patch to the changelog for reference?
I have one where I enforce what's documented (but that's of course not upstream and therefore not "real" yet)
commit 263da346cd3cf526de3f5138827fbc3520f2f8e0
Author: David Hildenbrand <[email protected]> Date: Fri Jun 21 12:05:39 2019 +0200
mm/memory_hotplug: Assert that the device_hotplug_lock is held
We currently need the device_hotplug_lock(), as documented. Let's assert that the lock is held when adding/removing/onlining/offlining memory.
Updated documentation to make this clearer.
That patch in return was the result of debugging a lockdep warning we can trigger right now (and I think it's a false positive prevented by the device_hotplug_lock - I think it is the tie breaker). Anyhow, this patch here didn't change it.
1. Start a guest with a DIMM attached
2. Online a memory block of that DIMM
3. Unplug the DIMM
:/# [ 22.616108] Offlined Pages 32768
[ 22.631567]
[ 22.632337] ======================================================
[ 22.635104] WARNING: possible circular locking dependency detected
[ 22.637475] 5.3.0-rc1-next-20190723+ #111 Not tainted
[ 22.639314] ------------------------------------------------------
[ 22.641276] kworker/u4:0/8 is trying to acquire lock:
[ 22.642578] (____ptrval____) (kn->count#39){++++}, at: kernfs_remove_by_name_ns+0x3c/0x80
[ 22.645004]
[ 22.645004] but task is already holding lock:
[ 22.646495] (____ptrval____) (mem_sysfs_mutex){+.+.}, at: remove_memory_block_devices+0x65/0xd0
[ 22.648649]
[ 22.648649] which lock already depends on the new lock.
[ 22.648649]
[ 22.650488]
[ 22.650488] the existing dependency chain (in reverse order) is:
[ 22.651987]
[ 22.651987] -> #4 (mem_sysfs_mutex){+.+.}:
[ 22.653254] __mutex_lock+0x8d/0x930
[ 22.654079] create_memory_block_devices+0xa4/0x140
[ 22.655292] add_memory_resource+0xd6/0x200
[ 22.656252] __add_memory+0x58/0x90
[ 22.657096] acpi_memory_device_add+0x149/0x2b0
[ 22.658126] acpi_bus_attach+0xf1/0x1f0
[ 22.658899] acpi_bus_attach+0x66/0x1f0
[ 22.659698] acpi_bus_attach+0x66/0x1f0
[ 22.660482] acpi_bus_attach+0x66/0x1f0
[ 22.661265] acpi_bus_scan+0x3e/0x90
[ 22.662098] acpi_scan_init+0x104/0x24d
[ 22.662920] acpi_init+0x2ab/0x30d
[ 22.663733] do_one_initcall+0x58/0x2cf
[ 22.664727] kernel_init_freeable+0x1b8/0x242
[ 22.665780] kernel_init+0x5/0xf1
[ 22.666494] ret_from_fork+0x3a/0x50
[ 22.667271]
[ 22.667271] -> #3 (mem_hotplug_lock.rw_sem){++++}:
[ 22.668378] get_online_mems+0x39/0xc0
[ 22.669327] kmem_cache_create_usercopy+0x29/0x280
[ 22.670369] kmem_cache_create+0xd/0x10
[ 22.671412] ptlock_cache_init+0x1b/0x23
[ 22.672206] start_kernel+0x225/0x501
[ 22.672979] secondary_startup_64+0xa4/0xb0
[ 22.673887]
[ 22.673887] -> #2 (cpu_hotplug_lock.rw_sem){++++}:
[ 22.675091] cpus_read_lock+0x39/0xc0
[ 22.675962] __offline_pages+0x3e/0x7c0
[ 22.676997] memory_subsys_offline+0x3a/0x60
[ 22.678073] device_offline+0x82/0xb0
[ 22.679039] acpi_bus_offline+0xdb/0x150
[ 22.679912] acpi_device_hotplug+0x1b4/0x3a0
[ 22.680939] acpi_hotplug_work_fn+0x15/0x20
[ 22.682025] process_one_work+0x26c/0x5a0
[ 22.683019] worker_thread+0x48/0x3e0
[ 22.683942] kthread+0x103/0x140
[ 22.684855] ret_from_fork+0x3a/0x50
[ 22.685841]
[ 22.685841] -> #1 (&device->physical_node_lock){+.+.}:
[ 22.687246] __mutex_lock+0x8d/0x930
[ 22.688179] acpi_get_first_physical_node+0x18/0x60
[ 22.689699] acpi_companion_match+0x3b/0x60
[ 22.690989] acpi_device_uevent_modalias+0x9/0x20
