- 22 Mar, 2022 40 commits
-
-
Oscar Salvador authored
Abhishek reported that after patch [1], hotplug operations are taking roughly double the expected time. [2] The reason behind is that the CPU callbacks that migrate_on_reclaim_init() sets always call set_migration_target_nodes() whenever a CPU is brought up/down. But we only care about numa nodes going from having cpus to become cpuless, and vice versa, as that influences the demotion_target order. We do already have two CPU callbacks (vmstat_cpu_online() and vmstat_cpu_dead()) that check exactly that, so get rid of the CPU callbacks in migrate_on_reclaim_init() and only call set_migration_target_nodes() from vmstat_cpu_{dead,online}() whenever a numa node change its N_CPU state. [1] https://lore.kernel.org/linux-mm/20210721063926.3024591-2-ying.huang@intel.com/ [2] https://lore.kernel.org/linux-mm/eb438ddd-2919-73d4-bd9f-b7eecdd9577a@linux.vnet.ibm.com/ [osalvador@suse.de: add feedback from Huang Ying] Link: https://lkml.kernel.org/r/20220314150945.12694-1-osalvador@suse.de Link: https://lkml.kernel.org/r/20220310120749.23077-1-osalvador@suse.de Fixes: 884a6e5d ("mm/migrate: update node demotion order on hotplug events") Signed-off-by:Oscar Salvador <osalvador@suse.de> Reviewed-by:
Baolin Wang <baolin.wang@linux.alibaba.com> Tested-by:
Abhishek Goel <huntbag@linux.vnet.ibm.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: "Huang, Ying" <ying.huang@intel.com> Cc: Abhishek Goel <huntbag@linux.vnet.ibm.com> Cc: <stable@vger.kernel.org> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
David Hildenbrand authored
Let's make it clearer at which places we actually add and remove memory blocks -- streamlining the terminology -- and highlight which memory block start out online and which start out as offline. * rename add_memory_block -> add_boot_memory_block * rename init_memory_block -> add_memory_block * rename unregister_memory -> remove_memory_block * rename register_memory -> __add_memory_block * add add_hotplug_memory_block * mark add_boot_memory_block with __init (suggested by Oscar) __add_memory_block() is a pure helper for add_memory_block(), remove the somewhat obvious comment. Link: https://lkml.kernel.org/r/20220221154531.11382-1-david@redhat.comSigned-off-by:
David Hildenbrand <david@redhat.com> Reviewed-by:
-
David Hildenbrand authored
test_pages_in_a_zone() is just another nasty PFN walker that can easily stumble over ZONE_DEVICE memory ranges falling into the same memory block as ordinary system RAM: the memmap of parts of these ranges might possibly be uninitialized. In fact, we observed (on an older kernel) with UBSAN: UBSAN: Undefined behaviour in ./include/linux/mm.h:1133:50 index 7 is out of range for type 'zone [5]' CPU: 121 PID: 35603 Comm: read_all Kdump: loaded Tainted: [...] Hardware name: Dell Inc. PowerEdge R7425/08V001, BIOS 1.12.2 11/15/2019 Call Trace: dump_stack+0x9a/0xf0 ubsan_epilogue+0x9/0x7a __ubsan_handle_out_of_bounds+0x13a/0x181 test_pages_in_a_zone+0x3c4/0x500 show_valid_zones+0x1fa/0x380 dev_attr_show+0x43/0xb0 sysfs_kf_seq_show+0x1c5/0x440 seq_read+0x49d/0x1190 vfs_read+0xff/0x300 ksys_read+0xb8/0x170 do_syscall_64+0xa5/0x4b0 entry_SYSCALL_64_after_hwframe+0x6a/0xdf RIP: 0033:0x7f01f4439b52 We seem to stumble over a memmap that contains a garbage zone id. While we could try inserting pfn_to_online_page() calls, it will just make memory offlining slower, because we use test_pages_in_a_zone() to make sure we're offlining pages that all belong to the same zone. Let's just get rid of this PFN walker and determine the single zone of a memory block -- if any -- for early memory blocks during boot. For memory onlining, we know the single zone already. Let's avoid any additional memmap scanning and just rely on the zone information available during boot. For memory hot(un)plug, we only really care about memory blocks that: * span a single zone (and, thereby, a single node) * are completely System RAM (IOW, no holes, no ZONE_DEVICE) If one of these conditions is not met, we reject memory offlining. Hotplugged memory blocks (starting out offline), always meet both conditions. There are three scenarios to handle: (1) Memory hot(un)plug A memory block with zone == NULL cannot be offlined, corresponding to our previous test_pages_in_a_zone() check. After successful memory onlining/offlining, we simply set the zone accordingly. * Memory onlining: set the zone we just used for onlining * Memory offlining: set zone = NULL So a hotplugged memory block starts with zone = NULL. Once memory onlining is done, we set the proper zone. (2) Boot memory with !CONFIG_NUMA We know that there is just a single pgdat, so we simply scan all zones of that pgdat for an intersection with our memory block PFN range when adding the memory block. If more than one zone intersects (e.g., DMA and DMA32 on x86 for the first memory block) we set zone = NULL and consequently mimic what test_pages_in_a_zone() used to do. (3) Boot memory with CONFIG_NUMA At the point in time we create the memory block devices during boot, we don't know yet which nodes *actually* span a memory block. While we could scan all zones of all nodes for intersections, overlapping nodes complicate the situation and scanning all nodes is possibly expensive. But that problem has already been solved by the code that sets the node of a memory block and creates the link in the sysfs -- do_register_memory_block_under_node(). So, we hook into the code that sets the node id for a memory block. If we already have a different node id set for the memory block, we know that multiple nodes *actually* have PFNs falling into our memory block: we set zone = NULL and consequently mimic what test_pages_in_a_zone() used to do. If there is no node id set, we do the same as (2) for the given node. Note that the call order in driver_init() is: -> memory_dev_init(): create memory block devices -> node_dev_init(): link memory block devices to the node and set the node id So in summary, we detect if there is a single zone responsible for this memory block and we consequently store the zone in that case in the memory block, updating it during memory onlining/offlining. Link: https://lkml.kernel.org/r/20220210184359.235565-3-david@redhat.comSigned-off-by:
