| Message ID | 20240611215544.2105970-2-jiaqiyan@google.com |
|---|---|
| State | New |
| Headers | show |
| Series | Userspace controls soft-offline pages | expand |
On 2024/6/12 5:55, Jiaqi Yan wrote: > Correctable memory errors are very common on servers with large > amount of memory, and are corrected by ECC. Soft offline is kernel's > additional recovery handling for memory pages having (excessive) > corrected memory errors. Impacted page is migrated to a healthy page > if inuse; the original page is discarded for any future use. > Thanks for your update. > The actual policy on whether (and when) to soft offline should be > maintained by userspace, especially in case of an 1G HugeTLB page. > Soft-offline dissolves the HugeTLB page, either in-use or free, into > chunks of 4K pages, reducing HugeTLB pool capacity by 1 hugepage. > If userspace has not acknowledged such behavior, it may be surprised > when later mmap hugepages MAP_FAILED due to lack of hugepages. s/mmap hugepages MAP_FAILED/fails to mmap hugepages/ ? > In case of a transparent hugepage, it will be split into 4K pages > as well; userspace will stop enjoying the transparent performance. > > In addition, discarding the entire 1G HugeTLB page only because of > corrected memory errors sounds very costly and kernel better not > doing under the hood. But today there are at least 2 such cases: s/doing/doing so/ ? > 1. GHES driver sees both GHES_SEV_CORRECTED and > CPER_SEC_ERROR_THRESHOLD_EXCEEDED after parsing CPER. > 2. RAS Correctable Errors Collector counts correctable errors per > PFN and when the counter for a PFN reaches threshold > In both cases, userspace has no control of the soft offline performed > by kernel's memory failure recovery. > > This commit gives userspace the control of softofflining any page: > kernel only soft offlines raw page / transparent hugepage / HugeTLB > hugepage if userspace has agreed to. The interface to userspace is a > new sysctl called enable_soft_offline under /proc/sys/vm. By default > enable_soft_line is 1 to preserve existing behavior in kernel. s/enable_soft_line/enable_soft_offline/ > > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com> > --- > mm/memory-failure.c | 16 ++++++++++++++++ > 1 file changed, 16 insertions(+) > > diff --git a/mm/memory-failure.c b/mm/memory-failure.c > index d3c830e817e3..23415fe03318 100644 > --- a/mm/memory-failure.c > +++ b/mm/memory-failure.c > @@ -68,6 +68,8 @@ static int sysctl_memory_failure_early_kill __read_mostly; > > static int sysctl_memory_failure_recovery __read_mostly = 1; > > +static int sysctl_enable_soft_offline __read_mostly = 1; > + > atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); > > static bool hw_memory_failure __read_mostly = false; > @@ -141,6 +143,15 @@ static struct ctl_table memory_failure_table[] = { > .extra1 = SYSCTL_ZERO, > .extra2 = SYSCTL_ONE, > }, > + { > + .procname = "enable_soft_offline", > + .data = &sysctl_enable_soft_offline, > + .maxlen = sizeof(sysctl_enable_soft_offline), > + .mode = 0644, > + .proc_handler = proc_dointvec_minmax, > + .extra1 = SYSCTL_ZERO, > + .extra2 = SYSCTL_ONE, > + } > }; > > /* > @@ -2771,6 +2782,11 @@ int soft_offline_page(unsigned long pfn, int flags) > bool try_again = true; > struct page *page; > > + if (!sysctl_enable_soft_offline) { > + pr_info("soft offline: %#lx: OS-wide disabled\n", pfn); > + return -EINVAL; > + } > + IMHO, callers might reach here with page refcnt increased. So we have to take care of releasing it first? Also will it be better to return -EOPNOTSUPP or some other better errno? Thanks. .