[ 22.692626] platform_uevent+0xd/0x40
[ 22.693832] dev_uevent+0x86/0x1c0
[ 22.695133] uevent_show+0x93/0x100
[ 22.695988] dev_attr_show+0x14/0x40
[ 22.697342] sysfs_kf_seq_show+0xb2/0xf0
[ 22.698845] seq_read+0xd0/0x3f0
[ 22.700066] vfs_read+0xc0/0x170
[ 22.701168] ksys_read+0x63/0xe0
[ 22.702392] do_syscall_64+0x4b/0x1b0
[ 22.703979] entry_SYSCALL_64_after_hwframe+0x49/0xbe
[ 22.705708]
[ 22.705708] -> #0 (kn->count#39){++++}:
[ 22.707658] __lock_acquire+0xe2f/0x1a20
[ 22.708877] lock_acquire+0x95/0x190
[ 22.710299] __kernfs_remove+0x253/0x2f0
[ 22.711936] kernfs_remove_by_name_ns+0x3c/0x80
[ 22.713392] device_del+0x148/0x360
[ 22.714685] device_unregister+0x9/0x20
[ 22.716414] remove_memory_block_devices+0x90/0xd0
[ 22.718135] try_remove_memory+0xc6/0x130
[ 22.719669] __remove_memory+0x5/0xc
[ 22.721178] acpi_memory_device_remove+0x72/0xf0
[ 22.723178] acpi_bus_trim+0x50/0x90
[ 22.724537] acpi_device_hotplug+0x222/0x3a0
[ 22.726257] acpi_hotplug_work_fn+0x15/0x20
[ 22.728044] process_one_work+0x26c/0x5a0
[ 22.729825] worker_thread+0x48/0x3e0
[ 22.731128] kthread+0x103/0x140
[ 22.732137] ret_from_fork+0x3a/0x50
[ 22.733368]
[ 22.733368] other info that might help us debug this:
[ 22.733368]
[ 22.736178] Chain exists of:
[ 22.736178] kn->count#39 --> mem_hotplug_lock.rw_sem --> mem_sysfs_mutex
[ 22.736178]
[ 22.739723] Possible unsafe locking scenario:
[ 22.739723]
[ 22.741143] CPU0 CPU1
[ 22.741788] ---- ----
[ 22.742653] lock(mem_sysfs_mutex);
[ 22.743990] lock(mem_hotplug_lock.rw_sem);
[ 22.746069] lock(mem_sysfs_mutex);
[ 22.747207] lock(kn->count#39);
[ 22.748132]
[ 22.748132] *** DEADLOCK ***
[ 22.748132]
[ 22.749182] 7 locks held by kworker/u4:0/8:
[ 22.750684] #0: (____ptrval____) ((wq_completion)kacpi_hotplug){+.+.}, at: process_one_work+0x1e9/0x5a0
[ 22.753966] #1: (____ptrval____) ((work_completion)(&hpw->work)){+.+.}, at: process_one_work+0x1e9/0x5a0
[ 22.756429] #2: (____ptrval____) (device_hotplug_lock){+.+.}, at: acpi_device_hotplug+0x2d/0x3a0
[ 22.758292] #3: (____ptrval____) (acpi_scan_lock){+.+.}, at: acpi_device_hotplug+0x3b/0x3a0
[ 22.759836] #4: (____ptrval____) (cpu_hotplug_lock.rw_sem){++++}, at: try_remove_memory+0x3b/0x130
[ 22.761463] #5: (____ptrval____) (mem_hotplug_lock.rw_sem){++++}, at: percpu_down_write+0x1b/0xf2
[ 22.763812] #6: (____ptrval____) (mem_sysfs_mutex){+.+.}, at: remove_memory_block_devices+0x65/0xd0Signature · 3 lines
Thanks, David / dhildenb
Quoted 53 lines · David Hildenbrand
Quoted 39 lines · Michal Hocko
Quoted 38 lines · David Hildenbrand
Quoted 21 lines · Michal Hocko
Quoted 20 lines · David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50Quoted 7 lines
Quoted 1 line
So perform the locking just like in
acpi_device_hotplug().While playing with the
device_hotplug_lock, can we actually document what it is protecting please? I have a bad feeling that we are adding this lock just because some other code path does rather than with a good idea why it is needed. This patch just confirms that. What exactly does the lock protect from here in an early boot stage.We have plenty of documentation already
mm/memory_hotplug.cgit grep -C5
device_hotplugmm/memory_hotplug.cAlso see
Quoted 8 lines
Quoted 1 line