Rafael Parra <rparrazo@redhat.com> Reviewed-by:
-
David Hildenbrand authored
Patch series "drivers/base/memory: determine and store zone for single-zone memory blocks", v2. I remember talking to Michal in the past about removing test_pages_in_a_zone(), which we use for: * verifying that a memory block we intend to offline is really only managed by a single zone. We don't support offlining of memory blocks that are managed by multiple zones (e.g., multiple nodes, DMA and DMA32) * exposing that zone to user space via /sys/devices/system/memory/memory*/valid_zones Now that I identified some more cases where test_pages_in_a_zone() might go wrong, and we received an UBSAN report (see patch #3), let's get rid of this PFN walker. So instead of detecting the zone at runtime with test_pages_in_a_zone() by scanning the memmap, let's determine and remember for each memory block if it's managed by a single zone. The stored zone can then be used for the above two cases, avoiding a manual lookup using test_pages_in_a_zone(). This avoids eventually stumbling over uninitialized memmaps in corner cases, especially when ZONE_DEVICE ranges partly fall into memory block (that are responsible for managing System RAM). Handling memory onlining is easy, because we online to exactly one zone. Handling boot memory is more tricky, because we want to avoid scanning all zones of all nodes to detect possible zones that overlap with the physical memory region of interest. Fortunately, we already have code that determines the applicable nodes for a memory block, to create sysfs links -- we'll hook into that. Patch #1 is a simple cleanup I had laying around for a longer time. Patch #2 contains the main logic to remove test_pages_in_a_zone() and further details. [1] https://lkml.kernel.org/r/20220128144540.153902-1-david@redhat.com [2] https://lkml.kernel.org/r/20220203105212.30385-1-david@redhat.com This patch (of 2): Let's adjust the stale terminology, making it match unregister_memory_block_under_nodes() and do_register_memory_block_under_node(). We're dealing with memory block devices, which span 1..X memory sections. Link: https://lkml.kernel.org/r/20220210184359.235565-1-david@redhat.com Link: https://lkml.kernel.org/r/20220210184359.235565-2-david@redhat.comSigned-off-by:
-
Miaohe Lin authored
It's misplaced since commit 79605093 ("mm, memory_hotplug: print reason for the offlining failure"). Move it to the right place. Link: https://lkml.kernel.org/r/20220207133643.23427-5-linmiaohe@huawei.comSigned-off-by:
Miaohe Lin <linmiaohe@huawei.com> Reviewed-by:
-
Miaohe Lin authored
We can use helper macro node_spanned_pages to check whether node spans pages. And we can change the parameter of check_cpu_on_node to nid as that's what it really cares. Thus we can further get rid of the local variable pgdat and improve the readability a bit. Link: https://lkml.kernel.org/r/20220207133643.23427-4-linmiaohe@huawei.comSigned-off-by:
-
Miaohe Lin authored
If zid reaches ZONE_NORMAL, the caller will always get the NORMAL zone no matter what zone_intersects() returns. So we can save some possible cpu cycles by avoid calling zone_intersects() for ZONE_NORMAL. Link: https://lkml.kernel.org/r/20220207133643.23427-3-linmiaohe@huawei.comSigned-off-by:
-
Miaohe Lin authored
Patch series "A few cleanup patches around memory_hotplug". This series contains a few patches to fix obsolete and misplaced comments, clean up the try_offline_node function and so on. This patch (of 4): Since commit f1dd2cd1 ("mm, memory_hotplug: do not associate hotadded memory to zones until online"), there is no need to pass in the zone. [akpm@linux-foundation.org: remove the comment altogether, per David] Link: https://lkml.kernel.org/r/20220207133643.23427-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20220207133643.23427-2-linmiaohe@huawei.comSigned-off-by:
-
David Hildenbrand authored
... and call node_dev_init() after memory_dev_init() from driver_init(), so before any of the existing arch/subsys calls. All online nodes should be known at that point: early during boot, arch code determines node and zone ranges and sets the relevant nodes online; usually this happens in setup_arch(). This is in line with memory_dev_init(), which initializes the memory device subsystem and creates all memory block devices. Similar to memory_dev_init(), panic() if anything goes wrong, we don't want to continue with such basic initialization errors. The important part is that node_dev_init() gets called after memory_dev_init() and after cpu_dev_init(), but before any of the relevant archs call register_cpu() to register the new cpu device under the node device. The latter should be the case for the current users of topology_init(). Link: https://lkml.kernel.org/r/20220203105212.30385-1-david@redhat.comSigned-off-by:
-
David Hildenbrand authored
If register_memory() fails, we freed the memory block but already added the memory block to the group list, not good. Let's defer adding the block to the memory group to after registering the memory block device. We do handle it properly during unregister_memory(), but that's not called when the registration fails. Link: https://lkml.kernel.org/r/20220128144540.153902-1-david@redhat.com Fixes: 028fc57a ("drivers/base/memory: introduce "memory groups" to logically group memory blocks") Signed-off-by:
Michal Hocko <mhocko@suse.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Rafael J. Wysocki" <rafael@kernel.org> Signed-off-by:
-
Wei Yang authored
alloc_mem_cgroup_per_node_info is allocated for each possible node and this used to be a problem because !node_online nodes didn't have appropriate data structure allocated. This has changed by "mm: handle uninitialized numa nodes gracefully" so we can drop the special casing here. Link: https://lkml.kernel.org/r/20220127085305.20890-7-mhocko@kernel.orgSigned-off-by:
Wei Yang <richard.weiyang@gmail.com> Signed-off-by:
-
Michal Hocko authored
free_area_init_node is also called from memory less node initialization path (free_area_init_memoryless_node). It doesn't really make much sense to display the physical memory range for those nodes: Initmem setup node XX [mem 0x0000000000000000-0x0000000000000000] Instead be explicit that the node is memoryless: Initmem setup node XX as memoryless Link: https://lkml.kernel.org/r/20220127085305.20890-6-mhocko@kernel.orgSigned-off-by:
Rafael Aquini <raquini@redhat.com> Acked-by:
Mike Rapoport <rppt@linux.ibm.com> Reviewed-by:
-
Michal Hocko authored
When a !node_online node is brought up it needs a hotplug specific initialization because the node could be either uninitialized yet or it could have been recycled after previous hotremove. hotadd_init_pgdat is responsible for that. Internal pgdat state is initialized at two places currently - hotadd_init_pgdat - free_area_init_core_hotplug There is no real clear cut what should go where but this patch's chosen to move the whole internal state initialization into free_area_init_core_hotplug. hotadd_init_pgdat is still responsible to pull all the parts together - most notably to initialize zonelists because those depend on the overall topology. This patch doesn't introduce any functional change. Link: https://lkml.kernel.org/r/20220127085305.20890-5-mhocko@kernel.orgSigned-off-by:
-
Michal Hocko authored
Prior to "mm: handle uninitialized numa nodes gracefully" memory hotplug used to allocate pgdat when memory has been added to a node (hotadd_init_pgdat) arch_free_nodedata has been only used in the failure path because once the pgdat is exported (to be visible by NODA_DATA(nid)) it cannot really be freed because there is no synchronization available for that. pgdat is allocated for each possible nodes now so the memory hotplug doesn't need to do the ever use arch_free_nodedata so drop it. This patch doesn't introduce any functional change. Link: https://lkml.kernel.org/r/20220127085305.20890-4-mhocko@kernel.orgSigned-off-by:
-
Michal Hocko authored
We have had several reports [1][2][3] that page allocator blows up when an allocation from a possible node is requested. The underlying reason is that NODE_DATA for the specific node is not allocated. NUMA specific initialization is arch specific and it can vary a lot. E.g. x86 tries to initialize all nodes that have some cpu affinity (see init_cpu_to_node) but this can be insufficient because the node might be cpuless for example. One way to address this problem would be to check for !node_online nodes when trying to get a zonelist and silently fall back to another node. That is unfortunately adding a branch into allocator hot path and it doesn't handle any other potential NODE_DATA users. This patch takes a different approach (following a lead of [3]) and it pre allocates pgdat for all possible nodes in an arch indipendent code - free_area_init. All uninitialized nodes are treated as memoryless nodes. node_state of the node is not changed because that would lead to other side effects - e.g. sysfs representation of such a node and from past discussions [4] it is known that some tools might have problems digesting that. Newly allocated pgdat only gets a minimal initialization and the rest of the work is expected to be done by the memory hotplug - hotadd_new_pgdat (renamed to hotadd_init_pgdat). generic_alloc_nodedata is changed to use the memblock allocator because neither page nor slab allocators are available at the stage when all pgdats are allocated. Hotplug doesn't allocate pgdat anymore so we can use the early boot allocator. The only arch specific implementation is ia64 and that is changed to use the early allocator as well. [1] http://lkml.kernel.org/r/20211101201312.11589-1-amakhalov@vmware.com [2] http://lkml.kernel.org/r/20211207224013.880775-1-npache@redhat.com [3] http://lkml.kernel.org/r/20190114082416.30939-1-mhocko@kernel.org [4] http://lkml.kernel.org/r/20200428093836.27190-1-srikar@linux.vnet.ibm.com [akpm@linux-foundation.org: replace comment, per Mike] Link: https://lkml.kernel.org/r/Yfe7RBeLCijnWBON@dhcp22.suse.czReported-by:
Alexey Makhalov <amakhalov@vmware.com> Tested-by:
Nico Pache <npache@redhat.com> Acked-by:
-
Michal Hocko authored