Hi Jiaqi, On Wed, Jun 12, 2024 at 5:56 AM Jiaqi Yan <jiaqiyan@google.com> wrote: > > Correctable memory errors are very common on servers with large > amount of memory, and are corrected by ECC. Soft offline is kernel's > additional recovery handling for memory pages having (excessive) > corrected memory errors. Impacted page is migrated to a healthy page > if inuse; the original page is discarded for any future use. > > The actual policy on whether (and when) to soft offline should be > maintained by userspace, especially in case of an 1G HugeTLB page. > Soft-offline dissolves the HugeTLB page, either in-use or free, into > chunks of 4K pages, reducing HugeTLB pool capacity by 1 hugepage. > If userspace has not acknowledged such behavior, it may be surprised > when later mmap hugepages MAP_FAILED due to lack of hugepages. > In case of a transparent hugepage, it will be split into 4K pages > as well; userspace will stop enjoying the transparent performance. > > In addition, discarding the entire 1G HugeTLB page only because of > corrected memory errors sounds very costly and kernel better not > doing under the hood. But today there are at least 2 such cases: > 1. GHES driver sees both GHES_SEV_CORRECTED and > CPER_SEC_ERROR_THRESHOLD_EXCEEDED after parsing CPER. > 2. RAS Correctable Errors Collector counts correctable errors per > PFN and when the counter for a PFN reaches threshold > In both cases, userspace has no control of the soft offline performed > by kernel's memory failure recovery. > > This commit gives userspace the control of softofflining any page: > kernel only soft offlines raw page / transparent hugepage / HugeTLB > hugepage if userspace has agreed to. The interface to userspace is a > new sysctl called enable_soft_offline under /proc/sys/vm. By default > enable_soft_line is 1 to preserve existing behavior in kernel. s/enable_soft_line/enable_soft_offline > > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com> > --- > mm/memory-failure.c | 16 ++++++++++++++++ > 1 file changed, 16 insertions(+) > > diff --git a/mm/memory-failure.c b/mm/memory-failure.c > index d3c830e817e3..23415fe03318 100644 > --- a/mm/memory-failure.c > +++ b/mm/memory-failure.c > @@ -68,6 +68,8 @@ static int sysctl_memory_failure_early_kill __read_mostly; > > static int sysctl_memory_failure_recovery __read_mostly = 1; > > +static int sysctl_enable_soft_offline __read_mostly = 1; > + > atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); > > static bool hw_memory_failure __read_mostly = false; > @@ -141,6 +143,15 @@ static struct ctl_table memory_failure_table[] = { > .extra1 = SYSCTL_ZERO, > .extra2 = SYSCTL_ONE, > }, > + { > + .procname = "enable_soft_offline", > + .data = &sysctl_enable_soft_offline, > + .maxlen = sizeof(sysctl_enable_soft_offline), > + .mode = 0644, > + .proc_handler = proc_dointvec_minmax, > + .extra1 = SYSCTL_ZERO, > + .extra2 = SYSCTL_ONE, > + } > }; > > /* > @@ -2771,6 +2782,11 @@ int soft_offline_page(unsigned long pfn, int flags) > bool try_again = true; > struct page *page; > > + if (!sysctl_enable_soft_offline) { > + pr_info("soft offline: %#lx: OS-wide disabled\n", pfn); > + return -EINVAL; IMO, "-EPERM" might sound better ;) Using "-EPERM" indicates that the operation is not permitted due to the OS-wide configuration. Thanks, Lance > + } > + > if (!pfn_valid(pfn)) { > WARN_ON_ONCE(flags & MF_COUNT_INCREASED); > return -ENXIO; > -- > 2.45.2.505.gda0bf45e8d-goog > >
On Fri, Jun 14, 2024 at 1:35 AM Lance Yang <ioworker0@gmail.com> wrote: > > Hi Jiaqi, > > On Wed, Jun 12, 2024 at 5:56 AM Jiaqi Yan <jiaqiyan@google.com> wrote: > > > > Correctable memory errors are very common on servers with large > > amount of memory, and are corrected by ECC. Soft offline is kernel's > > additional recovery handling for memory pages having (excessive) > > corrected memory errors. Impacted page is migrated to a healthy page > > if inuse; the original page is discarded for any future use. > > > > The actual policy on whether (and when) to soft offline should be > > maintained by userspace, especially in case of an 1G HugeTLB page. > > Soft-offline dissolves the HugeTLB page, either in-use or free, into > > chunks of 4K pages, reducing HugeTLB pool capacity