Documentation/core-api/memory-hotplug.rstOK, fair enough. I was more pointing to a documentation right there where the lock is declared because that is the place where people usually check for documentation. The core-api documentation looks quite nice. And based on that doc it seems that this patch is actually not needed because neither the online/offline or cpu hotplug should be possible that early unless I am missing something.
I really prefer to stick to locking rules as outlined on the interfaces if it doesn't hurt. Why it is not needed is not clear.
Quoted 1 line
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
Regarding the early stage: primarily lockdep as I mentioned.
Could you add a lockdep splat that would be fixed by this patch to the changelog for reference?
I have one where I enforce what's documented (but that's of course not upstream and therefore not "real" yet)
Then I suppose to not add locking for something that is not a problem. Really, think about it. People will look at this code and follow the lead without really knowing why the locking is needed. device_hotplug_lock has its purpose and if the code in question doesn't need synchronization for the documented scenarios then the locking simply shouldn't be there. Adding the lock just because of a non-existing, and IMHO dubious, lockdep splats is just wrong.
We need to rationalize the locking here, not to add more hacks.
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 71 lines · Michal Hocko
Quoted 53 lines · David Hildenbrand
Quoted 39 lines · Michal Hocko
Quoted 38 lines · David Hildenbrand
Quoted 21 lines · Michal Hocko
Quoted 20 lines · David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50Quoted 7 lines
Quoted 1 line
So perform the locking just like in
acpi_device_hotplug().While playing with the
device_hotplug_lock, can we actually document what it is protecting please? I have a bad feeling that we are adding this lock just because some other code path does rather than with a good idea why it is needed. This patch just confirms that. What exactly does the lock protect from here in an early boot stage.We have plenty of documentation already
mm/memory_hotplug.cgit grep -C5
device_hotplugmm/memory_hotplug.cAlso see
Quoted 8 lines
Quoted 1 line
Documentation/core-api/memory-hotplug.rstOK, fair enough. I was more pointing to a documentation right there where the lock is declared because that is the place where people usually check for documentation. The core-api documentation looks quite nice. And based on that doc it seems that this patch is actually not needed because neither the online/offline or cpu hotplug should be possible that early unless I am missing something.
I really prefer to stick to locking rules as outlined on the interfaces if it doesn't hurt. Why it is not needed is not clear.
Quoted 1 line
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
Regarding the early stage: primarily lockdep as I mentioned.
Could you add a lockdep splat that would be fixed by this patch to the changelog for reference?
I have one where I enforce what's documented (but that's of course not upstream and therefore not "real" yet)
Then I suppose to not add locking for something that is not a problem. Really, think about it. People will look at this code and follow the lead without really knowing why the locking is needed.
device_hotplug_lockhas its purpose and if the code in question doesn't need synchronization for the documented scenarios then the locking simply shouldn't be there. Adding the lock just because of a non-existing, and IMHO dubious, lockdep splats is just wrong.
