summaryrefslogtreecommitdiff
path: root/arch/x86/include
diff options
context:
space:
mode:
authorOmar Sandoval <osandov@fb.com>2022-04-15 05:14:01 +0300
committerLinus Torvalds <torvalds@linux-foundation.org>2022-04-16 00:49:56 +0300
commitc12cd77cb028255663810e6c4528f0325facff66 (patch)
tree668b9fc6442d7d27c19d6c33b98cfe08e767e4e3 /arch/x86/include
parentaeb7923733d100b86c6bc68e7ae32913b0cec9d8 (diff)
downloadlinux-c12cd77cb028255663810e6c4528f0325facff66.tar.xz
mm/vmalloc: fix spinning drain_vmap_work after reading from /proc/vmcore
Commit 3ee48b6af49c ("mm, x86: Saving vmcore with non-lazy freeing of vmas") introduced set_iounmap_nonlazy(), which sets vmap_lazy_nr to lazy_max_pages() + 1, ensuring that any future vunmaps() immediately purge the vmap areas instead of doing it lazily. Commit 690467c81b1a ("mm/vmalloc: Move draining areas out of caller context") moved the purging from the vunmap() caller to a worker thread. Unfortunately, set_iounmap_nonlazy() can cause the worker thread to spin (possibly forever). For example, consider the following scenario: 1. Thread reads from /proc/vmcore. This eventually calls __copy_oldmem_page() -> set_iounmap_nonlazy(), which sets vmap_lazy_nr to lazy_max_pages() + 1. 2. Then it calls free_vmap_area_noflush() (via iounmap()), which adds 2 pages (one page plus the guard page) to the purge list and vmap_lazy_nr. vmap_lazy_nr is now lazy_max_pages() + 3, so the drain_vmap_work is scheduled. 3. Thread returns from the kernel and is scheduled out. 4. Worker thread is scheduled in and calls drain_vmap_area_work(). It frees the 2 pages on the purge list. vmap_lazy_nr is now lazy_max_pages() + 1. 5. This is still over the threshold, so it tries to purge areas again, but doesn't find anything. 6. Repeat 5. If the system is running with only one CPU (which is typicial for kdump) and preemption is disabled, then this will never make forward progress: there aren't any more pages to purge, so it hangs. If there is more than one CPU or preemption is enabled, then the worker thread will spin forever in the background. (Note that if there were already pages to be purged at the time that set_iounmap_nonlazy() was called, this bug is avoided.) This can be reproduced with anything that reads from /proc/vmcore multiple times. E.g., vmcore-dmesg /proc/vmcore. It turns out that improvements to vmap() over the years have obsoleted the need for this "optimization". I benchmarked `dd if=/proc/vmcore of=/dev/null` with 4k and 1M read sizes on a system with a 32GB vmcore. The test was run on 5.17, 5.18-rc1 with a fix that avoided the hang, and 5.18-rc1 with set_iounmap_nonlazy() removed entirely: |5.17 |5.18+fix|5.18+removal 4k|40.86s| 40.09s| 26.73s 1M|24.47s| 23.98s| 21.84s The removal was the fastest (by a wide margin with 4k reads). This patch removes set_iounmap_nonlazy(). Link: https://lkml.kernel.org/r/52f819991051f9b865e9ce25605509bfdbacadcd.1649277321.git.osandov@fb.com Fixes: 690467c81b1a ("mm/vmalloc: Move draining areas out of caller context") Signed-off-by: Omar Sandoval <osandov@fb.com> Acked-by: Chris Down <chris@chrisdown.name> Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Acked-by: Baoquan He <bhe@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'arch/x86/include')
-rw-r--r--arch/x86/include/asm/io.h2
1 files changed, 0 insertions, 2 deletions
diff --git a/arch/x86/include/asm/io.h b/arch/x86/include/asm/io.h
index f6d91ecb8026..e9736af126b2 100644
--- a/arch/x86/include/asm/io.h
+++ b/arch/x86/include/asm/io.h
@@ -210,8 +210,6 @@ void __iomem *ioremap(resource_size_t offset, unsigned long size);
extern void iounmap(volatile void __iomem *addr);
#define iounmap iounmap
-extern void set_iounmap_nonlazy(void);
-
#ifdef __KERNEL__
void memcpy_fromio(void *, const volatile void __iomem *, size_t);