by 1 hugepage. > > If userspace has not acknowledged such behavior, it may be surprised > > when later mmap hugepages MAP_FAILED due to lack of hugepages. > > In case of a transparent hugepage, it will be split into 4K pages > > as well; userspace will stop enjoying the transparent performance. > > > > In addition, discarding the entire 1G HugeTLB page only because of > > corrected memory errors sounds very costly and kernel better not > > doing under the hood. But today there are at least 2 such cases: > > 1. GHES driver sees both GHES_SEV_CORRECTED and > > CPER_SEC_ERROR_THRESHOLD_EXCEEDED after parsing CPER. > > 2. RAS Correctable Errors Collector counts correctable errors per > > PFN and when the counter for a PFN reaches threshold > > In both cases, userspace has no control of the soft offline performed > > by kernel's memory failure recovery. > > > > This commit gives userspace the control of softofflining any page: > > kernel only soft offlines raw page / transparent hugepage / HugeTLB > > hugepage if userspace has agreed to. The interface to userspace is a > > new sysctl called enable_soft_offline under /proc/sys/vm. By default > > enable_soft_line is 1 to preserve existing behavior in kernel. > > s/enable_soft_line/enable_soft_offline Will fix this typo in v3. > > > > > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com> > > --- > > mm/memory-failure.c | 16 ++++++++++++++++ > > 1 file changed, 16 insertions(+) > > > > diff --git a/mm/memory-failure.c b/mm/memory-failure.c > > index d3c830e817e3..23415fe03318 100644 > > --- a/mm/memory-failure.c > > +++ b/mm/memory-failure.c > > @@ -68,6 +68,8 @@ static int sysctl_memory_failure_early_kill __read_mostly; > > > > static int sysctl_memory_failure_recovery __read_mostly = 1; > > > > +static int sysctl_enable_soft_offline __read_mostly = 1; > > + > > atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); > > > > static bool hw_memory_failure __read_mostly = false; > > @@ -141,6 +143,15 @@ static struct ctl_table memory_failure_table[] = { > > .extra1 = SYSCTL_ZERO, > > .extra2 = SYSCTL_ONE, > > }, > > + { > > + .procname = "enable_soft_offline", > > + .data = &sysctl_enable_soft_offline, > > + .maxlen = sizeof(sysctl_enable_soft_offline), > > + .mode = 0644, > > + .proc_handler = proc_dointvec_minmax, > > + .extra1 = SYSCTL_ZERO, > > + .extra2 = SYSCTL_ONE, > > + } > > }; > > > > /* > > @@ -2771,6 +2782,11 @@ int soft_offline_page(unsigned long pfn, int flags) > > bool try_again = true; > > struct page *page; > > > > + if (!sysctl_enable_soft_offline) { > > + pr_info("soft offline: %#lx: OS-wide disabled\n", pfn); > > + return -EINVAL; > > IMO, "-EPERM" might sound better ;) > > Using "-EPERM" indicates that the operation is not permitted due to > the OS-wide configuration. Miaohe suggested -EOPNOTSUPP. I agree both EOPNOTSUPP and EPERM may be better than EINVAL. But I wonder how about EAGAIN? With EAGAIN plus showing "disabled by /proc/sys/vm/enable_soft_offline" in dmesg, users now should be clear that they can try again with /proc/sys/vm/enable_soft_offline=1. > > Thanks, > Lance > > > + } > > + > > if (!pfn_valid(pfn)) { > > WARN_ON_ONCE(flags & MF_COUNT_INCREASED); > > return -ENXIO; > > -- > > 2.45.2.505.gda0bf45e8d-goog > > > >
On Thu, Jun 13, 2024 at 8:28 PM Miaohe Lin <linmiaohe@huawei.com> wrote: > > On 2024/6/12 5:55, Jiaqi Yan wrote: > > Correctable memory errors are very common on servers with large > > amount of memory, and are corrected by ECC. Soft offline is kernel's > > additional recovery handling for memory pages having (excessive) > > corrected memory errors. Impacted page is migrated to a healthy page > > if inuse; the original page is discarded for any future use. > > > > Thanks for your update. > > > The actual policy on whether (and when) to soft offline should be > > maintained by userspace, especially in case of an 1G HugeTLB page. > > Soft-offline dissolves the HugeTLB page, either in-use or free, into > > chunks of 4K pages, reducing HugeTLB pool capacity by 1 hugepage. > > If userspace has not acknowledged such behavior, it may be surprised > > when later mmap hugepages MAP_FAILED due to lack of hugepages. > > s/mmap hugepages MAP_FAILED/fails to mmap hugepages/ ? > > > In case of a transparent hugepage, it will be split into 4K pages > > as well; userspace will stop enjoying the transparent performance. > > > > In addition, discarding the entire 1G HugeTLB page only because of > > corrected memory errors sounds very costly and kernel better not > > doing under the hood. But today there are at least 2 such cases: > > s/doing/doing so/ ? > > > 1. GHES driver sees both GHES_SEV_CORRECTED and > > CPER_SEC_ERROR_THRESHOLD_EXCEEDED after parsing CPER. > > 2. RAS Correctable Errors Collector counts correctable errors per > > PFN and when the counter for a PFN reaches threshold > > In both cases, userspace has no control of the soft offline performed > > by kernel's memory failure recovery. > > > > This commit gives userspace the control of softofflining any page: > > kernel only soft offlines raw page / transparent hugepage / HugeTLB > > hugepage if userspace has agreed to. The interface to userspace is a > > new sysctl called enable_soft_offline under /proc/sys/vm. By default > > enable_soft_line is 1 to preserve existing behavior in kernel. > > s/enable_soft_line/enable_soft_offline/ Will fix these 3 typos in v3. > > > > > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com> > > --- > > mm/memory-failure.c | 16 ++++++++++++++++ > > 1 file changed, 16 insertions(+) > > > > diff --git a/mm/memory-failure.c b/mm/memory-failure.c > > index d3c830e817e3..23415fe03318 100644 > > --- a/mm/memory-failure.c > > +++ b/mm/memory-failure.c > > @@ -68,6 +68,8 @@ static int sysctl_memory_failure_early_kill __read_mostly; > > > > static int sysctl_memory_failure_recovery __read_mostly = 1; > > > > +static int sysctl_enable_soft_offline __read_mostly = 1; > > + > > atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); > > > > static bool hw_memory_failure __read_mostly = false; > > @@ -141,6 +143,15 @@ static struct ctl_table memory_failure_table[] = { > > .extra1 = SYSCTL_ZERO, > > .extra2 = SYSCTL_ONE, > > }, > > + { > > + .procname = "enable_soft_offline", > > + .data = &sysctl_enable_soft_offline, > > + .maxlen = sizeof(sysctl_enable_soft_offline), > > + .mode = 0644, > > + .proc_handler = proc_dointvec_minmax, > > + .extra1 = SYSCTL_ZERO, > > + .extra2 = SYSCTL_ONE, > > + } > > }; > > > > /* > > @@ -2771,6 +2782,11 @@ int soft_offline_page(unsigned long pfn, int flags) > > bool try_again = true; > > struct page *page; > > > > + if (!sysctl_enable_soft_offline) { > > + pr_info("soft offline: %#lx: OS-wide disabled\n", pfn); > > + return -EINVAL; > > + } > > + > > IMHO, callers might reach here with page refcnt increased. So we have to take care of releasing it first? Ah, I think you are right when MF_COUNT_INCREASED. I will move this after the pfn_to_online_page check, and release if disabled. > Also will it be better to return -EOPNOTSUPP or some other better errno? > > Thanks. > .
On 2024/6/15 0:30, Jiaqi Yan wrote: > On Fri, Jun 14, 2024 at 1:35 AM Lance Yang <ioworker0@gmail.com> wrote: >> >> Hi Jiaqi, >> >> On Wed, Jun 12, 2024 at 5:56 AM Jiaqi Yan <jiaqiyan@google.com> wrote: >>> >>> Correctable memory errors are very common on servers with large >>> amount of memory, and are corrected by ECC. Soft offline is kernel's >>> additional recovery handling for memory pages having (excessive) >>> corrected memory errors. Impacted page is migrated to a healthy page >>> if inuse; the original page is discarded for any future use. >>> >>> The actual policy on whether (and when) to soft offline should be >>> maintained by userspace, especially in case of an 1G HugeTLB page. >>> Soft-offline dissolves the HugeTLB page, either in-use or free, into >>> chunks of 4K pages, reducing HugeTLB pool capacity by 1 hugepage. >>> If userspace has not acknowledged such behavior, it may be surprised >>> when later mmap hugepages MAP_FAILED due to lack of hugepages. >>> In case of a transparent hugepage, it will be split into 4K pages >>> as well; userspace will stop enjoying the transparent performance. >>> >>> In addition, discarding the entire 1G HugeTLB page only because of >>> corrected memory errors sounds very costly and kernel better not >>> doing under the hood. But today there are at least 2 such cases: >>> 1. GHES driver sees both