Quoted 1 line
We need to rationalize the locking here, not to add more hacks.
No, sorry. The real hack is calling a function that is *documented* to be called under lock without it. That is an optimization for a special case. That is the black magic in the code.
The only alternative I see to this patch is adding a comment like
/*
* We end up calling __add_memory() without the device_hotplug_lock
* held. This is fine as we cannot race with other hotplug activities
* and userspace trying to online memory blocks.
*/
Personally, I don't think that's any better than just grabbing the lock as we are told to. (honestly, I don't see how optimizing away the lock here is of *any* help to optimize our overall memory hotplug locking)
@Rafael, what's your take? lock or comment?
Signature · 3 lines
Thanks, David / dhildenb
Rafael J. Wysocki
maintainerQuoted 87 lines · David Hildenbrand
Quoted 71 lines · Michal Hocko
Quoted 53 lines · David Hildenbrand
Quoted 39 lines · Michal Hocko
Quoted 38 lines · David Hildenbrand
Quoted 21 lines · Michal Hocko
Quoted 20 lines · David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50Quoted 7 lines
Quoted 1 line
So perform the locking just like in
acpi_device_hotplug().While playing with the
device_hotplug_lock, can we actually document what it is protecting please? I have a bad feeling that we are adding this lock just because some other code path does rather than with a good idea why it is needed. This patch just confirms that. What exactly does the lock protect from here in an early boot stage.We have plenty of documentation already
mm/memory_hotplug.cgit grep -C5
device_hotplugmm/memory_hotplug.cAlso see
Quoted 8 lines
Quoted 1 line
Documentation/core-api/memory-hotplug.rstOK, fair enough. I was more pointing to a documentation right there where the lock is declared because that is the place where people usually check for documentation. The core-api documentation looks quite nice. And based on that doc it seems that this patch is actually not needed because neither the online/offline or cpu hotplug should be possible that early unless I am missing something.
I really prefer to stick to locking rules as outlined on the interfaces if it doesn't hurt. Why it is not needed is not clear.
Quoted 1 line
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
Regarding the early stage: primarily lockdep as I mentioned.
Could you add a lockdep splat that would be fixed by this patch to the changelog for reference?
I have one where I enforce what's documented (but that's of course not upstream and therefore not "real" yet)
Then I suppose to not add locking for something that is not a problem. Really, think about it. People will look at this code and follow the lead without really knowing why the locking is needed.
device_hotplug_lockhas its purpose and if the code in question doesn't need synchronization for the documented scenarios then the locking simply shouldn't be there. Adding the lock just because of a non-existing, and IMHO dubious, lockdep splats is just wrong.Quoted 1 line
We need to rationalize the locking here, not to add more hacks.
No, sorry. The real hack is calling a function that is *documented* to be called under lock without it. That is an optimization for a special case. That is the black magic in the code.
The only alternative I see to this patch is adding a comment like
/*
* We end up calling__add_memory()without thedevice_hotplug_lock
* held. This is fine as we cannot race with other hotplug activities
* and userspace trying to online memory blocks.
*/
Quoted 5 lines
Personally, I don't think that's any better than just grabbing the lock as we are told to. (honestly, I don't see how optimizing away the lock here is of *any* help to optimize our overall memory hotplug locking)
@Rafael, what's your take? lock or comment?
Well, I have ACKed your patch already. :-)
That said, adding a comment stating that the lock is acquired mostly for consistency wouldn't hurt.
