hugetlbfs: Disable PMD sharing for large systems

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

Quoted 7 lines · Waiman Long

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

Can;t the application just specify MAP_POPULATE?

Quoted 8 lines · Matthew Wilcox

Quoted 7 lines · Waiman Long

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

Quoted 1 line

Can;t the application just specify MAP_POPULATE?

Originally, I thought that this happened in the startup phase when the pages were faulted in. The problem persists after steady state had been reached though. Every time you have a new user process created, it will have its own page table.

It is the sharing of the of huge page shared
memory that is causing problem. Of course, it depends on how the application is written.

Quoted 9 lines · Waiman Long

Quoted 8 lines · Matthew Wilcox

Quoted 7 lines · Waiman Long

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

Quoted 5 lines

Quoted 1 line

Can;t the application just specify MAP_POPULATE?

Originally, I thought that this happened in the startup phase when the pages were faulted in. The problem persists after steady state had been reached though. Every time you have a new user process created, it will have its own page table.

This is still at fault time. Although, for the particular application it may be after the 'startup phase'.

Quoted 3 lines

It is the sharing of the of huge page shared
memory that is causing problem. Of course, it depends on how the application is written.

It may be the case that some applications would find the delays acceptable for the benefit of shared pmds once they reach steady state. As you say, of course this depends on how the application is written.

I know that Oracle DB would not like it if PMD sharing is disabled for them. Based on what I know of their model, all processes which share PMDs perform faults (write or read) during the startup phase. This is in environments as big or bigger than you describe above. I have never looked at/for delays in these environments around pmd sharing (page faults), but that does not mean they do not exist. I will try to get the DB group to give me access to one of their large environments for analysis.

We may want to consider making the timeout value and disable threshold user configurable.

Quoted 33 lines · Mike Kravetz

Quoted 9 lines · Waiman Long

Quoted 8 lines · Matthew Wilcox

Quoted 7 lines · Waiman Long

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

Quoted 5 lines

Quoted 1 line

Can;t the application just specify MAP_POPULATE?

Originally, I thought that this happened in the startup phase when the pages were faulted in. The problem persists after steady state had been reached though. Every time you have a new user process created, it will have its own page table.

This is still at fault time. Although, for the particular application it may be after the 'startup phase'.

Quoted 3 lines

It is the sharing of the of huge page shared
memory that is causing problem. Of course, it depends on how the application is written.

It may be the case that some applications would find the delays acceptable for the benefit of shared pmds once they reach steady state. As you say, of course this depends on how the application is written.

I know that Oracle DB would not like it if PMD sharing is disabled for them. Based on what I know of their model, all processes which share PMDs perform faults (write or read) during the startup phase. This is in environments as big or bigger than you describe above. I have never looked at/for delays in these environments around pmd sharing (page faults), but that does not mean they do not exist. I will try to get the DB group to give me access to one of their large environments for analysis.

Quoted 2 lines

We may want to consider making the timeout value and disable threshold user configurable.

Making it configurable is certainly doable. They can be sysctl parameters so that the users can reenable PMD sharing by making those parameters larger.

Quoted 33 lines · Mike Kravetz

Quoted 9 lines · Waiman Long

Quoted 8 lines · Matthew Wilcox

Quoted 7 lines · Waiman Long

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

Quoted 5 lines

Quoted 1 line

Can;t the application just specify MAP_POPULATE?

Originally, I thought that this happened in the startup phase when the pages were faulted in. The problem persists after steady state had been reached though. Every time you have a new user process created, it will have its own page table.

This is still at fault time. Although, for the particular application it may be after the 'startup phase'.

Quoted 3 lines

It is the sharing of the of huge page shared
memory that is causing problem. Of course, it depends on how the application is written.

It may be the case that some applications would find the delays acceptable for the benefit of shared pmds once they reach steady state. As you say, of course this depends on how the application is written.

I know that Oracle DB would not like it if PMD sharing is disabled for them. Based on what I know of their model, all processes which share PMDs perform faults (write or read) during the startup phase. This is in environments as big or bigger than you describe above. I have never looked at/for delays in these environments around pmd sharing (page faults), but that does not mean they do not exist. I will try to get the DB group to give me access to one of their large environments for analysis.

Quoted 2 lines

We may want to consider making the timeout value and disable threshold user configurable.

