This series is a follow-up to [1], which adds mTHP support to khugepaged. mTHP khugepaged support is a "loose" dependency for the sysfs/sysctl configs to make sense. Without it global="defer" and mTHP="inherit" case is "undefined" behavior.
We've seen cases were customers switching from RHEL7 to RHEL8 see a significant increase in the memory footprint for the same workloads.
Through our investigations we found that a large contributing factor to the increase in RSS was an increase in THP usage.
For workloads like MySQL, or when using allocators like jemalloc, it is often recommended to set /transparent_hugepages/enabled=never. This is in part due to performance degradations and increased memory waste.
This series introduces enabled=defer, this setting acts as a middle ground between always and madvise. If the mapping is MADV_HUGEPAGE, the page fault handler will act normally, making a hugepage if possible. If the allocation is not MADV_HUGEPAGE, then the page fault handler will default to the base size allocation. The caveat is that khugepaged can still operate on pages that are not MADV_HUGEPAGE.
This allows for three things... one, applications specifically designed to use hugepages will get them, and two, applications that don't use hugepages can still benefit from them without aggressively inserting THPs at every possible chance. This curbs the memory waste, and defers the use of hugepages to khugepaged. Khugepaged can then scan the memory for eligible collapsing. Lastly there is the added benefit for those who want THPs but experience higher latency PFs. Now you can get base page performance at the PF handler and Hugepage performance for those mappings after they collapse.
Admins may want to lower max_ptes_none, if not, khugepaged may aggressively collapse single allocations into hugepages.
TESTING: - Built for x86_64, aarch64, ppc64le, and s390x - selftests mm - In [1] I provided a script [2] that has multiple access patterns - lots of general use. - redis testing. This test was my original case for the defer mode. What I was able to prove was that THP=always leads to increased max_latency cases; hence why it is recommended to disable THPs for redis servers. However with 'defer' we dont have the max_latency spikes and can still get the system to utilize THPs. I further tested this with the mTHP defer setting and found that redis (and probably other jmalloc users) can utilize THPs via defer (+mTHP defer) without a large latency penalty and some potential gains. I uploaded some mmtest results here[3] which compares: stock+thp=never stock+(m)thp=always khugepaged-mthp + defer (max_ptes_none=64)
The results show that (m)THPs can cause some throughput regression in some cases, but also has gains in other cases. The mTHP+defer results have more gains and less losses over the (m)THP=always case.
V6 Changes: - nits - rebased dependent series and added review tags
V5 Changes: - rebased dependent series - added reviewed-by tag on 2/4
V4 Changes: - Minor Documentation fixes - rebased the dependent series [1] onto mm-unstable commit 0e68b850b1d3 ("vmalloc: use atomic_long_add_return_relaxed()")
V3 Changes: - Combined the documentation commits into one, and moved a section to the khugepaged mthp patchset
V2 Changes: - base changes on mTHP khugepaged support - Fix selftests parsing issue - add mTHP defer option - add mTHP defer Documentation
[1] - https://lore.kernel.org/all/20250515032226.128900-1-npache@redhat.com/ [2] - https://gitlab.com/npache/khugepaged_mthp_test [3] - https://people.redhat.com/npache/mthp_khugepaged_defer/testoutput2/output.ht...
Nico Pache (4): mm: defer THP insertion to khugepaged mm: document (m)THP defer usage khugepaged: add defer option to mTHP options selftests: mm: add defer to thp setting parser
Documentation/admin-guide/mm/transhuge.rst | 31 +++++++--- include/linux/huge_mm.h | 18 +++++- mm/huge_memory.c | 69 +++++++++++++++++++--- mm/khugepaged.c | 8 +-- tools/testing/selftests/mm/thp_settings.c | 1 + tools/testing/selftests/mm/thp_settings.h | 1 + 6 files changed, 106 insertions(+), 22 deletions(-)
setting /transparent_hugepages/enabled=always allows applications to benefit from THPs without having to madvise. However, the page fault handler takes very few considerations to decide weather or not to actually use a THP. This can lead to a lot of wasted memory. khugepaged only operates on memory that was either allocated with enabled=always or MADV_HUGEPAGE.
Introduce the ability to set enabled=defer, which will prevent THPs from being allocated by the page fault handler unless madvise is set, leaving it up to khugepaged to decide which allocations will collapse to a THP. This should allow applications to benefits from THPs, while curbing some of the memory waste.
Acked-by: Zi Yan ziy@nvidia.com Co-developed-by: Rafael Aquini raquini@redhat.com Signed-off-by: Rafael Aquini raquini@redhat.com Signed-off-by: Nico Pache npache@redhat.com --- include/linux/huge_mm.h | 15 +++++++++++++-- mm/huge_memory.c | 31 +++++++++++++++++++++++++++---- 2 files changed, 40 insertions(+), 6 deletions(-)
diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h index e3d15c737008..02038e3db829 100644 --- a/include/linux/huge_mm.h +++ b/include/linux/huge_mm.h @@ -48,6 +48,7 @@ enum transparent_hugepage_flag { TRANSPARENT_HUGEPAGE_UNSUPPORTED, TRANSPARENT_HUGEPAGE_FLAG, TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, + TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG, TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, @@ -186,6 +187,7 @@ static inline bool hugepage_global_enabled(void) { return transparent_hugepage_flags & ((1<<TRANSPARENT_HUGEPAGE_FLAG) | + (1<<TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG) | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)); }
@@ -195,6 +197,12 @@ static inline bool hugepage_global_always(void) (1<<TRANSPARENT_HUGEPAGE_FLAG); }
+static inline bool hugepage_global_defer(void) +{ + return transparent_hugepage_flags & + (1<<TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG); +} + static inline int highest_order(unsigned long orders) { return fls_long(orders) - 1; @@ -291,13 +299,16 @@ unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma, unsigned long tva_flags, unsigned long orders) { + if ((tva_flags & TVA_IN_PF) && hugepage_global_defer() && + !(vm_flags & VM_HUGEPAGE)) + return 0; + /* Optimization to check if required orders are enabled early. */ if ((tva_flags & TVA_ENFORCE_SYSFS) && vma_is_anonymous(vma)) { unsigned long mask = READ_ONCE(huge_anon_orders_always); - if (vm_flags & VM_HUGEPAGE) mask |= READ_ONCE(huge_anon_orders_madvise); - if (hugepage_global_always() || + if (hugepage_global_always() || hugepage_global_defer() || ((vm_flags & VM_HUGEPAGE) && hugepage_global_enabled())) mask |= READ_ONCE(huge_anon_orders_inherit);
diff --git a/mm/huge_memory.c b/mm/huge_memory.c index 700988a0d5cf..ce0ee74753af 100644 --- a/mm/huge_memory.c +++ b/mm/huge_memory.c @@ -297,12 +297,15 @@ static ssize_t enabled_show(struct kobject *kobj, const char *output;
if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) - output = "[always] madvise never"; + output = "[always] madvise defer never"; else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) - output = "always [madvise] never"; + output = "always [madvise] defer never"; + else if (test_bit(TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG, + &transparent_hugepage_flags)) + output = "always madvise [defer] never"; else - output = "always madvise [never]"; + output = "always madvise defer [never]";
return sysfs_emit(buf, "%s\n", output); } @@ -315,13 +318,20 @@ static ssize_t enabled_store(struct kobject *kobj,
if (sysfs_streq(buf, "always")) { clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); + clear_bit(TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG, &transparent_hugepage_flags); set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); + } else if (sysfs_streq(buf, "defer")) { + clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); + clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); + set_bit(TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG, &transparent_hugepage_flags); } else if (sysfs_streq(buf, "madvise")) { clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); + clear_bit(TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG, &transparent_hugepage_flags); set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); } else if (sysfs_streq(buf, "never")) { clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); + clear_bit(TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG, &transparent_hugepage_flags); } else ret = -EINVAL;
@@ -954,18 +964,31 @@ static int __init setup_transparent_hugepage(char *str) &transparent_hugepage_flags); clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); + clear_bit(TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG, + &transparent_hugepage_flags); ret = 1; + } else if (!strcmp(str, "defer")) { + clear_bit(TRANSPARENT_HUGEPAGE_FLAG, + &transparent_hugepage_flags); + clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, + &transparent_hugepage_flags); + set_bit(TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG, + &transparent_hugepage_flags); } else if (!strcmp(str, "madvise")) { clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); + clear_bit(TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG, + &transparent_hugepage_flags); set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, - &transparent_hugepage_flags); + &transparent_hugepage_flags); ret = 1; } else if (!strcmp(str, "never")) { clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); + clear_bit(TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG, + &transparent_hugepage_flags); ret = 1; } out:
On Thu, May 15, 2025 at 12:39 PM Nico Pache npache@redhat.com wrote:
setting /transparent_hugepages/enabled=always allows applications to benefit from THPs without having to madvise. However, the page fault handler takes very few considerations to decide weather or not to actually use a THP. This can lead to a lot of wasted memory. khugepaged only operates on memory that was either allocated with enabled=always or MADV_HUGEPAGE.
