arm64 2M大页 kvm 页表建立 dcache刷到内存空洞宕机 · Pull Request !2247 · openEuler/kernel

https://gitee.com/openeuler/kernel/issues/I82UA0?from=project-issue

dmesg

```
[ 1596.322299] Unable to handle kernel paging request at virtual address ffff9dcb00000000
[ 1596.330664] Mem abort info:
[ 1596.333879]   ESR = 0x96000146
[ 1596.337381]   Exception class = DABT (current EL), IL = 32 bits
[ 1596.343720]   SET = 0, FnV = 0
[ 1596.347211]   EA = 0, S1PTW = 0
[ 1596.350776] Data abort info:
[ 1596.354077]   ISV = 0, ISS = 0x00000146
[ 1596.358346]   CM = 1, WnR = 1
[ 1596.361737] swapper pgtable: 64k pages, 48-bit VAs, pgdp = 00000000f95ddb1d
[ 1596.369114] [ffff9dcb00000000] pgd=0000037bfffc0803, pud=0000037bfffc0803, pmd=0000000000000000
[ 1596.378224] Internal error: Oops: 96000146 [#1] SMP
...
[ 1596.478233] Process kvm (pid: 101049, stack limit = 0x00000000be0eeeea)
[ 1596.485253] CPU: 57 PID: 101049 Comm: kvm Kdump: loaded Tainted: G           OE     4.19.90-25.30.v2101.osc.sfc.6.9.1.0031.ky10.aarch64 #1
[ 1596.498056] Hardware name: O.E.M RSE3231/EAT1720B1, BIOS 1.0.10 30/12/2021 15:30:23
[ 1596.506109] pstate: 80000005 (Nzcv daif -PAN -UAO)
[ 1596.511320] pc : __flush_dcache_area+0x20/0x38
[ 1596.516178] lr : clean_dcache_guest_page+0x54/0x78
[ 1596.521377] sp : ffff9da2c68f7900
[ 1596.525108] x29: ffff9da2c68f7900 x28: 0000000000000cc6
[ 1596.530829] x27: 00000003a0010000 x26: 0000000000200000
[ 1596.536548] x25: 0000000000000000 x24: 0000fffbd1a10000
[ 1596.542267] x23: 000000000173ffe1 x22: ffff9dd2c9e20008
[ 1596.547986] x21: ffff9dc2d059ce50 x20: 0000000000200000
[ 1596.553705] x19: 000000000173ffe1 x18: 0000000000000000
[ 1596.559423] x17: 1dffffc000008018 x16: 0000000000000000
[ 1596.565140] x15: 0000000000000000 x14: 0000000000000000
[ 1596.570857] x13: ffff7fe772bff800 x12: ffff4dffa9cdc000
[ 1596.576576] x11: 0000000000001fff x10: 0000000772bff800
[ 1596.582295] x9 : 0000000000000001 x8 : 0000000000000000
[ 1596.588013] x7 : 0000000400000000 x6 : 0000000000000018
[ 1596.593732] x5 : ffff4dffaa9c0598 x4 : ffff4dffaa9c01d8
[ 1596.599449] x3 : 000000000000003f x2 : 0000000000000040
[ 1596.605167] x1 : ffff9dcb00010000 x0 : ffff9dcb00000000
[ 1596.610886] Call trace:
[ 1596.613754]  __flush_dcache_area+0x20/0x38
[ 1596.618263]  user_mem_abort+0x8bc/0xab8
[ 1596.622513]  kvm_handle_guest_abort+0x198/0x460
[ 1596.627455]  handle_exit+0x13c/0x300
[ 1596.631446]  kvm_arch_vcpu_ioctl_run+0x24c/0x800
[ 1596.636475]  kvm_vcpu_ioctl+0x460/0x878
[ 1596.640729]  do_vfs_ioctl+0xb0/0x898
[ 1596.644720]  ksys_ioctl+0x8c/0xa0
[ 1596.648452]  __arm64_sys_ioctl+0x28/0x98
[ 1596.652793]  el0_svc_common+0x78/0x130
[ 1596.656956]  el0_svc_handler+0x38/0x78
[ 1596.661122]  el0_svc+0x8/0x1b0
[ 1596.664596] Code: 9ac32042 8b010001 d1000443 8a230000 (d50b7e20)
[ 1596.671203] SMP: stopping secondary CPUs
[ 1596.687971] Starting crashdump kernel...
[ 1596.692312] Bye!
```

问题原因:
user_mem_abort()是为虚拟机建立内存页表映射的(HVA->HPA)，即stage-2页表建立
对于arm64，stage-2页表只支持PUD、PMD和普通页大小
对于PAGE_SIZE=64K时的2M hugepage情况，vma_pagesize取得2M，映射处理不正确
导致clean_dcache_guest_page()越界到内存空洞而触发异常

问题修复:

1. upstream commit 2f40c46021bbb3ecd5c5f05764ecccbc276bc690

对于2M大页，会fallback到普通页大小建立stage-2页表，dcache也会变为普通页大小
    backport适配，定义自己的UMA_CONT_{PMD,PTE}_SHIFT

2. 依赖 upstream commit 6d674e28f642e3ff676fbae2d8d1b872814d32b6

pfn如果无效(is_iomap())，则不进行dcache和icache的操作

GVPopenEuler/kernel

搜索帮助

GVP openEuler/kernel