Making it configurable is certainly doable. They can be sysctl parameters so that the users can reenable PMD sharing by making those parameters larger.

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

These random delays, however, are deemed unacceptable. The cause of that delay is the long wait time in acquiring the mmap_sem when trying to share the huge PMDs.

To remove the unacceptable delays, we have to limit the amount of wait time on the mmap_sem. So the new down_write_timedlock() function is used to acquire the write lock on the mmap_sem with a timeout value of 10ms which should not cause a perceivable delay. If timeout happens, the task will abandon its effort to share the PMD and allocate its own copy instead.

When too many timeouts happens (threshold currently set at 256), the system may be too large for PMD sharing to be useful without undue delay. So the sharing will be disabled in this case.

Signed-off-by Waiman Long <[email protected]>

---
include/linux/fs.h | 7 +++++++
mm/hugetlb.c | 24 +++++++++++++++++++++---

2 files changed, 28 insertions(+), 3 deletions(-)

include/linux/fs.h +7 -0

index 997a530ff4e9..e9d3ad465a6b 100644
@@ -40,6 +40,7 @@
 #include <linux/fs_types.h>
 #include <linux/build_bug.h>
 #include <linux/stddef.h>
+#include <linux/ktime.h>
 
 #include <asm/byteorder.h>
 #include <uapi/linux/fs.h>

@@ -519,6 +520,12 @@ static inline void i_mmap_lock_write(struct address_space *mapping)
 	down_write(&mapping->i_mmap_rwsem);
 }
 
+static inline bool i_mmap_timedlock_write(struct address_space *mapping,
+					 ktime_t timeout)
+{
+	return down_write_timedlock(&mapping->i_mmap_rwsem, timeout);
+}
+
 static inline void i_mmap_unlock_write(struct address_space *mapping)
 {
 	up_write(&mapping->i_mmap_rwsem);

mm/hugetlb.c +2 -0

index 6d7296dd11b8..445af661ae29 100644
@@ -4750,6 +4750,8 @@ void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
 	}
 }
 
+#define PMD_SHARE_DISABLE_THRESHOLD	(1 << 8)
+
 /*
 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
 * and returns the corresponding pte. While this is not necessary for the

@@ -4770,11 +4772,24 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	pte_t *spte = NULL;

mm/hugetlb.c +14 -1

+	 */
+	if (!vma_shareable(vma, addr) ||
+	   (atomic_read(&timeout_cnt) >= PMD_SHARE_DISABLE_THRESHOLD))
+		goto out_no_share;
+
+	if (!i_mmap_timedlock_write(mapping, ms_to_ktime(10))) {
+		if (atomic_inc_return(&timeout_cnt) ==
+		    PMD_SHARE_DISABLE_THRESHOLD)
+			pr_info("Hugetlbfs PMD sharing disabled because of timeouts!\n");
+		goto out_no_share;
+	}
 
-	i_mmap_lock_write(mapping);
 	vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
 		if (svma == vma)
 			continue;

@@ -4806,6 +4821,9 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	pte = (pte_t *)pmd_alloc(mm, pud, addr);
 	i_mmap_unlock_write(mapping);
 	return pte;
+
+out_no_share:
+	return (pte_t *)pmd_alloc(mm, pud, addr);
 }
 
 /*

Signature · 1 line

2.18.1

Quoted 76 lines · Waiman Long

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

These random delays, however, are deemed unacceptable. The cause of that delay is the long wait time in acquiring the mmap_sem when trying to share the huge PMDs.

To remove the unacceptable delays, we have to limit the amount of wait time on the mmap_sem. So the new down_write_timedlock() function is used to acquire the write lock on the mmap_sem with a timeout value of 10ms which should not cause a perceivable delay. If timeout happens, the task will abandon its effort to share the PMD and allocate its own copy instead.

When too many timeouts happens (threshold currently set at 256), the system may be too large for PMD sharing to be useful without undue delay. So the sharing will be disabled in this case.