Introduce the ability to set enabled=defer, which will prevent THPs from being allocated by the page fault handler unless madvise is set, leaving it up to khugepaged to decide which allocations will collapse to a THP. This should allow applications to benefits from THPs, while curbing some of the memory waste.
Acked-by: Zi Yan ziy@nvidia.com Co-developed-by: Rafael Aquini raquini@redhat.com Signed-off-by: Rafael Aquini raquini@redhat.com Signed-off-by: Nico Pache npache@redhat.com
include/linux/huge_mm.h | 15 +++++++++++++-- mm/huge_memory.c | 31 +++++++++++++++++++++++++++---- 2 files changed, 40 insertions(+), 6 deletions(-)
diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h index e3d15c737008..02038e3db829 100644 --- a/include/linux/huge_mm.h +++ b/include/linux/huge_mm.h @@ -48,6 +48,7 @@ enum transparent_hugepage_flag { TRANSPARENT_HUGEPAGE_UNSUPPORTED, TRANSPARENT_HUGEPAGE_FLAG, TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG, TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,@@ -186,6 +187,7 @@ static inline bool hugepage_global_enabled(void) { return transparent_hugepage_flags & ((1<<TRANSPARENT_HUGEPAGE_FLAG) |
(1<<TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG) | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG));}
@@ -195,6 +197,12 @@ static inline bool hugepage_global_always(void) (1<<TRANSPARENT_HUGEPAGE_FLAG); }
+static inline bool hugepage_global_defer(void) +{
return transparent_hugepage_flags &(1<<TRANSPARENT_HUGEPAGE_DEFER_PF_FLAG);+}
static inline int highest_order(unsigned long orders) { return fls_long(orders) - 1; @@ -291,13 +299,16 @@ unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma, unsigned long tva_flags, unsigned long orders) {
if ((tva_flags & TVA_IN_PF) && hugepage_global_defer() &&!(vm_flags & VM_HUGEPAGE))return 0;/* Optimization to check if required orders are enabled early. */ if ((tva_flags & TVA_ENFORCE_SYSFS) && vma_is_anonymous(vma)) { unsigned long mask = READ_ONCE(huge_anon_orders_always);
if (vm_flags & VM_HUGEPAGE) mask |= READ_ONCE(huge_anon_orders_madvise);if (hugepage_global_always() ||
if (hugepage_global_always() || hugepage_global_defer() || ((vm_flags & VM_HUGEPAGE) && hugepage_global_enabled())) mask |= READ_ONCE(huge_anon_orders_inherit);diff --git a/mm/huge_memory.c b/mm/huge_memory.c index 700988a0d5cf..ce0ee74753af 100644 --- a/mm/huge_memory.c +++ b/mm/huge_memory.c @@ -297,12 +297,15 @@ static ssize_t enabled_show(struct kobject *kobj, const char *output;
if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
output = "[always] madvise never";
output = "[always] madvise defer never";
a small nit: alphabetical ordering might improve readability here.
The new defer option for (m)THPs allows for a more conservative approach to (m)THPs. Document its usage in the transhuge admin-guide.
Reviewed-by: Bagas Sanjaya bagasdotme@gmail.com Signed-off-by: Nico Pache npache@redhat.com --- Documentation/admin-guide/mm/transhuge.rst | 31 ++++++++++++++++------ 1 file changed, 23 insertions(+), 8 deletions(-)
diff --git a/Documentation/admin-guide/mm/transhuge.rst b/Documentation/admin-guide/mm/transhuge.rst index 5c63fe51b3ad..7e87ef317add 100644 --- a/Documentation/admin-guide/mm/transhuge.rst +++ b/Documentation/admin-guide/mm/transhuge.rst @@ -88,8 +88,9 @@ In certain cases when hugepages are enabled system wide, application may end up allocating more memory resources. An application may mmap a large region but only touch 1 byte of it, in that case a 2M page might be allocated instead of a 4k page for no good. This is why it's -possible to disable hugepages system-wide and to only have them inside -MADV_HUGEPAGE madvise regions. +possible to disable hugepages system-wide, only have them inside +MADV_HUGEPAGE madvise regions, or defer them away from the page fault +handler to khugepaged.
Embedded systems should enable hugepages only inside madvise regions to eliminate any risk of wasting any precious byte of memory and to @@ -99,6 +100,15 @@ Applications that gets a lot of benefit from hugepages and that don't risk to lose memory by using hugepages, should use madvise(MADV_HUGEPAGE) on their critical mmapped regions.
+Applications that would like to benefit from THPs but would still like a +more memory conservative approach can choose 'defer'. This avoids +inserting THPs at the page fault handler unless they are MADV_HUGEPAGE. +Khugepaged will then scan all mappings, even those not explicitly marked +with MADV_HUGEPAGE, for potential collapses into (m)THPs. Admins using +this the 'defer' setting should consider tweaking max_ptes_none. The +current default of 511 may aggressively collapse your PTEs into PMDs. +Lower this value to conserve more memory (i.e., max_ptes_none=64). + .. _thp_sysfs:
sysfs @@ -109,11 +119,14 @@ Global THP controls
Transparent Hugepage Support for anonymous memory can be entirely disabled (mostly for debugging purposes) or only enabled inside MADV_HUGEPAGE -regions (to avoid the risk of consuming more memory resources) or enabled -system wide. This can be achieved per-supported-THP-size with one of:: +regions (to avoid the risk of consuming more memory resources), deferred to +khugepaged, or enabled system wide. + +This can be achieved per-supported-THP-size with one of::
echo always >/sys/kernel/mm/transparent_hugepage/hugepages-<size>kB/enabled echo madvise >/sys/kernel/mm/transparent_hugepage/hugepages-<size>kB/enabled + echo defer >/sys/kernel/mm/transparent_hugepage/hugepages-<size>kB/enabled echo never >/sys/kernel/mm/transparent_hugepage/hugepages-<size>kB/enabled
where <size> is the hugepage size being addressed, the available sizes @@ -136,6 +149,7 @@ The top-level setting (for use with "inherit") can be set by issuing one of the following commands::
echo always >/sys/kernel/mm/transparent_hugepage/enabled + echo defer >/sys/kernel/mm/transparent_hugepage/enabled echo madvise >/sys/kernel/mm/transparent_hugepage/enabled echo never >/sys/kernel/mm/transparent_hugepage/enabled
@@ -286,7 +300,8 @@ of small pages into one large page:: A higher value leads to use additional memory for programs. A lower value leads to gain less thp performance. Value of max_ptes_none can waste cpu time very little, you can -ignore it. +ignore it. Consider lowering this value when using +``transparent_hugepage=defer``
``max_ptes_swap`` specifies how many pages can be brought in from swap when collapsing a group of pages into a transparent huge page:: @@ -311,14 +326,14 @@ Boot parameters
You can change the sysfs boot time default for the top-level "enabled" control by passing the parameter ``transparent_hugepage=always`` or -``transparent_hugepage=madvise`` or ``transparent_hugepage=never`` to the -kernel command line. +``transparent_hugepage=madvise`` or ``transparent_hugepage=defer`` or +``transparent_hugepage=never`` to the kernel command line.
Alternatively, each supported anonymous THP size can be controlled by passing ``thp_anon=<size>[KMG],<size>[KMG]:<state>;<size>[KMG]-<size>[KMG]:<state>``, where ``<size>`` is the THP size (must be a power of 2 of PAGE_SIZE and supported anonymous THP) and ``<state>`` is one of ``always``, ``madvise``, -``never`` or ``inherit``. +``defer``, ``never`` or ``inherit``.
For example, the following will set 16K, 32K, 64K THP to ``always``, set 128K, 512K to ``inherit``, set 256K to ``madvise`` and 1M, 2M
Now that we have defer to globally disable THPs at fault time, lets add a defer setting to the mTHP options. This will allow khugepaged to operate at that order, while avoiding it at PF time.