Quoted 88 lines · Rafael J. Wysocki
Quoted 87 lines · David Hildenbrand
Quoted 71 lines · Michal Hocko
Quoted 53 lines · David Hildenbrand
Quoted 39 lines · Michal Hocko
Quoted 38 lines · David Hildenbrand
Quoted 21 lines · Michal Hocko
Quoted 20 lines · David Hildenbrand
We end up calling
__add_memory()without the device hotplug lock held. (I used a local patch to assert in__add_memory()that the
device_hotplug_lockis held - I might upstream that as well soon)[ 26.771684] create_memory_block_devices+0xa4/0x140 [ 26.772952] add_memory_resource+0xde/0x200 [ 26.773987] __add_memory+0x6e/0xa0 [ 26.775161] acpi_memory_device_add+0x149/0x2b0 [ 26.776263] acpi_bus_attach+0xf1/0x1f0 [ 26.777247] acpi_bus_attach+0x66/0x1f0 [ 26.778268] acpi_bus_attach+0x66/0x1f0 [ 26.779073] acpi_bus_attach+0x66/0x1f0 [ 26.780143] acpi_bus_scan+0x3e/0x90 [ 26.780844] acpi_scan_init+0x109/0x257 [ 26.781638] acpi_init+0x2ab/0x30d [ 26.782248] do_one_initcall+0x58/0x2cf [ 26.783181] kernel_init_freeable+0x1bd/0x247 [ 26.784345] kernel_init+0x5/0xf1 [ 26.785314] ret_from_fork+0x3a/0x50Quoted 7 lines
Quoted 1 line
So perform the locking just like in
acpi_device_hotplug().While playing with the
device_hotplug_lock, can we actually document what it is protecting please? I have a bad feeling that we are adding this lock just because some other code path does rather than with a good idea why it is needed. This patch just confirms that. What exactly does the lock protect from here in an early boot stage.We have plenty of documentation already
mm/memory_hotplug.cgit grep -C5
device_hotplugmm/memory_hotplug.cAlso see
Quoted 8 lines
Quoted 1 line
Documentation/core-api/memory-hotplug.rstOK, fair enough. I was more pointing to a documentation right there where the lock is declared because that is the place where people usually check for documentation. The core-api documentation looks quite nice. And based on that doc it seems that this patch is actually not needed because neither the online/offline or cpu hotplug should be possible that early unless I am missing something.
I really prefer to stick to locking rules as outlined on the interfaces if it doesn't hurt. Why it is not needed is not clear.
Quoted 1 line
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
Regarding the early stage: primarily lockdep as I mentioned.
Could you add a lockdep splat that would be fixed by this patch to the changelog for reference?
I have one where I enforce what's documented (but that's of course not upstream and therefore not "real" yet)
Then I suppose to not add locking for something that is not a problem. Really, think about it. People will look at this code and follow the lead without really knowing why the locking is needed.
device_hotplug_lockhas its purpose and if the code in question doesn't need synchronization for the documented scenarios then the locking simply shouldn't be there. Adding the lock just because of a non-existing, and IMHO dubious, lockdep splats is just wrong.Quoted 1 line
We need to rationalize the locking here, not to add more hacks.
No, sorry. The real hack is calling a function that is *documented* to be called under lock without it. That is an optimization for a special case. That is the black magic in the code.
The only alternative I see to this patch is adding a comment like
/*
* We end up calling__add_memory()without thedevice_hotplug_lock
* held. This is fine as we cannot race with other hotplug activities
* and userspace trying to online memory blocks.
*/
Quoted 7 lines
Quoted 5 lines
Personally, I don't think that's any better than just grabbing the lock as we are told to. (honestly, I don't see how optimizing away the lock here is of *any* help to optimize our overall memory hotplug locking)
@Rafael, what's your take? lock or comment?
Well, I have ACKed your patch already. :-)
It's never to late to un-ACK if you changed your mind :)
Quoted 4 lines
That said, adding a comment stating that the lock is acquired mostly for consistency wouldn't hurt.
I can certainly do that. Thanks!
Signature · 3 lines
Thanks, David / dhildenb
Quoted 1 line · David Hildenbrand
On 25.07.19 21:19, Michal Hocko wrote:
[...]
Quoted 5 lines
Quoted 1 line
We need to rationalize the locking here, not to add more hacks.
No, sorry. The real hack is calling a function that is *documented* to be called under lock without it. That is an optimization for a special case. That is the black magic in the code.
OK, let me ask differently. What does the device_hotplug_lock actually protects from in the add_memory path? (Which data structures)
This function is meant to be used when struct pages and node/zone data structures should be updated. Why should we even care about some device concept here? This should all be handled a layer up. Not all memory will have user space API to control online/offline state.