Signed-off-by Waiman Long <[email protected]>

---
include/linux/fs.h | 7 +++++++
mm/hugetlb.c | 24 +++++++++++++++++++++---

2 files changed, 28 insertions(+), 3 deletions(-)

include/linux/fs.h +7 -0

index 997a530ff4e9..e9d3ad465a6b 100644
@@ -40,6 +40,7 @@
 #include <linux/fs_types.h>
 #include <linux/build_bug.h>
 #include <linux/stddef.h>
+#include <linux/ktime.h>
 
 #include <asm/byteorder.h>
 #include <uapi/linux/fs.h>

@@ -519,6 +520,12 @@ static inline void i_mmap_lock_write(struct address_space *mapping)
 	down_write(&mapping->i_mmap_rwsem);
 }
 
+static inline bool i_mmap_timedlock_write(struct address_space *mapping,
+					 ktime_t timeout)
+{
+	return down_write_timedlock(&mapping->i_mmap_rwsem, timeout);
+}
+
 static inline void i_mmap_unlock_write(struct address_space *mapping)
 {
 	up_write(&mapping->i_mmap_rwsem);

mm/hugetlb.c +3 -0

index 6d7296dd11b8..445af661ae29 100644
@@ -4750,6 +4750,8 @@ void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
 	}
 }
 
+#define PMD_SHARE_DISABLE_THRESHOLD	(1 << 8)
+
 /*
 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
 * and returns the corresponding pte. While this is not necessary for the

@@ -4770,11 +4772,24 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	pte_t *spte = NULL;
 	pte_t *pte;
 	spinlock_t *ptl;
+	static atomic_t timeout_cnt;
 

mm/hugetlb.c pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) · line 4780

Quoted 6 lines

mm/hugetlb.c +4 -2

-	if (!vma_shareable(vma, addr))
-		return (pte_t *)pmd_alloc(mm, pud, addr);
+	/*
+	 * Don't share if it is not sharable or locking attempt timed out
+	 * after 10ms. After 256 timeouts, PMD sharing will be permanently
+	 * disabled as it is just too slow.

It looks like this kind of policy interacts with kernel debug options like KASAN (which is going to slow the system down anyway) could introduce tricky issues due to different timings on a debug kernel.

Quoted 76 lines · Qian Cai

Quoted 76 lines · Waiman Long

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

These random delays, however, are deemed unacceptable. The cause of that delay is the long wait time in acquiring the mmap_sem when trying to share the huge PMDs.

To remove the unacceptable delays, we have to limit the amount of wait time on the mmap_sem. So the new down_write_timedlock() function is used to acquire the write lock on the mmap_sem with a timeout value of 10ms which should not cause a perceivable delay. If timeout happens, the task will abandon its effort to share the PMD and allocate its own copy instead.

When too many timeouts happens (threshold currently set at 256), the system may be too large for PMD sharing to be useful without undue delay. So the sharing will be disabled in this case.

Signed-off-by Waiman Long <[email protected]>

---
include/linux/fs.h | 7 +++++++
mm/hugetlb.c | 24 +++++++++++++++++++++---

2 files changed, 28 insertions(+), 3 deletions(-)

include/linux/fs.h +7 -0

index 997a530ff4e9..e9d3ad465a6b 100644
@@ -40,6 +40,7 @@
 #include <linux/fs_types.h>
 #include <linux/build_bug.h>
 #include <linux/stddef.h>
+#include <linux/ktime.h>
 
 #include <asm/byteorder.h>
 #include <uapi/linux/fs.h>

@@ -519,6 +520,12 @@ static inline void i_mmap_lock_write(struct address_space *mapping)
 	down_write(&mapping->i_mmap_rwsem);
 }
 
+static inline bool i_mmap_timedlock_write(struct address_space *mapping,
+					 ktime_t timeout)
+{
+	return down_write_timedlock(&mapping->i_mmap_rwsem, timeout);
+}
+
 static inline void i_mmap_unlock_write(struct address_space *mapping)
 {
 	up_write(&mapping->i_mmap_rwsem);

mm/hugetlb.c +3 -0

index 6d7296dd11b8..445af661ae29 100644
@@ -4750,6 +4750,8 @@ void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
 	}
 }
 
+#define PMD_SHARE_DISABLE_THRESHOLD	(1 << 8)
+
 /*
 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
 * and returns the corresponding pte. While this is not necessary for the

@@ -4770,11 +4772,24 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	pte_t *spte = NULL;
 	pte_t *pte;
 	spinlock_t *ptl;
+	static atomic_t timeout_cnt;
 

Quoted 8 lines

mm/hugetlb.c pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) · line 4780

Quoted 6 lines

mm/hugetlb.c +4 -2

-	if (!vma_shareable(vma, addr))
-		return (pte_t *)pmd_alloc(mm, pud, addr);
+	/*
+	 * Don't share if it is not sharable or locking attempt timed out
+	 * after 10ms. After 256 timeouts, PMD sharing will be permanently
+	 * disabled as it is just too slow.