Signed-off-by: Nico Pache npache@redhat.com --- include/linux/huge_mm.h | 5 +++++ mm/huge_memory.c | 38 +++++++++++++++++++++++++++++++++----- mm/khugepaged.c | 8 ++++---- 3 files changed, 42 insertions(+), 9 deletions(-)
diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h index 02038e3db829..71a1edb5062e 100644 --- a/include/linux/huge_mm.h +++ b/include/linux/huge_mm.h @@ -96,6 +96,7 @@ extern struct kobj_attribute thpsize_shmem_enabled_attr; #define TVA_SMAPS (1 << 0) /* Will be used for procfs */ #define TVA_IN_PF (1 << 1) /* Page fault handler */ #define TVA_ENFORCE_SYSFS (1 << 2) /* Obey sysfs configuration */ +#define TVA_IN_KHUGEPAGE ((1 << 2) | (1 << 3)) /* Khugepaged defer support */
#define thp_vma_allowable_order(vma, vm_flags, tva_flags, order) \ (!!thp_vma_allowable_orders(vma, vm_flags, tva_flags, BIT(order))) @@ -182,6 +183,7 @@ extern unsigned long transparent_hugepage_flags; extern unsigned long huge_anon_orders_always; extern unsigned long huge_anon_orders_madvise; extern unsigned long huge_anon_orders_inherit; +extern unsigned long huge_anon_orders_defer;
static inline bool hugepage_global_enabled(void) { @@ -306,6 +308,9 @@ unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma, /* Optimization to check if required orders are enabled early. */ if ((tva_flags & TVA_ENFORCE_SYSFS) && vma_is_anonymous(vma)) { unsigned long mask = READ_ONCE(huge_anon_orders_always); + + if ((tva_flags & TVA_IN_KHUGEPAGE) == TVA_IN_KHUGEPAGE) + mask |= READ_ONCE(huge_anon_orders_defer); if (vm_flags & VM_HUGEPAGE) mask |= READ_ONCE(huge_anon_orders_madvise); if (hugepage_global_always() || hugepage_global_defer() || diff --git a/mm/huge_memory.c b/mm/huge_memory.c index ce0ee74753af..addf4c16c91d 100644 --- a/mm/huge_memory.c +++ b/mm/huge_memory.c @@ -81,6 +81,7 @@ unsigned long huge_zero_pfn __read_mostly = ~0UL; unsigned long huge_anon_orders_always __read_mostly; unsigned long huge_anon_orders_madvise __read_mostly; unsigned long huge_anon_orders_inherit __read_mostly; +unsigned long huge_anon_orders_defer __read_mostly; static bool anon_orders_configured __initdata;
static inline bool file_thp_enabled(struct vm_area_struct *vma) @@ -505,13 +506,15 @@ static ssize_t anon_enabled_show(struct kobject *kobj, const char *output;
if (test_bit(order, &huge_anon_orders_always)) - output = "[always] inherit madvise never"; + output = "[always] inherit madvise defer never"; else if (test_bit(order, &huge_anon_orders_inherit)) - output = "always [inherit] madvise never"; + output = "always [inherit] madvise defer never"; else if (test_bit(order, &huge_anon_orders_madvise)) - output = "always inherit [madvise] never"; + output = "always inherit [madvise] defer never"; + else if (test_bit(order, &huge_anon_orders_defer)) + output = "always inherit madvise [defer] never"; else - output = "always inherit madvise [never]"; + output = "always inherit madvise defer [never]";
return sysfs_emit(buf, "%s\n", output); } @@ -527,25 +530,36 @@ static ssize_t anon_enabled_store(struct kobject *kobj, spin_lock(&huge_anon_orders_lock); clear_bit(order, &huge_anon_orders_inherit); clear_bit(order, &huge_anon_orders_madvise); + clear_bit(order, &huge_anon_orders_defer); set_bit(order, &huge_anon_orders_always); spin_unlock(&huge_anon_orders_lock); } else if (sysfs_streq(buf, "inherit")) { spin_lock(&huge_anon_orders_lock); clear_bit(order, &huge_anon_orders_always); clear_bit(order, &huge_anon_orders_madvise); + clear_bit(order, &huge_anon_orders_defer); set_bit(order, &huge_anon_orders_inherit); spin_unlock(&huge_anon_orders_lock); } else if (sysfs_streq(buf, "madvise")) { spin_lock(&huge_anon_orders_lock); clear_bit(order, &huge_anon_orders_always); clear_bit(order, &huge_anon_orders_inherit); + clear_bit(order, &huge_anon_orders_defer); set_bit(order, &huge_anon_orders_madvise); spin_unlock(&huge_anon_orders_lock); + } else if (sysfs_streq(buf, "defer")) { + spin_lock(&huge_anon_orders_lock); + clear_bit(order, &huge_anon_orders_always); + clear_bit(order, &huge_anon_orders_inherit); + clear_bit(order, &huge_anon_orders_madvise); + set_bit(order, &huge_anon_orders_defer); + spin_unlock(&huge_anon_orders_lock); } else if (sysfs_streq(buf, "never")) { spin_lock(&huge_anon_orders_lock); clear_bit(order, &huge_anon_orders_always); clear_bit(order, &huge_anon_orders_inherit); clear_bit(order, &huge_anon_orders_madvise); + clear_bit(order, &huge_anon_orders_defer); spin_unlock(&huge_anon_orders_lock); } else ret = -EINVAL; @@ -1002,7 +1016,7 @@ static char str_dup[PAGE_SIZE] __initdata; static int __init setup_thp_anon(char *str) { char *token, *range, *policy, *subtoken; - unsigned long always, inherit, madvise; + unsigned long always, inherit, madvise, defer; char *start_size, *end_size; int start, end, nr; char *p; @@ -1014,6 +1028,8 @@ static int __init setup_thp_anon(char *str) always = huge_anon_orders_always; madvise = huge_anon_orders_madvise; inherit = huge_anon_orders_inherit; + defer = huge_anon_orders_defer; + p = str_dup; while ((token = strsep(&p, ";")) != NULL) { range = strsep(&token, ":"); @@ -1053,18 +1069,28 @@ static int __init setup_thp_anon(char *str) bitmap_set(&always, start, nr); bitmap_clear(&inherit, start, nr); bitmap_clear(&madvise, start, nr); + bitmap_clear(&defer, start, nr); } else if (!strcmp(policy, "madvise")) { bitmap_set(&madvise, start, nr); bitmap_clear(&inherit, start, nr); bitmap_clear(&always, start, nr); + bitmap_clear(&defer, start, nr); } else if (!strcmp(policy, "inherit")) { bitmap_set(&inherit, start, nr); bitmap_clear(&madvise, start, nr); bitmap_clear(&always, start, nr); + bitmap_clear(&defer, start, nr); + } else if (!strcmp(policy, "defer")) { + bitmap_set(&defer, start, nr); + bitmap_clear(&madvise, start, nr); + bitmap_clear(&always, start, nr); + bitmap_clear(&inherit, start, nr); } else if (!strcmp(policy, "never")) { bitmap_clear(&inherit, start, nr); bitmap_clear(&madvise, start, nr); bitmap_clear(&always, start, nr); + bitmap_clear(&defer, start, nr); + } else { pr_err("invalid policy %s in thp_anon boot parameter\n", policy); goto err; @@ -1075,6 +1101,8 @@ static int __init setup_thp_anon(char *str) huge_anon_orders_always = always; huge_anon_orders_madvise = madvise; huge_anon_orders_inherit = inherit; + huge_anon_orders_defer = defer; + anon_orders_configured = true; return 1;
diff --git a/mm/khugepaged.c b/mm/khugepaged.c index 0723b184c7a4..428060495c49 100644 --- a/mm/khugepaged.c +++ b/mm/khugepaged.c @@ -491,7 +491,7 @@ void khugepaged_enter_vma(struct vm_area_struct *vma, { if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) && hugepage_pmd_enabled()) { - if (thp_vma_allowable_order(vma, vm_flags, TVA_ENFORCE_SYSFS, + if (thp_vma_allowable_order(vma, vm_flags, TVA_IN_KHUGEPAGE, PMD_ORDER)) __khugepaged_enter(vma->vm_mm); } @@ -955,7 +955,7 @@ static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, struct collapse_control *cc, int order) { struct vm_area_struct *vma; - unsigned long tva_flags = cc->is_khugepaged ? TVA_ENFORCE_SYSFS : 0; + unsigned long tva_flags = cc->is_khugepaged ? TVA_IN_KHUGEPAGE : 0;
if (unlikely(khugepaged_test_exit_or_disable(mm))) return SCAN_ANY_PROCESS; @@ -1434,7 +1434,7 @@ static int khugepaged_scan_pmd(struct mm_struct *mm, bool writable = false; int chunk_none_count = 0; int scaled_none = khugepaged_max_ptes_none >> (HPAGE_PMD_ORDER - KHUGEPAGED_MIN_MTHP_ORDER); - unsigned long tva_flags = cc->is_khugepaged ? TVA_ENFORCE_SYSFS : 0; + unsigned long tva_flags = cc->is_khugepaged ? TVA_IN_KHUGEPAGE : 0; VM_BUG_ON(address & ~HPAGE_PMD_MASK);
result = find_pmd_or_thp_or_none(mm, address, &pmd); @@ -2626,7 +2626,7 @@ static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result, break; } if (!thp_vma_allowable_order(vma, vma->vm_flags, - TVA_ENFORCE_SYSFS, PMD_ORDER)) { + TVA_IN_KHUGEPAGE, PMD_ORDER)) { skip: progress++; continue;
add the defer setting to the selftests library for reading thp settings.