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 9 lines · Michal Hocko
Quoted 1 line · David Hildenbrand
On 25.07.19 21:19, Michal Hocko wrote:
[...]
Quoted 5 lines
Quoted 1 line
We need to rationalize the locking here, not to add more hacks.
No, sorry. The real hack is calling a function that is *documented* to be called under lock without it. That is an optimization for a special case. That is the black magic in the code.
Quoted 7 lines
OK, let me ask differently. What does the
device_hotplug_lockactually protects from in theadd_memorypath? (Which data structures)This function is meant to be used when struct pages and node/zone data structures should be updated. Why should we even care about some device concept here? This should all be handled a layer up. Not all memory will have user space API to control online/offline state.
Via add_memory()/__add_memory() we create memory block devices for all memory. So all memory we create via this function (IOW, hotplug) will have user space APIs.
Sorry, I can't follow what you are saying here - are you confusing the function we are talking about with arch_add_memory() ? (where I pulled out the creation of memory block devices)
Signature · 3 lines
Thanks, David / dhildenb
Quoted 13 lines · David Hildenbrand
Quoted 12 lines · Michal Hocko
Quoted 1 line · David Hildenbrand
On 25.07.19 21:19, Michal Hocko wrote:
[...]
Quoted 5 lines
Quoted 1 line
We need to rationalize the locking here, not to add more hacks.
No, sorry. The real hack is calling a function that is *documented* to be called under lock without it. That is an optimization for a special case. That is the black magic in the code.
OK, let me ask differently. What does the
device_hotplug_lockactually protects from in theadd_memorypath? (Which data structures)
Quoted 8 lines
Quoted 4 lines
This function is meant to be used when struct pages and node/zone data structures should be updated. Why should we even care about some device concept here? This should all be handled a layer up. Not all memory will have user space API to control online/offline state.
Via add_memory()/__add_memory() we create memory block devices for all memory. So all memory we create via this function (IOW, hotplug) will have user space APIs.
Ups, I have mixed add_memory with add_pages which I've had in mind while writing that. Sorry about the confusion.
Anyway, my dislike of the device_hotplug_lock persists. I would really love to see it go rather than grow even more to the hotplug code. We should be really striving for mem hotplug internal and ideally range defined locking longterm.
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 19 lines · Michal Hocko
Quoted 18 lines · David Hildenbrand
Quoted 16 lines · Michal Hocko
Quoted 1 line · David Hildenbrand
On 25.07.19 21:19, Michal Hocko wrote:
[...]
Quoted 5 lines
Quoted 1 line
We need to rationalize the locking here, not to add more hacks.
No, sorry. The real hack is calling a function that is *documented* to be called under lock without it. That is an optimization for a special case. That is the black magic in the code.
OK, let me ask differently. What does the
device_hotplug_lockactually protects from in theadd_memorypath? (Which data structures)This function is meant to be used when struct pages and node/zone data structures should be updated. Why should we even care about some device concept here? This should all be handled a layer up. Not all memory will have user space API to control online/offline state.
Quoted 6 lines
Quoted 3 lines
Via add_memory()/__add_memory() we create memory block devices for all memory. So all memory we create via this function (IOW, hotplug) will have user space APIs.
Ups, I have mixed
add_memorywithadd_pageswhich I've had in mind while writing that. Sorry about the confusion.
No worries :)
Quoted 5 lines
Anyway, my dislike of the
device_hotplug_lockpersists. I would really love to see it go rather than grow even more to the hotplug code. We should be really striving for mem hotplug internal and ideally range defined locking longterm.
Yes, and that is a different story, because it will require major changes to all add_memory() users. (esp, due to the documented race conditions). Having that said, memory hotplug locking is not ideal yet.
Signature · 3 lines
Thanks, David / dhildenb
Quoted 1 line · David Hildenbrand
On 26.07.19 10:31, Michal Hocko wrote:
[...]