GHES_SEV_CORRECTED and >>> CPER_SEC_ERROR_THRESHOLD_EXCEEDED after parsing CPER. >>> 2. RAS Correctable Errors Collector counts correctable errors per >>> PFN and when the counter for a PFN reaches threshold >>> In both cases, userspace has no control of the soft offline performed >>> by kernel's memory failure recovery. >>> >>> This commit gives userspace the control of softofflining any page: >>> kernel only soft offlines raw page / transparent hugepage / HugeTLB >>> hugepage if userspace has agreed to. The interface to userspace is a >>> new sysctl called enable_soft_offline under /proc/sys/vm. By default >>> enable_soft_line is 1 to preserve existing behavior in kernel. >> >> s/enable_soft_line/enable_soft_offline > > Will fix this typo in v3. > >> >>> >>> Signed-off-by: Jiaqi Yan <jiaqiyan@google.com> >>> --- >>> mm/memory-failure.c | 16 ++++++++++++++++ >>> 1 file changed, 16 insertions(+) >>> >>> diff --git a/mm/memory-failure.c b/mm/memory-failure.c >>> index d3c830e817e3..23415fe03318 100644 >>> --- a/mm/memory-failure.c >>> +++ b/mm/memory-failure.c >>> @@ -68,6 +68,8 @@ static int sysctl_memory_failure_early_kill __read_mostly; >>> >>> static int sysctl_memory_failure_recovery __read_mostly = 1; >>> >>> +static int sysctl_enable_soft_offline __read_mostly = 1; >>> + >>> atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); >>> >>> static bool hw_memory_failure __read_mostly = false; >>> @@ -141,6 +143,15 @@ static struct ctl_table memory_failure_table[] = { >>> .extra1 = SYSCTL_ZERO, >>> .extra2 = SYSCTL_ONE, >>> }, >>> + { >>> + .procname = "enable_soft_offline", >>> + .data = &sysctl_enable_soft_offline, >>> + .maxlen = sizeof(sysctl_enable_soft_offline), >>> + .mode = 0644, >>> + .proc_handler = proc_dointvec_minmax, >>> + .extra1 = SYSCTL_ZERO, >>> + .extra2 = SYSCTL_ONE, >>> + } >>> }; >>> >>> /* >>> @@ -2771,6 +2782,11 @@ int soft_offline_page(unsigned long pfn, int flags) >>> bool try_again = true; >>> struct page *page; >>> >>> + if (!sysctl_enable_soft_offline) { >>> + pr_info("soft offline: %#lx: OS-wide disabled\n", pfn); >>> + return -EINVAL; >> >> IMO, "-EPERM" might sound better ;) >> >> Using "-EPERM" indicates that the operation is not permitted due to >> the OS-wide configuration. > > Miaohe suggested -EOPNOTSUPP. I agree both EOPNOTSUPP and EPERM may be > better than EINVAL. But I wonder how about EAGAIN? With EAGAIN plus > showing "disabled by /proc/sys/vm/enable_soft_offline" in dmesg, users > now should be clear that they can try again with > /proc/sys/vm/enable_soft_offline=1. IMHO, it might not be suitable to use EAGAIN. Because it means "Resource temporarily unavailable" and this can be solved by simply retry later without any further actions taken. But I might be wrong. Thanks. .
On 2024-06-17 09:31, Miaohe Lin wrote: > IMHO, it might not be suitable to use EAGAIN. Because it means > "Resource temporarily unavailable" and > this can be solved by simply retry later without any further actions > taken. But I might be wrong. We usually use 'EOPNOTSUPP' when we fail due to a setting not being set. EPERM is more for a capability matter.
On 2024/6/17 15:51, Oscar Salvador wrote: > On 2024-06-17 09:31, Miaohe Lin wrote: > >> IMHO, it might not be suitable to use EAGAIN. Because it means >> "Resource temporarily unavailable" and >> this can be solved by simply retry later without any further actions >> taken. But I might be wrong. > > We usually use 'EOPNOTSUPP' when we fail due to a setting not being set. > EPERM is more for a capability matter. Thanks Oscar. So we should return EOPNOTSUPP here. Thanks. .
On Mon, Jun 17, 2024 at 4:16 PM Miaohe Lin <linmiaohe@huawei.com> wrote: > > On 2024/6/17 15:51, Oscar Salvador wrote: > > On 2024-06-17 09:31, Miaohe Lin wrote: > > > >> IMHO, it might not be suitable to use EAGAIN. Because it means > >> "Resource temporarily unavailable" and > >> this can be solved by simply retry later without any further actions > >> taken. But I might be wrong. > > > > We usually use 'EOPNOTSUPP' when we fail due to a setting not being set. > > EPERM is more for a capability matter. Agreed. Let's use 'EOPNOTSUPP' here ;) Thanks, Lance > > Thanks Oscar. So we should return EOPNOTSUPP here. > > Thanks. > .