Patch series "mm, memory_hotplug: handle unitialized numa node gracefully". The core of the fix is patch 2 which also links existing bug reports. The high level goal is to have all possible numa nodes have their pgdat allocated and initialized so for_each_possible_node(nid) NODE_DATA(nid) will never return garbage. This has proven to be problem in several places when an offline numa node is used for an allocation just to realize that node_data and therefore allocation fallback zonelists are not initialized and such an allocation request blows up. There were attempts to address that by checking node_online in several places including the page allocator. This patchset approaches the problem from a different perspective and instead of special casing, which just adds a runtime overhead, it allocates pglist_data for each possible node. This can add some memory overhead for platforms with high number of possible nodes if they do not contain any memory. This should be a rather rare configuration though. How to test this? David has provided and excellent howto: http://lkml.kernel.org/r/6e5ebc19-890c-b6dd-1924-9f25c441010d@redhat.com Patches 1 and 3-6 are mostly cleanups. The patchset has been reviewed by Rafael (thanks!) and the core fix tested by Rafael and Alexey (thanks to both). David has tested as per instructions above and hasn't found any fallouts in the memory hotplug scenarios. This patch (of 6): This is a preparatory patch and it doesn't introduce any functional change. It merely pulls out arch_alloc_nodedata (and co) outside of CONFIG_MEMORY_HOTPLUG because the following patch will need to call this from the generic MM code. Link: https://lkml.kernel.org/r/20220127085305.20890-1-mhocko@kernel.org Link: https://lkml.kernel.org/r/20220127085305.20890-2-mhocko@kernel.orgSigned-off-by:
-
Charan Teja Kalla authored
The process_madvise() system call is expected to skip holes in vma passed through 'struct iovec' vector list. But do_madvise, which process_madvise() calls for each vma, returns ENOMEM in case of unmapped holes, despite the VMA is processed. Thus process_madvise() should treat ENOMEM as expected and consider the VMA passed to as processed and continue processing other vma's in the vector list. Returning -ENOMEM to user, despite the VMA is processed, will be unable to figure out where to start the next madvise. Link: https://lkml.kernel.org/r/4f091776142f2ebf7b94018146de72318474e686.1647008754.git.quic_charante@quicinc.com Fixes: ecb8ac8b("mm/madvise: introduce process_madvise() syscall: an external memory hinting API") Signed-off-by:
Charan Teja Kalla <quic_charante@quicinc.com> Cc: David Rientjes <rientjes@google.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: <stable@vger.kernel.org> Signed-off-by:
-
Charan Teja Kalla authored
Patch series "mm: madvise: return correct bytes processed with process_madvise", v2. With the process_madvise(), always choose to return non zero processed bytes over an error. This can help the user to know on which VMA, passed in the 'struct iovec' vector list, is failed to advise thus can take the decission of retrying/skipping on that VMA. This patch (of 2): The process_madvise() system call returns error even after processing some VMA's passed in the 'struct iovec' vector list which leaves the user confused to know where to restart the advise next. It is also against this syscall man page[1] documentation where it mentions that "return value may be less than the total number of requested bytes, if an error occurred after some iovec elements were already processed.". Consider a user passed 10 VMA's in the 'struct iovec' vector list of which 9 are processed but one. Then it just returns the error caused on that failed VMA despite the first 9 VMA's processed, leaving the user confused about on which VMA it is failed. Returning the number of bytes processed here can help the user to know which VMA it is failed on and thus can retry/skip the advise on that VMA. [1]https://man7.org/linux/man-pages/man2/process_madvise.2.html. Link: https://lkml.kernel.org/r/cover.1647008754.git.quic_charante@quicinc.com Link: https://lkml.kernel.org/r/125b61a0edcee5c2db8658aed9d06a43a19ccafc.1647008754.git.quic_charante@quicinc.com Fixes: ecb8ac8b("mm/madvise: introduce process_madvise() syscall: an external memory hinting API") Signed-off-by:
-
Miaohe Lin authored
Using vma_lookup() verifies the start address is contained in the found vma. This results in easier to read the code. Link: https://lkml.kernel.org/r/20220311082731.63513-1-linmiaohe@huawei.comSigned-off-by:
-
Matthew Wilcox (Oracle) authored
Hardware poison is tracked on a per-page basis, not on the head page. Link: https://lkml.kernel.org/r/20220130013042.1906881-1-willy@infradead.orgSigned-off-by:
Matthew Wilcox (Oracle) <willy@infradead.org> Acked-by:
Naoya Horiguchi <naoya.horiguchi@nec.com> Reviewed-by:
Yang Shi <shy828301@gmail.com> Cc: David Rientjes <rientjes@google.com> Cc: Mike Kravetz <mike.kravetz@oracle.com> Signed-off-by:
-
Miaohe Lin authored
Use helper macro __ATTR_RW to define KSM_ATTR to make code more clear. Minor readability improvement. Link: https://lkml.kernel.org/r/20220221115809.26381-1-linmiaohe@huawei.comSigned-off-by:
-
Yang Yang authored
When faults in from swap what used to be a KSM page and that page had been swapped in before, system has to make a copy, and leaves remerging the pages to a later pass of ksmd. That is not good for performace, we'd better to reduce this kind of copy. There are some ways to reduce it, for example lessen swappiness or madvise(, , MADV_MERGEABLE) range. So add this event to support doing this tuning. Just like this patch: "mm, THP, swap: add THP swapping out fallback counting". Link: https://lkml.kernel.org/r/20220113023839.758845-1-yang.yang29@zte.com.cnSigned-off-by:
Yang Yang <yang.yang29@zte.com.cn> Reviewed-by:
Ran Xiaokai <ran.xiaokai@zte.com.cn> Cc: Hugh Dickins <hughd@google.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Saravanan D <saravanand@fb.com> Signed-off-by:
-
Johannes Weiner authored