Signed-off-by: Nico Pache npache@redhat.com --- tools/testing/selftests/mm/thp_settings.c | 1 + tools/testing/selftests/mm/thp_settings.h | 1 + 2 files changed, 2 insertions(+)
diff --git a/tools/testing/selftests/mm/thp_settings.c b/tools/testing/selftests/mm/thp_settings.c index ad872af1c81a..b2f9f62b302a 100644 --- a/tools/testing/selftests/mm/thp_settings.c +++ b/tools/testing/selftests/mm/thp_settings.c @@ -20,6 +20,7 @@ static const char * const thp_enabled_strings[] = { "always", "inherit", "madvise", + "defer", NULL };
diff --git a/tools/testing/selftests/mm/thp_settings.h b/tools/testing/selftests/mm/thp_settings.h index fc131d23d593..0d52e6d4f754 100644 --- a/tools/testing/selftests/mm/thp_settings.h +++ b/tools/testing/selftests/mm/thp_settings.h @@ -11,6 +11,7 @@ enum thp_enabled { THP_ALWAYS, THP_INHERIT, THP_MADVISE, + THP_DEFER, };
enum thp_defrag {
On Thu, May 15, 2025 at 11:41 AM Nico Pache npache@redhat.com wrote:
This series is a follow-up to [1], which adds mTHP support to khugepaged. mTHP khugepaged support is a "loose" dependency for the sysfs/sysctl configs to make sense. Without it global="defer" and mTHP="inherit" case is "undefined" behavior.
We've seen cases were customers switching from RHEL7 to RHEL8 see a significant increase in the memory footprint for the same workloads.
Through our investigations we found that a large contributing factor to the increase in RSS was an increase in THP usage.
For workloads like MySQL, or when using allocators like jemalloc, it is often recommended to set /transparent_hugepages/enabled=never. This is in part due to performance degradations and increased memory waste.
This series introduces enabled=defer, this setting acts as a middle ground between always and madvise. If the mapping is MADV_HUGEPAGE, the page fault handler will act normally, making a hugepage if possible. If the allocation is not MADV_HUGEPAGE, then the page fault handler will default to the base size allocation. The caveat is that khugepaged can still operate on pages that are not MADV_HUGEPAGE.
This allows for three things... one, applications specifically designed to use hugepages will get them, and two, applications that don't use hugepages can still benefit from them without aggressively inserting THPs at every possible chance. This curbs the memory waste, and defers the use of hugepages to khugepaged. Khugepaged can then scan the memory for eligible collapsing. Lastly there is the added benefit for those who want THPs but experience higher latency PFs. Now you can get base page performance at the PF handler and Hugepage performance for those mappings after they collapse.
Admins may want to lower max_ptes_none, if not, khugepaged may aggressively collapse single allocations into hugepages.
TESTING:
Built for x86_64, aarch64, ppc64le, and s390x
selftests mm
In [1] I provided a script [2] that has multiple access patterns
lots of general use.
redis testing. This test was my original case for the defer mode. What I was able to prove was that THP=always leads to increased max_latency cases; hence why it is recommended to disable THPs for redis servers. However with 'defer' we dont have the max_latency spikes and can still get the system to utilize THPs. I further tested this with the mTHP defer setting and found that redis (and probably other jmalloc users) can utilize THPs via defer (+mTHP defer) without a large latency penalty and some potential gains. I uploaded some mmtest results here[3] which compares: stock+thp=never stock+(m)thp=always khugepaged-mthp + defer (max_ptes_none=64)
The results show that (m)THPs can cause some throughput regression in some cases, but also has gains in other cases. The mTHP+defer results have more gains and less losses over the (m)THP=always case.
V6 Changes:
- nits
- rebased dependent series and added review tags
V5 Changes:
- rebased dependent series
- added reviewed-by tag on 2/4
V4 Changes:
- Minor Documentation fixes
- rebased the dependent series [1] onto mm-unstable commit 0e68b850b1d3 ("vmalloc: use atomic_long_add_return_relaxed()")
V3 Changes:
- Combined the documentation commits into one, and moved a section to the khugepaged mthp patchset
V2 Changes:
- base changes on mTHP khugepaged support
- Fix selftests parsing issue
- add mTHP defer option
- add mTHP defer Documentation
[1] - https://lore.kernel.org/all/20250515032226.128900-1-npache@redhat.com/ [2] - https://gitlab.com/npache/khugepaged_mthp_test [3] - https://people.redhat.com/npache/mthp_khugepaged_defer/testoutput2/output.ht...
Nico Pache (4): mm: defer THP insertion to khugepaged mm: document (m)THP defer usage khugepaged: add defer option to mTHP options selftests: mm: add defer to thp setting parser
Documentation/admin-guide/mm/transhuge.rst | 31 +++++++--- include/linux/huge_mm.h | 18 +++++- mm/huge_memory.c | 69 +++++++++++++++++++--- mm/khugepaged.c | 8 +-- tools/testing/selftests/mm/thp_settings.c | 1 + tools/testing/selftests/mm/thp_settings.h | 1 + 6 files changed, 106 insertions(+), 22 deletions(-)
-- 2.49.0
Hello Nico,
Upon reviewing the series, it occurred to me that BPF could solve this more cleanly. Adding a 'tva_flags' parameter to the BPF hook would handle this case and future scenarios without requiring new modes. The BPF mode could then serve as a unified solution.
On Tue, May 20, 2025 at 3:25 AM Yafang Shao laoar.shao@gmail.com wrote:
On Thu, May 15, 2025 at 11:41 AM Nico Pache npache@redhat.com wrote:
This series is a follow-up to [1], which adds mTHP support to khugepaged. mTHP khugepaged support is a "loose" dependency for the sysfs/sysctl configs to make sense. Without it global="defer" and mTHP="inherit" case is "undefined" behavior.
We've seen cases were customers switching from RHEL7 to RHEL8 see a significant increase in the memory footprint for the same workloads.
Through our investigations we found that a large contributing factor to the increase in RSS was an increase in THP usage.
For workloads like MySQL, or when using allocators like jemalloc, it is often recommended to set /transparent_hugepages/enabled=never. This is in part due to performance degradations and increased memory waste.
This series introduces enabled=defer, this setting acts as a middle ground between always and madvise. If the mapping is MADV_HUGEPAGE, the page fault handler will act normally, making a hugepage if possible. If the allocation is not MADV_HUGEPAGE, then the page fault handler will default to the base size allocation. The caveat is that khugepaged can still operate on pages that are not MADV_HUGEPAGE.
This allows for three things... one, applications specifically designed to use hugepages will get them, and two, applications that don't use hugepages can still benefit from them without aggressively inserting THPs at every possible chance. This curbs the memory waste, and defers the use of hugepages to khugepaged. Khugepaged can then scan the memory for eligible collapsing. Lastly there is the added benefit for those who want THPs but experience higher latency PFs. Now you can get base page performance at the PF handler and Hugepage performance for those mappings after they collapse.
Admins may want to lower max_ptes_none, if not, khugepaged may aggressively collapse single allocations into hugepages.
TESTING:
Built for x86_64, aarch64, ppc64le, and s390x
selftests mm
In [1] I provided a script [2] that has multiple access patterns
lots of general use.
redis testing. This test was my original case for the defer mode. What I was able to prove was that THP=always leads to increased max_latency cases; hence why it is recommended to disable THPs for redis servers. However with 'defer' we dont have the max_latency spikes and can still get the system to utilize THPs. I further tested this with the mTHP defer setting and found that redis (and probably other jmalloc users) can utilize THPs via defer (+mTHP defer) without a large latency penalty and some potential gains. I uploaded some mmtest results here[3] which compares: stock+thp=never stock+(m)thp=always khugepaged-mthp + defer (max_ptes_none=64)
The results show that (m)THPs can cause some throughput regression in some cases, but also has gains in other cases. The mTHP+defer results have more gains and less losses over the (m)THP=always case.