It looks like this kind of policy interacts with kernel debug options like KASAN (which is going to slow the system down anyway) could introduce tricky issues due to different timings on a debug kernel.

With respect to lockdep, down_write_timedlock() works like a trylock. So a lot of checking will be skipped. Also the lockdep code won't be run until the lock is acquired. So its execution time has no effect on the timeout.

Quoted 87 lines · Waiman Long

Quoted 83 lines · Qian Cai

Quoted 71 lines · Waiman Long

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

These random delays, however, are deemed unacceptable. The cause of that delay is the long wait time in acquiring the mmap_sem when trying to share the huge PMDs.

To remove the unacceptable delays, we have to limit the amount of wait time on the mmap_sem. So the new down_write_timedlock() function is used to acquire the write lock on the mmap_sem with a timeout value of 10ms which should not cause a perceivable delay. If timeout happens, the task will abandon its effort to share the PMD and allocate its own copy instead.

When too many timeouts happens (threshold currently set at 256), the system may be too large for PMD sharing to be useful without undue delay. So the sharing will be disabled in this case.

Signed-off-by Waiman Long <[email protected]>

---
include/linux/fs.h | 7 +++++++
mm/hugetlb.c | 24 +++++++++++++++++++++---

2 files changed, 28 insertions(+), 3 deletions(-)

include/linux/fs.h +7 -0

index 997a530ff4e9..e9d3ad465a6b 100644
@@ -40,6 +40,7 @@
 #include <linux/fs_types.h>
 #include <linux/build_bug.h>
 #include <linux/stddef.h>
+#include <linux/ktime.h>
 
 #include <asm/byteorder.h>
 #include <uapi/linux/fs.h>

@@ -519,6 +520,12 @@ static inline void i_mmap_lock_write(struct address_space *mapping)
 	down_write(&mapping->i_mmap_rwsem);
 }
 
+static inline bool i_mmap_timedlock_write(struct address_space *mapping,
+					 ktime_t timeout)
+{
+	return down_write_timedlock(&mapping->i_mmap_rwsem, timeout);
+}
+
 static inline void i_mmap_unlock_write(struct address_space *mapping)
 {
 	up_write(&mapping->i_mmap_rwsem);

mm/hugetlb.c +2 -0

index 6d7296dd11b8..445af661ae29 100644
@@ -4750,6 +4750,8 @@ void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
 	}
 }
 
+#define PMD_SHARE_DISABLE_THRESHOLD	(1 << 8)
+
 /*
 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
 * and returns the corresponding pte. While this is not necessary for the

@@ -4770,11 +4772,24 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	pte_t *spte = NULL;

mm/hugetlb.c pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) · line 4780

Quoted 10 lines

mm/hugetlb.c +5 -2

 	pte_t *pte;
 	spinlock_t *ptl;
+	static atomic_t timeout_cnt;
 
-	if (!vma_shareable(vma, addr))
-		return (pte_t *)pmd_alloc(mm, pud, addr);
+	/*
+	 * Don't share if it is not sharable or locking attempt timed out
+	 * after 10ms. After 256 timeouts, PMD sharing will be permanently
+	 * disabled as it is just too slow.

It looks like this kind of policy interacts with kernel debug options like KASAN (which is going to slow the system down anyway) could introduce tricky issues due to different timings on a debug kernel.

Quoted 4 lines

With respect to lockdep, down_write_timedlock() works like a trylock. So a lot of checking will be skipped. Also the lockdep code won't be run until the lock is acquired. So its execution time has no effect on the timeout.

No only lockdep, but also things like KASAN, debug_pagealloc, page_poison, kmemleak, debug objects etc that all going to slow down things in huge_pmd_share(), and make it tricky to get a right timeout value for those debug kernels without changing the previous behavior.

Quoted 91 lines · Qian Cai

Quoted 91 lines · Waiman Long

Quoted 75 lines · Qian Cai

Quoted 75 lines · Waiman Long

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

These random delays, however, are deemed unacceptable. The cause of that delay is the long wait time in acquiring the mmap_sem when trying to share the huge PMDs.