Quoted 8 lines
Quoted 4 lines
Anyway, my dislike of the
device_hotplug_lockpersists. I would really love to see it go rather than grow even more to the hotplug code. We should be really striving for mem hotplug internal and ideally range defined locking longterm.Yes, and that is a different story, because it will require major changes to all
add_memory()users. (esp, due to the documented race conditions). Having that said, memory hotplug locking is not ideal yet.
I am really happy to hear that we are on the same page here. Do we have any document (I am sorry but I am lacking behind recent development in this area) that describes roadblocks to remove device_hotplug_lock?
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 8 lines · Michal Hocko
Quoted 1 line · David Hildenbrand
On 26.07.19 10:31, Michal Hocko wrote:
[...]
Quoted 5 lines
Quoted 4 lines
Anyway, my dislike of the
device_hotplug_lockpersists. I would really love to see it go rather than grow even more to the hotplug code. We should be really striving for mem hotplug internal and ideally range defined locking longterm.
Quoted 7 lines
Quoted 3 lines
Yes, and that is a different story, because it will require major changes to all
add_memory()users. (esp, due to the documented race conditions). Having that said, memory hotplug locking is not ideal yet.I am really happy to hear that we are on the same page here. Do we have any document (I am sorry but I am lacking behind recent development in this area) that describes roadblocks to remove
device_hotplug_lock?
Only the core-api document I mentioned (I documented there quite some current conditions I identified back then).
I am not sure if we can remove it completely from add_memory()/remove_memory(): We actually create/delete devices which can otherwise create races with user space.
Besides that:
- try_offline_node() needs the lock to synchronize against cpu hotplug
- I *assume* try_online_node() needs it as well
Then, there is the possible race condition with user space onlining memory avoided by the lock. Also, currently the lock protects the "online_type" when onlining memory.
Then, there might be other global variables (eventually zone/node/section related) that might need this lock right now - no details known.
IOW, we have to be very carefully and it is more involved than it might seem.
Locking is definitely better (and more reliably!) than one year ago, but there is definitely a lot to do. (unfortunately, just like in many areas in memory hotplug code :( - say zone handling when offlining/failing to online memory).
Signature · 3 lines
Thanks, David / dhildenb
Quoted 13 lines · David Hildenbrand
Quoted 12 lines · Michal Hocko
Quoted 1 line · David Hildenbrand
On 26.07.19 10:31, Michal Hocko wrote:
[...]
Quoted 8 lines
Quoted 4 lines
Anyway, my dislike of the
device_hotplug_lockpersists. I would really love to see it go rather than grow even more to the hotplug code. We should be really striving for mem hotplug internal and ideally range defined locking longterm.Yes, and that is a different story, because it will require major changes to all
add_memory()users. (esp, due to the documented race conditions). Having that said, memory hotplug locking is not ideal yet.
Quoted 6 lines
Quoted 3 lines
I am really happy to hear that we are on the same page here. Do we have any document (I am sorry but I am lacking behind recent development in this area) that describes roadblocks to remove
device_hotplug_lock?Only the core-api document I mentioned (I documented there quite some current conditions I identified back then).
That document doesn't describe which _data structures_ are protected by the lock though. It documents only the current state of locking.
Quoted 3 lines
I am not sure if we can remove it completely from add_memory()/remove_memory(): We actually create/delete devices which can otherwise create races with user space.
More details would be really appreciated.
Quoted 3 lines
Besides that:
-try_offline_node()needs the lock to synchronize against cpu hotplug
- I *assume*try_online_node()needs it as well
more details on why would be great.
Quoted 3 lines
Then, there is the possible race condition with user space onlining memory avoided by the lock. Also, currently the lock protects the "
online_type" when onlining memory.
I do not see the race, if the user API triggered online/offline takes a range lock on the affected physical memory range
Quoted 3 lines
Then, there might be other global variables (eventually zone/node/section related) that might need this lock right now - no details known.
zones/nodes have their own locking for spans. Sections should be using a low level locking but I am not really sure this is needed if there is a mem hotplug lock in place (range or global)
Quoted 2 lines
IOW, we have to be very carefully and it is more involved than it might seem.