V6 Changes:
- nits
- rebased dependent series and added review tags
V5 Changes:
- rebased dependent series
- added reviewed-by tag on 2/4
V4 Changes:
- Minor Documentation fixes
- rebased the dependent series [1] onto mm-unstable commit 0e68b850b1d3 ("vmalloc: use atomic_long_add_return_relaxed()")
V3 Changes:
- Combined the documentation commits into one, and moved a section to the khugepaged mthp patchset
V2 Changes:
- base changes on mTHP khugepaged support
- Fix selftests parsing issue
- add mTHP defer option
- add mTHP defer Documentation
[1] - https://lore.kernel.org/all/20250515032226.128900-1-npache@redhat.com/ [2] - https://gitlab.com/npache/khugepaged_mthp_test [3] - https://people.redhat.com/npache/mthp_khugepaged_defer/testoutput2/output.ht...
Nico Pache (4): mm: defer THP insertion to khugepaged mm: document (m)THP defer usage khugepaged: add defer option to mTHP options selftests: mm: add defer to thp setting parser
Documentation/admin-guide/mm/transhuge.rst | 31 +++++++--- include/linux/huge_mm.h | 18 +++++- mm/huge_memory.c | 69 +++++++++++++++++++--- mm/khugepaged.c | 8 +-- tools/testing/selftests/mm/thp_settings.c | 1 + tools/testing/selftests/mm/thp_settings.h | 1 + 6 files changed, 106 insertions(+), 22 deletions(-)
-- 2.49.0
Hello Nico,
Upon reviewing the series, it occurred to me that BPF could solve this more cleanly. Adding a 'tva_flags' parameter to the BPF hook would handle this case and future scenarios without requiring new modes. The BPF mode could then serve as a unified solution.
Hi Yafang,
I dont see how this is the case? This would require users to modify/add functionality rather than configuring the system in this manner. What if BPF is not configured or being used? Having to use an additional technology that requires precise configuration doesn't seem cleaner.
Either way, thank you for taking a look into the series !
-- Nico
-- Regards Yafang
On Wed, May 21, 2025 at 6:19 PM Nico Pache npache@redhat.com wrote:
On Tue, May 20, 2025 at 3:25 AM Yafang Shao laoar.shao@gmail.com wrote:
On Thu, May 15, 2025 at 11:41 AM Nico Pache npache@redhat.com wrote:
This series is a follow-up to [1], which adds mTHP support to khugepaged. mTHP khugepaged support is a "loose" dependency for the sysfs/sysctl configs to make sense. Without it global="defer" and mTHP="inherit" case is "undefined" behavior.
We've seen cases were customers switching from RHEL7 to RHEL8 see a significant increase in the memory footprint for the same workloads.
Through our investigations we found that a large contributing factor to the increase in RSS was an increase in THP usage.
For workloads like MySQL, or when using allocators like jemalloc, it is often recommended to set /transparent_hugepages/enabled=never. This is in part due to performance degradations and increased memory waste.
This series introduces enabled=defer, this setting acts as a middle ground between always and madvise. If the mapping is MADV_HUGEPAGE, the page fault handler will act normally, making a hugepage if possible. If the allocation is not MADV_HUGEPAGE, then the page fault handler will default to the base size allocation. The caveat is that khugepaged can still operate on pages that are not MADV_HUGEPAGE.
This allows for three things... one, applications specifically designed to use hugepages will get them, and two, applications that don't use hugepages can still benefit from them without aggressively inserting THPs at every possible chance. This curbs the memory waste, and defers the use of hugepages to khugepaged. Khugepaged can then scan the memory for eligible collapsing. Lastly there is the added benefit for those who want THPs but experience higher latency PFs. Now you can get base page performance at the PF handler and Hugepage performance for those mappings after they collapse.
Admins may want to lower max_ptes_none, if not, khugepaged may aggressively collapse single allocations into hugepages.
TESTING:
Built for x86_64, aarch64, ppc64le, and s390x
selftests mm
In [1] I provided a script [2] that has multiple access patterns
lots of general use.
redis testing. This test was my original case for the defer mode. What I was able to prove was that THP=always leads to increased max_latency cases; hence why it is recommended to disable THPs for redis servers. However with 'defer' we dont have the max_latency spikes and can still get the system to utilize THPs. I further tested this with the mTHP defer setting and found that redis (and probably other jmalloc users) can utilize THPs via defer (+mTHP defer) without a large latency penalty and some potential gains. I uploaded some mmtest results here[3] which compares: stock+thp=never stock+(m)thp=always khugepaged-mthp + defer (max_ptes_none=64)
The results show that (m)THPs can cause some throughput regression in some cases, but also has gains in other cases. The mTHP+defer results have more gains and less losses over the (m)THP=always case.
V6 Changes:
- nits
- rebased dependent series and added review tags
V5 Changes:
- rebased dependent series
- added reviewed-by tag on 2/4
V4 Changes:
- Minor Documentation fixes
- rebased the dependent series [1] onto mm-unstable commit 0e68b850b1d3 ("vmalloc: use atomic_long_add_return_relaxed()")
V3 Changes:
- Combined the documentation commits into one, and moved a section to the khugepaged mthp patchset
V2 Changes:
- base changes on mTHP khugepaged support
- Fix selftests parsing issue
- add mTHP defer option
- add mTHP defer Documentation
[1] - https://lore.kernel.org/all/20250515032226.128900-1-npache@redhat.com/ [2] - https://gitlab.com/npache/khugepaged_mthp_test [3] - https://people.redhat.com/npache/mthp_khugepaged_defer/testoutput2/output.ht...
Nico Pache (4): mm: defer THP insertion to khugepaged mm: document (m)THP defer usage khugepaged: add defer option to mTHP options selftests: mm: add defer to thp setting parser
Documentation/admin-guide/mm/transhuge.rst | 31 +++++++--- include/linux/huge_mm.h | 18 +++++- mm/huge_memory.c | 69 +++++++++++++++++++--- mm/khugepaged.c | 8 +-- tools/testing/selftests/mm/thp_settings.c | 1 + tools/testing/selftests/mm/thp_settings.h | 1 + 6 files changed, 106 insertions(+), 22 deletions(-)
-- 2.49.0
Hello Nico,
Upon reviewing the series, it occurred to me that BPF could solve this more cleanly. Adding a 'tva_flags' parameter to the BPF hook would handle this case and future scenarios without requiring new modes. The BPF mode could then serve as a unified solution.
Hi Yafang,
I dont see how this is the case? This would require users to modify/add functionality rather than configuring the system in this manner. What if BPF is not configured or being used? Having to use an additional technology that requires precise configuration doesn't seem cleaner.
The core challenge remains: while certain tasks benefit from this new mode, others see no improvement—or may even regress. For that reason, implementing it globally seems unwise—per-task control would be far more effective.
On Wed, May 14, 2025 at 09:38:53PM -0600, Nico Pache wrote:
This series is a follow-up to [1], which adds mTHP support to khugepaged. mTHP khugepaged support is a "loose" dependency for the sysfs/sysctl configs to make sense. Without it global="defer" and mTHP="inherit" case is "undefined" behavior.
How can this be a follow up to an unmerged series? I'm confused by that.
And you're saying that you're introducing 'undefined behaviour' on the assumption that another series which seems to have quite a bit of discussion let to run will be merged?
While I'd understand if this was an RFC just to put the idea out there, you're not proposing it as such?
Unless there's a really good reason we're doing this way (I may be missing something), can we just have this as an RFC until the series it depends on is settled?
We've seen cases were customers switching from RHEL7 to RHEL8 see a significant increase in the memory footprint for the same workloads.
Through our investigations we found that a large contributing factor to the increase in RSS was an increase in THP usage.
For workloads like MySQL, or when using allocators like jemalloc, it is often recommended to set /transparent_hugepages/enabled=never. This is in part due to performance degradations and increased memory waste.
This series introduces enabled=defer, this setting acts as a middle ground between always and madvise. If the mapping is MADV_HUGEPAGE, the page fault handler will act normally, making a hugepage if possible. If the allocation is not MADV_HUGEPAGE, then the page fault handler will default to the base size allocation. The caveat is that khugepaged can still operate on pages that are not MADV_HUGEPAGE.
This allows for three things... one, applications specifically designed to use hugepages will get them, and two, applications that don't use hugepages can still benefit from them without aggressively inserting THPs at every possible chance. This curbs the memory waste, and defers the use of hugepages to khugepaged. Khugepaged can then scan the memory for eligible collapsing. Lastly there is the added benefit for those who want THPs but experience higher latency PFs. Now you can get base page performance at the PF handler and Hugepage performance for those mappings after they collapse.
Admins may want to lower max_ptes_none, if not, khugepaged may aggressively collapse single allocations into hugepages.