Once upon a time, all swapins counted toward memory pressure[1]. Then Joonsoo introduced workingset detection for anonymous pages and we gained the ability to distinguish hot from cold swapins[2][3]. But we failed to update swap_readpage() accordingly, and now we account partial memory pressure in the swapin path of cold memory. Not for all situations - which adds more inconsistency: paths using the conventional submit_bio() and lock_page() route will not see much pressure - unless storage itself is heavily congested and the bio submissions stall. ZRAM and ZSWAP do most of the work directly from swap_readpage() and will see all swapins reflected as pressure. IOW, a workload doing cold swapins could see little to no pressure reported with on-disk swap, but potentially high pressure with a zram or zswap backend. That confuses any psi-based health monitoring, load shedding, proactive reclaim, or userspace OOM killing schemes that might be in place for the workload. Restore consistency by making all swapin stall accounting conditional on the page actually being part of the workingset. [1] commit 93779069 ("mm/page_io.c: annotate refault stalls from swap_readpage") [2] commit aae466b0 ("mm/swap: implement workingset detection for anonymous LRU") [3] commit cad8320b ("mm/swap: don't SetPageWorkingset unconditionally during swapin") Link: https://lkml.kernel.org/r/20220214214921.419687-1-hannes@cmpxchg.orgSigned-off-by:
Johannes Weiner <hannes@cmpxchg.org> Reported-by:
CGEL <cgel.zte@gmail.com> Acked-by:
Minchan Kim <minchan@kernel.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by:
-
Huang Ying authored
If the NUMA balancing isn't used to optimize the page placement among sockets but only among memory types, the hot pages in the fast memory node couldn't be migrated (promoted) to anywhere. So it's unnecessary to scan the pages in the fast memory node via changing their PTE/PMD mapping to be PROT_NONE. So that the page faults could be avoided too. In the test, if only the memory tiering NUMA balancing mode is enabled, the number of the NUMA balancing hint faults for the DRAM node is reduced to almost 0 with the patch. While the benchmark score doesn't change visibly. Link: https://lkml.kernel.org/r/20220221084529.1052339-4-ying.huang@intel.comSigned-off-by:
"Huang, Ying" <ying.huang@intel.com> Suggested-by:
Dave Hansen <dave.hansen@linux.intel.com> Tested-by:
-
Huang Ying authored
With the advent of various new memory types, some machines will have multiple types of memory, e.g. DRAM and PMEM (persistent memory). The memory subsystem of these machines can be called memory tiering system, because the performance of the different types of memory are usually different. In such system, because of the memory accessing pattern changing etc, some pages in the slow memory may become hot globally. So in this patch, the NUMA balancing mechanism is enhanced to optimize the page placement among the different memory types according to hot/cold dynamically. In a typical memory tiering system, there are CPUs, fast memory and slow memory in each physical NUMA node. The CPUs and the fast memory will be put in one logical node (called fast memory node), while the slow memory will be put in another (faked) logical node (called slow memory node). That is, the fast memory is regarded as local while the slow memory is regarded as remote. So it's possible for the recently accessed pages in the slow memory node to be promoted to the fast memory node via the existing NUMA balancing mechanism. The original NUMA balancing mechanism will stop to migrate pages if the free memory of the target node becomes below the high watermark. This is a reasonable policy if there's only one memory type. But this makes the original NUMA balancing mechanism almost do not work to optimize page placement among different memory types. Details are as follows. It's the common cases that the working-set size of the workload is larger than the size of the fast memory nodes. Otherwise, it's unnecessary to use the slow memory at all. So, there are almost always no enough free pages in the fast memory nodes, so that the globally hot pages in the slow memory node cannot be promoted to the fast memory node. To solve the issue, we have 2 choices as follows, a. Ignore the free pages watermark checking when promoting hot pages from the slow memory node to the fast memory node. This will create some memory pressure in the fast memory node, thus trigger the memory reclaiming. So that, the cold pages in the fast memory node will be demoted to the slow memory node. b. Define a new watermark called wmark_promo which is higher than wmark_high, and have kswapd reclaiming pages until free pages reach such watermark. The scenario is as follows: when we want to promote hot-pages from a slow memory to a fast memory, but fast memory's free pages would go lower than high watermark with such promotion, we wake up kswapd with wmark_promo watermark in order to demote cold pages and free us up some space. So, next time we want to promote hot-pages we might have a chance of doing so. The choice "a" may create high memory pressure in the fast memory node. If the memory pressure of the workload is high, the memory pressure may become so high that the memory allocation latency of the workload is influenced, e.g. the direct reclaiming may be triggered. The choice "b" works much better at this aspect. If the memory pressure of the workload is high, the hot pages promotion will stop earlier because its allocation watermark is higher than that of the normal memory allocation. So in this patch, choice "b" is implemented. A new zone watermark (WMARK_PROMO) is added. Which is larger than the high watermark and can be controlled via watermark_scale_factor. In addition to the original page placement optimization among sockets, the NUMA balancing mechanism is extended to be used to optimize page placement according to hot/cold among different memory types. So the sysctl user space interface (numa_balancing) is extended in a backward compatible way as follow, so that the users can enable/disable these functionality individually. The sysctl is converted from a Boolean value to a bits field. The definition of the flags is, - 0: NUMA_BALANCING_DISABLED - 1: NUMA_BALANCING_NORMAL - 2: NUMA_BALANCING_MEMORY_TIERING We have tested the patch with the pmbench memory accessing benchmark with the 80:20 read/write ratio and the Gauss access address distribution on a 2 socket Intel server with Optane DC Persistent Memory Model. The test results shows that the pmbench score can improve up to 95.9%. Thanks Andrew Morton to help fix the document format error. Link: https://lkml.kernel.org/r/20220221084529.1052339-3-ying.huang@intel.comSigned-off-by:
-
Huang Ying authored
Patch series "NUMA balancing: optimize memory placement for memory tiering system", v13 With the advent of various new memory types, some machines will have multiple types of memory, e.g. DRAM and PMEM (persistent memory). The memory subsystem of these machines can be called memory tiering system, because the performance of the different types of memory are different. After commit c221c0b0 ("device-dax: "Hotplug" persistent memory for use like normal RAM"), the PMEM could be used as the cost-effective volatile memory in separate NUMA nodes. In a typical memory tiering system, there are CPUs, DRAM and PMEM in each physical NUMA node. The CPUs and the DRAM will be put in one logical node, while the PMEM will be put in another (faked) logical node. To optimize the system overall performance, the hot pages should be placed in DRAM node. To do that, we need to identify the hot pages in the PMEM node and migrate them to DRAM node via NUMA migration. In the original NUMA balancing, there are already a set of existing mechanisms to identify the pages recently accessed by the CPUs in a node and migrate the pages to the node. So we can reuse these mechanisms to build the mechanisms to optimize the page placement in the memory tiering system. This is implemented in this patchset. At the other hand, the cold pages should be placed in PMEM node. So, we also need to identify the cold pages in the DRAM node and migrate them to PMEM node. In commit 26aa2d19 ("mm/migrate: demote pages during reclaim"), a mechanism to demote the cold DRAM pages to PMEM node under memory pressure is implemented. Based on that, the cold DRAM pages can be demoted to PMEM node proactively to free some memory space on DRAM node to accommodate the promoted hot PMEM pages. This is implemented in this patchset too. We have tested the solution with the pmbench memory accessing benchmark with the 80:20 read/write ratio and the Gauss access address distribution on a 2 socket Intel server with Optane DC Persistent Memory Model. The test results shows that the pmbench score can improve up to 95.9%. This patch (of 3): In a system with multiple memory types, e.g. DRAM and PMEM, the CPU and DRAM in one socket will be put in one NUMA node as before, while the PMEM will be put in another NUMA node as described in the description of the commit c221c0b0 ("device-dax: "Hotplug" persistent memory for use like normal RAM"). So, the NUMA balancing mechanism will identify all PMEM accesses as remote access and try to promote the PMEM pages to DRAM. To distinguish the number of the inter-type promoted pages from that of the inter-socket migrated pages. A new vmstat count is added. The counter is per-node (count in the target node). So this can be used to identify promotion imbalance among the NUMA nodes. Link: https://lkml.kernel.org/r/20220301085329.3210428-1-ying.huang@intel.com Link: https://lkml.kernel.org/r/20220221084529.1052339-1-ying.huang@intel.com Link: https://lkml.kernel.org/r/20220221084529.1052339-2-ying.huang@intel.comSigned-off-by:
-
Hari Bathini authored
With commit a4e92ce8 ("powerpc/fadump: Reservationless firmware assisted dump"), Linux kernel's Contiguous Memory Allocator (CMA) based reservation was introduced in fadump. That change was aimed at using CMA to let applications utilize the memory reserved for fadump while blocking it from being used for kernel pages. The assumption was, even if CMA activation fails for whatever reason, the memory still remains reserved to avoid it from being used for kernel pages. But commit 072355c1 ("mm/cma: expose all pages to the buddy if activation of an area fails") breaks this assumption as it started exposing all pages to buddy allocator on CMA activation failure. It led to warning messages like below while running crash-utility on vmcore of a kernel having above two commits: crash: seek error: kernel virtual address: <from reserved region> To fix this problem, opt out from exposing pages to buddy allocator on CMA activation failure for fadump reserved memory. Link: https://lkml.kernel.org/r/20220117075246.36072-3-hbathini@linux.ibm.comSigned-off-by:
Hari Bathini <hbathini@linux.ibm.com> Acked-by:
Michael Ellerman <mpe@ellerman.id.au> Cc: Mahesh Salgaonkar <mahesh@linux.ibm.com> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Oscar Salvador <osalvador@suse.de> Cc: Sourabh Jain <sourabhjain@linux.ibm.com> Signed-off-by:
-
Hari Bathini authored
Patch series "powerpc/fadump: handle CMA activation failure appropriately", v3. Commit 072355c1 ("mm/cma: expose all pages to the buddy if activation of an area fails") started exposing all pages to buddy allocator on CMA activation failure. But there can be CMA users that want to handle the reserved memory differently on CMA allocation failure. Provide an option to opt out from exposing pages to buddy for such cases. Link: https://lkml.kernel.org/r/20220117075246.36072-1-hbathini@linux.ibm.com Link: https://lkml.kernel.org/r/20220117075246.36072-2-hbathini@linux.ibm.comSigned-off-by:
-
Hugh Dickins authored
Migration entries do not contribute to a page's reference count: move __split_huge_pmd_locked()'s page_ref_add() into pmd_migration's else block (along with the page_count() check - a page is quite likely to have reference count frozen to 0 when a migration entry is found). This will fix a very rare anonymous memory leak, after a split_huge_pmd() raced with an anon split_huge_page() or an anon THP migrate_pages(): since the wrongly raised refcount stopped the page (perhaps small, perhaps huge, depending on when the race hit) from ever being freed. At first I thought there were worse risks, from prematurely unfreezing a frozen page: but now think that would only affect page cache pages, which do not come this way (except for anonymous pages in swap cache, perhaps). Link: https://lkml.kernel.org/r/84792468-f512-e48f-378c-e34c3641e97@google.com Fixes: ec0abae6 ("mm/thp: fix __split_huge_pmd_locked() for migration PMD") Signed-off-by:
Hugh Dickins <hughd@google.com> Reviewed-by:
-
andrew.yang authored
When memory is tight, system may start to compact memory for large continuous memory demands. If one process tries to lock a memory page that is being locked and isolated for compaction, it may wait a long time or even forever. This is because compaction will perform non-atomic PG_Isolated clear while holding page lock, this may overwrite PG_waiters set by the process that can't obtain the page lock and add itself to the waiting queue to wait for the lock to be unlocked. CPU1 CPU2 lock_page(page); (successful) lock_page(); (failed) __ClearPageIsolated(page); SetPageWaiters(page) (may be overwritten) unlock_page(page); The solution is to not perform non-atomic operation on page flags while holding page lock. Link: https://lkml.kernel.org/r/20220315030515.20263-1-andrew.yang@mediatek.comSigned-off-by:andrew.yang <andrew.yang@mediatek.com> Cc: Matthias Brugger <matthias.bgg@gmail.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: "Vlastimil Babka" <vbabka@suse.cz> Cc: David Howells <dhowells@redhat.com> Cc: "William Kucharski" <william.kucharski@oracle.com> Cc: David Hildenbrand <david@redhat.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Marc Zyngier <maz@kernel.org> Cc: Nicholas Tang <nicholas.tang@mediatek.com> Cc: Kuan-Ying Lee <Kuan-Ying.Lee@mediatek.com> Signed-off-by:
-
Huang Ying authored
In commit ac16ec83 ("mm: migrate: support multiple target nodes demotion"), after the first demotion target node is found, we will continue to check the next candidate obtained via find_next_best_node(). This is to find all demotion target nodes with same NUMA distance. But one side effect of find_next_best_node() is that the candidate node returned will be set in "used" parameter, even if the candidate node isn't passed in the following NUMA distance checking, the candidate node will not be used as demotion target node for the following nodes. For example, for system as follows, node distances: node 0 1 2 3 0: 10 21 17 28 1: 21 10 28 17 2: 17 28 10 28 3: 28 17 28 10 when we establish demotion target node for node 0, in the first round node 2 is added to the demotion target node set. Then in the second round, node 3 is checked and failed because distance(0, 3) > distance(0, 2). But node 3 is set in "used" nodemask too. When we establish demotion target node for node 1, there is no available node. This is wrong, node 3 should be set as the demotion target of node 1. To fix this, if the candidate node is failed to pass the distance checking, it will be cleared in "used" nodemask. So that it can be used for the following node. The bug can be reproduced and fixed with this patch on a 2 socket server machine with DRAM and PMEM. Link: https://lkml.kernel.org/r/20220128055940.1792614-1-ying.huang@intel.com Fixes: ac16ec83 ("mm: migrate: support multiple target nodes demotion") Signed-off-by:
-
Miaohe Lin authored
oom_cpuset_eligible() is always called when !is_memcg_oom(). Remove this unnecessary check. Link: https://lkml.kernel.org/r/20220224115933.20154-1-linmiaohe@huawei.comSigned-off-by:
David Rientjes <rientjes@google.com> Acked-by:
-
Hugh Dickins authored
v2.6.34 commit 9d8cebd4 ("mm: fix mbind vma merge problem") introduced vma_merge() to mbind_range(); but unlike madvise, mlock and mprotect, it put a "continue" to next vma where its precedents go to update flags on current vma before advancing: that left vma with the wrong setting in the infamous vma_merge() case 8. v3.10 commit 1444f92c ("mm: merging memory blocks resets mempolicy") tried to fix that in vma_adjust(), without fully understanding the issue. v3.11 commit 3964acd0 ("mm: mempolicy: fix mbind_range() && vma_adjust() interaction") reverted that, and went about the fix in the right way, but chose to optimize out an unnecessary mpol_dup() with a prior mpol_equal() test. But on tmpfs, that also pessimized out the vital call to its ->set_policy(), leaving the new mbind unenforced. The user visible effect was that the pages got allocated on the local node (happened to be 0), after the mbind() caller had specifically asked for them to be allocated on node 1. There was not any page migration involved in the case reported: the pages simply got allocated on the wrong node. Just delete that optimization now (though it could be made conditional on vma not having a set_policy). Also remove the "next" variable: it turned out to be blameless, but also pointless. Link: https://lkml.kernel.org/r/319e4db9-64ae-4bca-92f0-ade85d342ff@google.com Fixes: 3964acd0 ("mm: mempolicy: fix mbind_range() && vma_adjust() interaction") Signed-off-by:
Oleg Nesterov <oleg@redhat.com> Reviewed-by:
Liam R. Howlett <Liam.Howlett@oracle.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: <stable@vger.kernel.org> Signed-off-by:
-
Baolin Wang authored
As Steven suggested [1], we should access the pointers from the trace event to avoid dereferencing them to the tracepoint function when the tracepoint is disabled. [1] https://lkml.org/lkml/2021/11/3/409 Link: https://lkml.kernel.org/r/4cd393b4d57f8f01ed72c001509b28e3a3b1a8c1.1646985115.git.baolin.wang@linux.alibaba.comSigned-off-by:
-
Charan Teja Kalla authored
Commit b518154e ("mm/vmscan: protect the workingset on anonymous LRU") requires to look twice for both mapped anon/file pages are used more than once to take the decission of reclaim or activation. Correct the documentation accordingly. Link: https://lkml.kernel.org/r/1646925640-21324-1-git-send-email-quic_charante@quicinc.comSigned-off-by:
-
Sebastian Andrzej Siewior authored
Commit 68d48e6a ("mm: workingset: add vmstat counter for shadow nodes") introduced an IRQ-off check to ensure that a lock is held which also disabled interrupts. This does not work the same way on PREEMPT_RT because none of the locks, that are held, disable interrupts. Replace this check with a lockdep assert which ensures that the lock is held. Link: https://lkml.kernel.org/r/20220301122143.1521823-3-bigeasy@linutronix.deSigned-off-by:
Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tejun Heo <tj@kernel.org> Cc: Zefan Li <lizefan.x@bytedance.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by:
-
Marcelo Tosatti authored
On systems that run FIFO:1 applications that busy loop, any SCHED_OTHER task that attempts to execute on such a CPU (such as work threads) will not be scheduled, which leads to system hangs. Commit d479960e ("mm: disable LRU pagevec during the migration temporarily") relies on queueing work items on all online CPUs to ensure visibility of lru_disable_count. To fix this, replace the usage of work items with synchronize_rcu, which provides the same guarantees. Readers of lru_disable_count are protected by either disabling preemption or rcu_read_lock: preempt_disable, local_irq_disable [bh_lru_lock()] rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT] preempt_disable [local_lock !CONFIG_PREEMPT_RT] Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on preempt_disable() regions of code. So any CPU which sees lru_disable_count = 0 will have exited the critical section when synchronize_rcu() returns. Link: https://lkml.kernel.org/r/Yin7hDxdt0s/x+fp@fuller.cnetSigned-off-by:
Marcelo Tosatti <mtosatti@redhat.com> Reviewed-by:
Nicolas Saenz Julienne <nsaenzju@redhat.com> Acked-by:
-
Waiman Long authored
Since commit 2c80cd57 ("mm/list_lru.c: fix list_lru_count_node() to be race free"), we are tracking the total number of lru entries in a list_lru_node in its nr_items field. In the case of memcg_reparent_list_lru_node(), there is nothing to be done if nr_items is 0. We don't even need to take the nlru->lock as no new lru entry could be added by a racing list_lru_add() to the draining src_idx memcg at this point. On systems that serve a lot of containers, it is possible that there can be thousands of list_lru's present due to the fact that each container may mount its own container specific filesystems. As a typical container uses only a few cpus, it is likely that only the list_lru_node that contains those cpus will be utilized while the rests may be empty. In other words, there can be a lot of list_lru_node with 0 nr_items. By skipping a lock/unlock operation and loading a cacheline from memcg_lrus, a sizeable number of cpu cycles can be saved. That can be substantial if we are talking about thousands of list_lru_node's with 0 nr_items. Link: https://lkml.kernel.org/r/20220309144000.1470138-1-longman@redhat.comSigned-off-by:
Waiman Long <longman@redhat.com> Reviewed-by:
Roman Gushchin <roman.gushchin@linux.dev> Cc: Muchun Song <songmuchun@bytedance.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shakeel Butt <shakeelb@google.com> Signed-off-by:
-
Hugh Dickins authored
__isolate_lru_page_prepare() conflates two unrelated functions, with the flags to one disjoint from the flags to the other; and hides some of the important checks outside of isolate_migratepages_block(), where the sequence is better to be visible. It comes from the days of lumpy reclaim, before compaction, when the combination made more sense. Move what's needed by mm/compaction.c isolate_migratepages_block() inline there, and what's needed by mm/vmscan.c isolate_lru_pages() inline there. Shorten "isolate_mode" to "mode", so the sequence of conditions is easier to read. Declare a "mapping" variable, to save one call to page_mapping() (but not another: calling again after page is locked is necessary). Simplify isolate_lru_pages() with a "move_to" list pointer. Link: https://lkml.kernel.org/r/879d62a8-91cc-d3c6-fb3b-69768236df68@google.comSigned-off-by:
Alex Shi <alexs@kernel.org> Cc: Alexander Duyck <alexander.duyck@gmail.com> Signed-off-by:
-
Hugh Dickins authored
PF_SWAPWRITE has been redundant since v3.2 commit ee72886d ("mm: vmscan: do not writeback filesystem pages in direct reclaim"). Coincidentally, NeilBrown's current patch "remove inode_congested()" deletes may_write_to_inode(), which appeared to be the one function which took notice of PF_SWAPWRITE. But if you study the old logic, and the conditions under which may_write_to_inode() was called, you discover that flag and function have been pointless for a decade. Link: https://lkml.kernel.org/r/75e80e7-742d-e3bd-531-614db8961e4@google.comSigned-off-by:
-