I am not questioning that. And that is why I am asking about a todo list for that transition.
Quoted 4 lines
Locking is definitely better (and more reliably!) than one year ago, but there is definitely a lot to do. (unfortunately, just like in many areas in memory hotplug code :( - say zone handling when offlining/failing to online memory).
Yeah, the code is shaping up. And I am happy to see that happening. But please try to understand that I really do not like to see some ad-hoc locking enforcement without a clear locking model in place. This patch is an example of it. Whoever would like to rationalize locking further will have to stumble over this and scratch head why the hack the locking is there and my experience tells me that people usually go along with existing code and make further assumptions based on that so we are unlikely to get rid of the locking...
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 18 lines · Michal Hocko
Quoted 17 lines · David Hildenbrand
Quoted 15 lines · Michal Hocko
Quoted 1 line · David Hildenbrand
On 26.07.19 10:31, Michal Hocko wrote:
[...]
Quoted 8 lines
Quoted 4 lines
Anyway, my dislike of the
device_hotplug_lockpersists. I would really love to see it go rather than grow even more to the hotplug code. We should be really striving for mem hotplug internal and ideally range defined locking longterm.Yes, and that is a different story, because it will require major changes to all
add_memory()users. (esp, due to the documented race conditions). Having that said, memory hotplug locking is not ideal yet.I am really happy to hear that we are on the same page here. Do we have any document (I am sorry but I am lacking behind recent development in this area) that describes roadblocks to remove
device_hotplug_lock?
Quoted 5 lines
Quoted 2 lines
Only the core-api document I mentioned (I documented there quite some current conditions I identified back then).
That document doesn't describe which
_datastructures_are protected by the lock though. It documents only the current state of locking.
Yeah, I also thing we should find out more and document it. Unfortunately, optimize the locking is not very high on my priority list (there are more critical things to figure out than optimizing locking that at least seems to work :) ). It is on my list, though.
Quoted 11 lines
Quoted 4 lines
I am not sure if we can remove it completely from add_memory()/remove_memory(): We actually create/delete devices which can otherwise create races with user space.
More details would be really appreciated.
Quoted 3 lines
Besides that:
-try_offline_node()needs the lock to synchronize against cpu hotplug
- I *assume*try_online_node()needs it as well
Quoted 8 lines
more details on why would be great.
Quoted 3 lines
Then, there is the possible race condition with user space onlining memory avoided by the lock. Also, currently the lock protects the "
online_type" when onlining memory.I do not see the race, if the user API triggered online/offline takes a range lock on the affected physical memory range
Yeah, and that's still future work. Another item on the list.
Quoted 9 lines
Quoted 4 lines
Then, there might be other global variables (eventually zone/node/section related) that might need this lock right now - no details known.
zones/nodes have their own locking for spans. Sections should be using a low level locking but I am not really sure this is needed if there is a mem hotplug lock in place (range or global)
Quoted 5 lines
Quoted 2 lines
IOW, we have to be very carefully and it is more involved than it might seem.
I am not questioning that. And that is why I am asking about a todo list for that transition.
I think somebody will have to invest quite some effort to create that todo list first :) (I'd love to provide more information right now, but I don't really have more)
Quoted 6 lines
Quoted 5 lines
Locking is definitely better (and more reliably!) than one year ago, but there is definitely a lot to do. (unfortunately, just like in many areas in memory hotplug code :( - say zone handling when offlining/failing to online memory).
Quoted 8 lines
Yeah, the code is shaping up. And I am happy to see that happening. But please try to understand that I really do not like to see some ad-hoc locking enforcement without a clear locking model in place. This patch is an example of it. Whoever would like to rationalize locking further will have to stumble over this and scratch head why the hack the locking is there and my experience tells me that people usually go along with existing code and make further assumptions based on that so we are unlikely to get rid of the locking...
I do understand, but we really have to rethink locking in a more broad sense and document it. Here, I am going to add a comment as requested by Rafael.
Signature · 3 lines
Thanks, David / dhildenb