TESTING:
- Built for x86_64, aarch64, ppc64le, and s390x
- selftests mm
- In [1] I provided a script [2] that has multiple access patterns
- lots of general use.
OK so this truly is dependent on the unmerged series? Or isn't it?
Is your testing based on that?
Because again... that surely makes this series a no-go until we land the prior (which might be changed, and thus necessitate re-testing).
Are you going to provide any of these numbers/data anywhere?
redis testing. This test was my original case for the defer mode. What I was able to prove was that THP=always leads to increased max_latency cases; hence why it is recommended to disable THPs for redis servers. However with 'defer' we dont have the max_latency spikes and can still get the system to utilize THPs. I further tested this with the mTHP defer setting and found that redis (and probably other jmalloc users) can utilize THPs via defer (+mTHP defer) without a large latency penalty and some potential gains. I uploaded some mmtest results here[3] which compares: stock+thp=never stock+(m)thp=always khugepaged-mthp + defer (max_ptes_none=64)
The results show that (m)THPs can cause some throughput regression in some cases, but also has gains in other cases. The mTHP+defer results have more gains and less losses over the (m)THP=always case.
V6 Changes:
- nits
- rebased dependent series and added review tags
V5 Changes:
- rebased dependent series
- added reviewed-by tag on 2/4
V4 Changes:
- Minor Documentation fixes
- rebased the dependent series [1] onto mm-unstable commit 0e68b850b1d3 ("vmalloc: use atomic_long_add_return_relaxed()")
V3 Changes:
- Combined the documentation commits into one, and moved a section to the khugepaged mthp patchset
V2 Changes:
- base changes on mTHP khugepaged support
- Fix selftests parsing issue
- add mTHP defer option
- add mTHP defer Documentation
[1] - https://lore.kernel.org/all/20250515032226.128900-1-npache@redhat.com/ [2] - https://gitlab.com/npache/khugepaged_mthp_test [3] - https://people.redhat.com/npache/mthp_khugepaged_defer/testoutput2/output.ht...
Nico Pache (4): mm: defer THP insertion to khugepaged mm: document (m)THP defer usage khugepaged: add defer option to mTHP options selftests: mm: add defer to thp setting parser
Documentation/admin-guide/mm/transhuge.rst | 31 +++++++--- include/linux/huge_mm.h | 18 +++++- mm/huge_memory.c | 69 +++++++++++++++++++--- mm/khugepaged.c | 8 +-- tools/testing/selftests/mm/thp_settings.c | 1 + tools/testing/selftests/mm/thp_settings.h | 1 + 6 files changed, 106 insertions(+), 22 deletions(-)
-- 2.49.0
On Tue, May 20, 2025 at 3:43 AM Lorenzo Stoakes lorenzo.stoakes@oracle.com wrote:
On Wed, May 14, 2025 at 09:38:53PM -0600, Nico Pache wrote:
This series is a follow-up to [1], which adds mTHP support to khugepaged. mTHP khugepaged support is a "loose" dependency for the sysfs/sysctl configs to make sense. Without it global="defer" and mTHP="inherit" case is "undefined" behavior.
How can this be a follow up to an unmerged series? I'm confused by that.
Hi Lorenzo,
follow up or loose dependency. Not sure the correct terminology.
Either way, as I was developing this as a potential solution for the THP internal fragmentation issue, upstream was working on adding mTHPs. By adding a new THP sysctl entry I noticed mTHP would now be missing the same entry. Furthermore I was told mTHP support for khugepaged was a desire, so I began working on it in conjunction. So given the undefined behavior of defer globally while any mix of mTHP settings, it became dependent on the khugepaged support. Either way patch 1 of this series is the core functionality. The rest is to fill the undefined behavior gap.
And you're saying that you're introducing 'undefined behaviour' on the assumption that another series which seems to have quite a bit of discussion let to run will be merged?
This could technically get merged without the mTHP khugepaged changes, but then the reviews would probably all be pointing out what I pointed out above. Chicken or Egg problem...
While I'd understand if this was an RFC just to put the idea out there, you're not proposing it as such?
Nope we've already discussed this in both the MM alignment and thp upstream meetings, no one was opposing it, and a lot of testing was done-- by me, RH's CI, and our perf teams. Ive posted several RFCs before posting a patchset.
Unless there's a really good reason we're doing this way (I may be missing something), can we just have this as an RFC until the series it depends on is settled?
Hopefully paragraph one clears this up! They were built in conjunction, but posting them as one series didn't feel right (and IIRC this was also discussed, and this was decided).
We've seen cases were customers switching from RHEL7 to RHEL8 see a significant increase in the memory footprint for the same workloads.
Through our investigations we found that a large contributing factor to the increase in RSS was an increase in THP usage.
For workloads like MySQL, or when using allocators like jemalloc, it is often recommended to set /transparent_hugepages/enabled=never. This is in part due to performance degradations and increased memory waste.
This series introduces enabled=defer, this setting acts as a middle ground between always and madvise. If the mapping is MADV_HUGEPAGE, the page fault handler will act normally, making a hugepage if possible. If the allocation is not MADV_HUGEPAGE, then the page fault handler will default to the base size allocation. The caveat is that khugepaged can still operate on pages that are not MADV_HUGEPAGE.
This allows for three things... one, applications specifically designed to use hugepages will get them, and two, applications that don't use hugepages can still benefit from them without aggressively inserting THPs at every possible chance. This curbs the memory waste, and defers the use of hugepages to khugepaged. Khugepaged can then scan the memory for eligible collapsing. Lastly there is the added benefit for those who want THPs but experience higher latency PFs. Now you can get base page performance at the PF handler and Hugepage performance for those mappings after they collapse.
Admins may want to lower max_ptes_none, if not, khugepaged may aggressively collapse single allocations into hugepages.
TESTING:
- Built for x86_64, aarch64, ppc64le, and s390x
- selftests mm
- In [1] I provided a script [2] that has multiple access patterns
- lots of general use.
OK so this truly is dependent on the unmerged series? Or isn't it?
Is your testing based on that?
Most of the testing was done in conjunction, but independent testing was also done on this series (including by a large customer that was itching to try the changes, and they were very satisfied with the results).
Because again... that surely makes this series a no-go until we land the prior (which might be changed, and thus necessitate re-testing).
Are you going to provide any of these numbers/data anywhere?
There is a link to the results in this cover letter [3] - https://people.redhat.com/npache/mthp_khugepaged_defer/testoutput2/output.ht...
redis testing. This test was my original case for the defer mode. What I was able to prove was that THP=always leads to increased max_latency cases; hence why it is recommended to disable THPs for redis servers. However with 'defer' we dont have the max_latency spikes and can still get the system to utilize THPs. I further tested this with the mTHP defer setting and found that redis (and probably other jmalloc users) can utilize THPs via defer (+mTHP defer) without a large latency penalty and some potential gains. I uploaded some mmtest results here[3] which compares: stock+thp=never stock+(m)thp=always khugepaged-mthp + defer (max_ptes_none=64)
The results show that (m)THPs can cause some throughput regression in some cases, but also has gains in other cases. The mTHP+defer results have more gains and less losses over the (m)THP=always case.
V6 Changes:
- nits
- rebased dependent series and added review tags
V5 Changes:
- rebased dependent series
- added reviewed-by tag on 2/4
V4 Changes:
- Minor Documentation fixes
- rebased the dependent series [1] onto mm-unstable commit 0e68b850b1d3 ("vmalloc: use atomic_long_add_return_relaxed()")
V3 Changes:
- Combined the documentation commits into one, and moved a section to the khugepaged mthp patchset
V2 Changes:
- base changes on mTHP khugepaged support
- Fix selftests parsing issue
- add mTHP defer option
- add mTHP defer Documentation
[1] - https://lore.kernel.org/all/20250515032226.128900-1-npache@redhat.com/ [2] - https://gitlab.com/npache/khugepaged_mthp_test [3] - https://people.redhat.com/npache/mthp_khugepaged_defer/testoutput2/output.ht...
Nico Pache (4): mm: defer THP insertion to khugepaged mm: document (m)THP defer usage khugepaged: add defer option to mTHP options selftests: mm: add defer to thp setting parser
Documentation/admin-guide/mm/transhuge.rst | 31 +++++++--- include/linux/huge_mm.h | 18 +++++- mm/huge_memory.c | 69 +++++++++++++++++++--- mm/khugepaged.c | 8 +-- tools/testing/selftests/mm/thp_settings.c | 1 + tools/testing/selftests/mm/thp_settings.h | 1 + 6 files changed, 106 insertions(+), 22 deletions(-)
-- 2.49.0
To start with I do apologise for coming to this at v6, I realise it's irritating to have push back at this late stage. This is more so my attempt to understand where this series -sits- so I can properly review it.