On Mon, Jun 17, 2024 at 3:34 AM Lance Yang <ioworker0@gmail.com> wrote: > > On Mon, Jun 17, 2024 at 4:16 PM Miaohe Lin <linmiaohe@huawei.com> wrote: > > > > On 2024/6/17 15:51, Oscar Salvador wrote: > > > On 2024-06-17 09:31, Miaohe Lin wrote: > > > > > >> IMHO, it might not be suitable to use EAGAIN. Because it means > > >> "Resource temporarily unavailable" and > > >> this can be solved by simply retry later without any further actions > > >> taken. But I might be wrong. > > > > > > We usually use 'EOPNOTSUPP' when we fail due to a setting not being set. > > > EPERM is more for a capability matter. > > Agreed. Let's use 'EOPNOTSUPP' here ;) Thanks Oscar and Lance. Agreed, will do in v3. > > Thanks, > Lance > > > > > Thanks Oscar. So we should return EOPNOTSUPP here. > > > > Thanks. > > .
diff --git a/mm/memory-failure.c b/mm/memory-failure.c index d3c830e817e3..23415fe03318 100644 --- a/mm/memory-failure.c +++ b/mm/memory-failure.c @@ -68,6 +68,8 @@ static int sysctl_memory_failure_early_kill __read_mostly; static int sysctl_memory_failure_recovery __read_mostly = 1; +static int sysctl_enable_soft_offline __read_mostly = 1; + atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); static bool hw_memory_failure __read_mostly = false; @@ -141,6 +143,15 @@ static struct ctl_table memory_failure_table[] = { .extra1 = SYSCTL_ZERO, .extra2 = SYSCTL_ONE, }, + { + .procname = "enable_soft_offline", + .data = &sysctl_enable_soft_offline, + .maxlen = sizeof(sysctl_enable_soft_offline), + .mode = 0644, + .proc_handler = proc_dointvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = SYSCTL_ONE, + } }; /* @@ -2771,6 +2782,11 @@ int soft_offline_page(unsigned long pfn, int flags) bool try_again = true; struct page *page; + if (!sysctl_enable_soft_offline) { + pr_info("soft offline: %#lx: OS-wide disabled\n", pfn); + return -EINVAL; + } + if (!pfn_valid(pfn)) { WARN_ON_ONCE(flags & MF_COUNT_INCREASED); return -ENXIO;
Correctable memory errors are very common on servers with large amount of memory, and are corrected by ECC. Soft offline is kernel's additional recovery handling for memory pages having (excessive) corrected memory errors. Impacted page is migrated to a healthy page if inuse; the original page is discarded for any future use. The actual policy on whether (and when) to soft offline should be maintained by userspace, especially in case of an 1G HugeTLB page. Soft-offline dissolves the HugeTLB page, either in-use or free, into chunks of 4K pages, reducing HugeTLB pool capacity by 1 hugepage. If userspace has not acknowledged such behavior, it may be surprised when later mmap hugepages MAP_FAILED due to lack of hugepages. In case of a transparent hugepage, it will be split into 4K pages as well; userspace will stop enjoying the transparent performance. In addition, discarding the entire 1G HugeTLB page only because of corrected memory errors sounds very costly and kernel better not doing under the hood. But today there are at least 2 such cases: 1. GHES driver sees both GHES_SEV_CORRECTED and CPER_SEC_ERROR_THRESHOLD_EXCEEDED after parsing CPER. 2. RAS Correctable Errors Collector counts correctable errors per PFN and when the counter for a PFN reaches threshold In both cases, userspace has no control of the soft offline performed by kernel's memory failure recovery. This commit gives userspace the control of softofflining any page: kernel only soft offlines raw page / transparent hugepage / HugeTLB hugepage if userspace has agreed to. The interface to userspace is a new sysctl called enable_soft_offline under /proc/sys/vm. By default enable_soft_line is 1 to preserve existing behavior in kernel. Signed-off-by: Jiaqi Yan <jiaqiyan@google.com> --- mm/memory-failure.c | 16 ++++++++++++++++ 1 file changed, 16 insertions(+)