So please bear with me here :)
So, I remain very confused. This may just be a _me_ thing here :)
So let me check my understanding:
1. This series introduces this new THP deferred mode. 2. By 'follow-up' really you mean 'inspired by' or 'related to' right? 3. If this series lands before [1], commits 2 - 4 are 'undefined behaviour'.
In my view if 3 is true this series should be RFC until [1] merges.
If I've got it wrong and this needs to land first, we should RFC [1].
That way we can un-RFC once the dependency is met.
We have about 5 [m]THP series in flight at the moment, all touching at least vaguely related stuff, so any help for reviewers would be hugely appreciated thanks :)
On Wed, May 21, 2025 at 04:41:54AM -0600, Nico Pache wrote:
On Tue, May 20, 2025 at 3:43 AM Lorenzo Stoakes lorenzo.stoakes@oracle.com wrote:
On Wed, May 14, 2025 at 09:38:53PM -0600, Nico Pache wrote:
This series is a follow-up to [1], which adds mTHP support to khugepaged. mTHP khugepaged support is a "loose" dependency for the sysfs/sysctl configs to make sense. Without it global="defer" and mTHP="inherit" case is "undefined" behavior.
How can this be a follow up to an unmerged series? I'm confused by that.
Hi Lorenzo,
follow up or loose dependency. Not sure the correct terminology.
See above. Let's nail this down please.
Either way, as I was developing this as a potential solution for the THP internal fragmentation issue, upstream was working on adding mTHPs. By adding a new THP sysctl entry I noticed mTHP would now be missing the same entry. Furthermore I was told mTHP support for khugepaged was a desire, so I began working on it in conjunction. So given the undefined behavior of defer globally while any mix of mTHP settings, it became dependent on the khugepaged support. Either way patch 1 of this series is the core functionality. The rest is to fill the undefined behavior gap.
And you're saying that you're introducing 'undefined behaviour' on the assumption that another series which seems to have quite a bit of discussion let to run will be merged?
This could technically get merged without the mTHP khugepaged changes, but then the reviews would probably all be pointing out what I pointed out above. Chicken or Egg problem...
While I'd understand if this was an RFC just to put the idea out there, you're not proposing it as such?
Nope we've already discussed this in both the MM alignment and thp upstream meetings, no one was opposing it, and a lot of testing was done-- by me, RH's CI, and our perf teams. Ive posted several RFCs before posting a patchset.
Unless there's a really good reason we're doing this way (I may be missing something), can we just have this as an RFC until the series it depends on is settled?
Hopefully paragraph one clears this up! They were built in conjunction, but posting them as one series didn't feel right (and IIRC this was also discussed, and this was decided).
'This was also discussed and this was decided' :)
I'm guessing rather you mean discussion was had with other reviewers and of course our earstwhile THP maintainer David, and you guys decided this made more sense?
Obviously upstream discussion is what counts, but as annoying as it is, one does have to address the concerns of reviewers even if late to a series (again, apologies for this).
So, to be clear - I'm not intending to hold up or block the series, I just want to understand how things are, this is the purpose here.
Thanks!
We've seen cases were customers switching from RHEL7 to RHEL8 see a significant increase in the memory footprint for the same workloads.
Through our investigations we found that a large contributing factor to the increase in RSS was an increase in THP usage.
For workloads like MySQL, or when using allocators like jemalloc, it is often recommended to set /transparent_hugepages/enabled=never. This is in part due to performance degradations and increased memory waste.
This series introduces enabled=defer, this setting acts as a middle ground between always and madvise. If the mapping is MADV_HUGEPAGE, the page fault handler will act normally, making a hugepage if possible. If the allocation is not MADV_HUGEPAGE, then the page fault handler will default to the base size allocation. The caveat is that khugepaged can still operate on pages that are not MADV_HUGEPAGE.
This allows for three things... one, applications specifically designed to use hugepages will get them, and two, applications that don't use hugepages can still benefit from them without aggressively inserting THPs at every possible chance. This curbs the memory waste, and defers the use of hugepages to khugepaged. Khugepaged can then scan the memory for eligible collapsing. Lastly there is the added benefit for those who want THPs but experience higher latency PFs. Now you can get base page performance at the PF handler and Hugepage performance for those mappings after they collapse.
Admins may want to lower max_ptes_none, if not, khugepaged may aggressively collapse single allocations into hugepages.
TESTING:
- Built for x86_64, aarch64, ppc64le, and s390x
- selftests mm
- In [1] I provided a script [2] that has multiple access patterns
- lots of general use.
OK so this truly is dependent on the unmerged series? Or isn't it?
Is your testing based on that?
Most of the testing was done in conjunction, but independent testing was also done on this series (including by a large customer that was itching to try the changes, and they were very satisfied with the results).
You should make this very clear in the cover letter.
Because again... that surely makes this series a no-go until we land the prior (which might be changed, and thus necessitate re-testing).
Are you going to provide any of these numbers/data anywhere?
There is a link to the results in this cover letter [3] - https://people.redhat.com/npache/mthp_khugepaged_defer/testoutput2/output.ht...
Ultimately it's not ok in mm to have a link to a website that might go away any time, these cover letters are 'baked in' to the commit log. Are you sure this website with 'testoutput2' will exist in 10 years time? :)
You should at the very least add a summary of this data in the cover letter, perhaps referring back to this link as 'at the time of writing full results are available at' or something like this.
redis testing. This test was my original case for the defer mode. What I was able to prove was that THP=always leads to increased max_latency cases; hence why it is recommended to disable THPs for redis servers. However with 'defer' we dont have the max_latency spikes and can still get the system to utilize THPs. I further tested this with the mTHP defer setting and found that redis (and probably other jmalloc users) can utilize THPs via defer (+mTHP defer) without a large latency penalty and some potential gains. I uploaded some mmtest results here[3] which compares: stock+thp=never stock+(m)thp=always khugepaged-mthp + defer (max_ptes_none=64)
The results show that (m)THPs can cause some throughput regression in some cases, but also has gains in other cases. The mTHP+defer results have more gains and less losses over the (m)THP=always case.
V6 Changes:
- nits
- rebased dependent series and added review tags
V5 Changes:
- rebased dependent series
- added reviewed-by tag on 2/4
V4 Changes:
- Minor Documentation fixes
- rebased the dependent series [1] onto mm-unstable commit 0e68b850b1d3 ("vmalloc: use atomic_long_add_return_relaxed()")
V3 Changes:
- Combined the documentation commits into one, and moved a section to the khugepaged mthp patchset
V2 Changes:
- base changes on mTHP khugepaged support
- Fix selftests parsing issue
- add mTHP defer option
- add mTHP defer Documentation
[1] - https://lore.kernel.org/all/20250515032226.128900-1-npache@redhat.com/ [2] - https://gitlab.com/npache/khugepaged_mthp_test [3] - https://people.redhat.com/npache/mthp_khugepaged_defer/testoutput2/output.ht...
Nico Pache (4): mm: defer THP insertion to khugepaged mm: document (m)THP defer usage khugepaged: add defer option to mTHP options selftests: mm: add defer to thp setting parser
Documentation/admin-guide/mm/transhuge.rst | 31 +++++++--- include/linux/huge_mm.h | 18 +++++- mm/huge_memory.c | 69 +++++++++++++++++++--- mm/khugepaged.c | 8 +-- tools/testing/selftests/mm/thp_settings.c | 1 + tools/testing/selftests/mm/thp_settings.h | 1 + 6 files changed, 106 insertions(+), 22 deletions(-)
-- 2.49.0
Cheers, Lorenzo
On 21.05.25 13:24, Lorenzo Stoakes wrote:
To start with I do apologise for coming to this at v6, I realise it's irritating to have push back at this late stage. This is more so my attempt to understand where this series -sits- so I can properly review it.
So please bear with me here :)
So, I remain very confused. This may just be a _me_ thing here :)
So let me check my understanding:
- This series introduces this new THP deferred mode.
- By 'follow-up' really you mean 'inspired by' or 'related to' right?
- If this series lands before [1], commits 2 - 4 are 'undefined behaviour'.
In my view if 3 is true this series should be RFC until [1] merges.
If I've got it wrong and this needs to land first, we should RFC [1].
That way we can un-RFC once the dependency is met.
I really don't have a strong opinion on the RFC vs. !RFC like others here -- as long as the dependency is obvious. I treat RFC more as a "rough idea" than well tested work.
Anyhow, to me the dependency is obvious, but I've followed the MM meeting discussions, development etc.
I interpret "follow up" as "depends on" here. Likely we should have spelled out "This series depends on the patch series X that was not merged yet, and likely a new version will be required once merged.".
In this particular case, maybe we should just have sent one initial RFC, and then rebased it on top of the other work on a public git branch (linked from the RFC cover letter).
Once the dependency gets merged, we could just resend the series. Looking at the changelog, only minor stuff changed (mostly rebasing etc).
Moving forward, I don't think there is the need to resend as long as the dependency isn't merged upstream (or close to being merged upstream) yet.
Because again... that surely makes this series a no-go until we land the prior (which might be changed, and thus necessitate re-testing).
Are you going to provide any of these numbers/data anywhere?
There is a link to the results in this cover letter [3] - https://people.redhat.com/npache/mthp_khugepaged_defer/testoutput2/output.ht...
Ultimately it's not ok in mm to have a link to a website that might go away any time, these cover letters are 'baked in' to the commit log. Are you sure this website with 'testoutput2' will exist in 10 years time? :)
You should at the very least add a summary of this data in the cover letter, perhaps referring back to this link as 'at the time of writing full results are available at' or something like this.
Yeah, or of they were included in some other mail, we can link to that mail in lore.
I think the TL;DR here to avoid too much back and forth is - let's please make this super super simple :)
I would prefer anything that has a dependency to just sit in RFC until the dependency is merged.
Or, alternatively, to have a big note at the top:
ANDREW - Please do not merge in mm-unstable until series [1] is merged, and when that is merged please ping for a resend.
Or whatever it might be.
On Wed, May 21, 2025 at 01:46:38PM +0200, David Hildenbrand wrote:
On 21.05.25 13:24, Lorenzo Stoakes wrote:
To start with I do apologise for coming to this at v6, I realise it's irritating to have push back at this late stage. This is more so my attempt to understand where this series -sits- so I can properly review it.
So please bear with me here :)
So, I remain very confused. This may just be a _me_ thing here :)
So let me check my understanding:
- This series introduces this new THP deferred mode.
- By 'follow-up' really you mean 'inspired by' or 'related to' right?
- If this series lands before [1], commits 2 - 4 are 'undefined behaviour'.
In my view if 3 is true this series should be RFC until [1] merges.
If I've got it wrong and this needs to land first, we should RFC [1].
That way we can un-RFC once the dependency is met.
I really don't have a strong opinion on the RFC vs. !RFC like others here -- as long as the dependency is obvious. I treat RFC more as a "rough idea" than well tested work.
Anyhow, to me the dependency is obvious, but I've followed the MM meeting discussions, development etc.
Right but is it clear to Andrew? I mean the cover letter was super unclear to me.
What's to prevent things getting merged out of order? And do people 'just have to remember' to resend? And a resend doesn't necessarily mean patch set X will come after patch set Y.
If there's a requirement related to the ordering of these series it really has to be expressed very clearly.
(by the way, I feel expressing things like this is a kind of area where we have _some kind_ of a break down in kernel process or it'd be nice to have tags or something to properly express this sort of thing. But maybe another discussion :)
I interpret "follow up" as "depends on" here. Likely we should have spelled out "This series depends on the patch series X that was not merged yet, and likely a new version will be required once merged.".
In this particular case, maybe we should just have sent one initial RFC, and then rebased it on top of the other work on a public git branch (linked from the RFC cover letter).
Once the dependency gets merged, we could just resend the series. Looking at the changelog, only minor stuff changed (mostly rebasing etc).
Moving forward, I don't think there is the need to resend as long as the dependency isn't merged upstream (or close to being merged upstream) yet.
I mean this is still 'just have to remember' stuff :)
Do we need patches 2-4 if the dependency isn't merged? That was unclear to me.
Because again... that surely makes this series a no-go until we land the prior (which might be changed, and thus necessitate re-testing).
Are you going to provide any of these numbers/data anywhere?
There is a link to the results in this cover letter [3] - https://people.redhat.com/npache/mthp_khugepaged_defer/testoutput2/output.ht...
Ultimately it's not ok in mm to have a link to a website that might go away any time, these cover letters are 'baked in' to the commit log. Are you sure this website with 'testoutput2' will exist in 10 years time? :)
You should at the very least add a summary of this data in the cover letter, perhaps referring back to this link as 'at the time of writing full results are available at' or something like this.
Yeah, or of they were included in some other mail, we can link to that mail in lore.
-- Cheers,
David / dhildenb
Anyhow, to me the dependency is obvious, but I've followed the MM meeting discussions, development etc.
Right but is it clear to Andrew? I mean the cover letter was super unclear to me.
I mean, assuming that it would not be clear to Andrew (and I think it is clear to Andrew), I we would get CCed on these emails and could immediately scream STOOOOOP :)
And until this would hit mm-stable, a bit more time would pass.
What's to prevent things getting merged out of order?
Fortunately, there are still people working here and not machines (at least, that's what I hope).
And do people 'just have to remember' to resend?
Yes, in this case Nico wants to get his stuff upstream and must drive it once the dependencies are met IMHO.
If there's a requirement related to the ordering of these series it really has to be expressed very clearly.
Jup. I'll note that for now there was no strict rule what to tag as RFC and what not that I know of. Of course, if people send broken, half-implemented, untested ... crap, it should *clearly* be RFC.
People should be spelling out dependencies in any case (especially for non-RFC versions) clearly.
I'll note that even if there would be a rule, I'm afraid we don't have a good place to document it (and not sure if people would find it or even try finding it ...) :/
A big problem is when some subsystems have their own rules for how to handle such things. That causes major pain for contributors ...
Fundamentally I trust you to make sure this all goes correctly so let's not belabour the point or delay things here :)
So in that vein, Nico - I would sugesst for future respins adding a really clear bit to the header as David suggested :) also update the cover letter tests so it isn't reliant on a possibly ephemeral web link.
But otherwise let's proceed as was.
On Wed, May 21, 2025 at 02:24:45PM +0200, David Hildenbrand wrote:
Anyhow, to me the dependency is obvious, but I've followed the MM meeting discussions, development etc.
Right but is it clear to Andrew? I mean the cover letter was super unclear to me.
I mean, assuming that it would not be clear to Andrew (and I think it is clear to Andrew), I we would get CCed on these emails and could immediately scream STOOOOOP :)
And until this would hit mm-stable, a bit more time would pass.
What's to prevent things getting merged out of order?
Fortunately, there are still people working here and not machines (at least, that's what I hope).
Obligatory link to this :P
https://www.youtube.com/watch?v=5lsExRvJTAI
And do people 'just have to remember' to resend?
Yes, in this case Nico wants to get his stuff upstream and must drive it once the dependencies are met IMHO.
If there's a requirement related to the ordering of these series it really has to be expressed very clearly.
Jup. I'll note that for now there was no strict rule what to tag as RFC and what not that I know of. Of course, if people send broken, half-implemented, untested ... crap, it should *clearly* be RFC.
People should be spelling out dependencies in any case (especially for non-RFC versions) clearly.
I'll note that even if there would be a rule, I'm afraid we don't have a good place to document it (and not sure if people would find it or even try finding it ...) :/
Yeah... :)
A big problem is when some subsystems have their own rules for how to handle such things. That causes major pain for contributors ...
Yeah, I wish there was something more general.
-- Cheers,
David / dhildenb
Cheers, Lorenzo
On 21.05.25 14:33, Lorenzo Stoakes wrote:
Fundamentally I trust you to make sure this all goes correctly so let's not belabour the point or delay things here :)
So in that vein, Nico - I would sugesst for future respins adding a really clear bit to the header as David suggested :) also update the cover letter tests so it isn't reliant on a possibly ephemeral web link.
But otherwise let's proceed as was.
Right, and maybe only post this series if there was a major change, otherwise wait until the other thing is on it's way upstream.
On Wed, May 21, 2025 at 02:24:45PM +0200, David Hildenbrand wrote:
Anyhow, to me the dependency is obvious, but I've followed the MM meeting discussions, development etc.
Right but is it clear to Andrew? I mean the cover letter was super unclear to me.
I mean, assuming that it would not be clear to Andrew (and I think it is clear to Andrew), I we would get CCed on these emails and could immediately scream STOOOOOP :)
And until this would hit mm-stable, a bit more time would pass.
What's to prevent things getting merged out of order?
Fortunately, there are still people working here and not machines (at least, that's what I hope).
Obligatory link to this :P
It's scary how relevant that has become lately :D
... fortunately, whenever I tell the chatbots that they are wrong (IOW, everytime I use them) they reply with "Oh yes, you are right." ... so far ...
linux-kselftest-mirror@lists.linaro.org
-
David Hildenbrand -
Lorenzo Stoakes -
Nico Pache -
Yafang Shao