To remove the unacceptable delays, we have to limit the amount of wait time on the mmap_sem. So the new down_write_timedlock() function is used to acquire the write lock on the mmap_sem with a timeout value of 10ms which should not cause a perceivable delay. If timeout happens, the task will abandon its effort to share the PMD and allocate its own copy instead.

When too many timeouts happens (threshold currently set at 256), the system may be too large for PMD sharing to be useful without undue delay. So the sharing will be disabled in this case.

Signed-off-by Waiman Long <[email protected]>

---
include/linux/fs.h | 7 +++++++
mm/hugetlb.c | 24 +++++++++++++++++++++---

2 files changed, 28 insertions(+), 3 deletions(-)

include/linux/fs.h +7 -0

index 997a530ff4e9..e9d3ad465a6b 100644
@@ -40,6 +40,7 @@
 #include <linux/fs_types.h>
 #include <linux/build_bug.h>
 #include <linux/stddef.h>
+#include <linux/ktime.h>
 
 #include <asm/byteorder.h>
 #include <uapi/linux/fs.h>

@@ -519,6 +520,12 @@ static inline void i_mmap_lock_write(struct address_space *mapping)
 	down_write(&mapping->i_mmap_rwsem);
 }
 
+static inline bool i_mmap_timedlock_write(struct address_space *mapping,
+					 ktime_t timeout)
+{
+	return down_write_timedlock(&mapping->i_mmap_rwsem, timeout);
+}
+
 static inline void i_mmap_unlock_write(struct address_space *mapping)
 {
 	up_write(&mapping->i_mmap_rwsem);

mm/hugetlb.c +3 -0

index 6d7296dd11b8..445af661ae29 100644
@@ -4750,6 +4750,8 @@ void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
 	}
 }
 
+#define PMD_SHARE_DISABLE_THRESHOLD	(1 << 8)
+
 /*
 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
 * and returns the corresponding pte. While this is not necessary for the

@@ -4770,11 +4772,24 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	pte_t *spte = NULL;
 	pte_t *pte;
 	spinlock_t *ptl;
+	static atomic_t timeout_cnt;
 

Quoted 8 lines

mm/hugetlb.c pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) · line 4780

Quoted 6 lines

mm/hugetlb.c +4 -2

-	if (!vma_shareable(vma, addr))
-		return (pte_t *)pmd_alloc(mm, pud, addr);
+	/*
+	 * Don't share if it is not sharable or locking attempt timed out
+	 * after 10ms. After 256 timeouts, PMD sharing will be permanently
+	 * disabled as it is just too slow.

It looks like this kind of policy interacts with kernel debug options like KASAN (which is going to slow the system down anyway) could introduce tricky issues due to different timings on a debug kernel.

With respect to lockdep, down_write_timedlock() works like a trylock. So a lot of checking will be skipped. Also the lockdep code won't be run until the lock is acquired. So its execution time has no effect on the timeout.

Quoted 3 lines

No only lockdep, but also things like KASAN, debug_pagealloc, page_poison, kmemleak, debug objects etc that all going to slow down things in huge_pmd_share(), and make it tricky to get a right timeout value for those debug kernels without changing the previous behavior.

Right, I understand that. I will move to use a sysctl parameters for the timeout and then set its default value to either 10ms or 20ms if some debug options are detected. Usually the slower than should not be more than 2X.

To remove the unacceptable delays, we have to limit the amount of wait time on the mmap_sem. So the new down_write_timedlock() function is used to acquire the write lock on the mmap_sem with a timeout value of 10ms which should not cause a perceivable delay. If timeout happens, the task will abandon its effort to share the PMD and allocate its own copy instead.

include/linux/fs.h +5 -0

+static inline bool i_mmap_timedlock_write(struct address_space *mapping,
+					 ktime_t timeout)
+{
+	return down_write_timedlock(&mapping->i_mmap_rwsem, timeout);
+}

Quoted 6 lines · Waiman Long

To remove the unacceptable delays, we have to limit the amount of wait time on the mmap_sem. So the new down_write_timedlock() function is used to acquire the write lock on the mmap_sem with a timeout value of 10ms which should not cause a perceivable delay. If timeout happens, the task will abandon its effort to share the PMD and allocate its own copy instead.

If you do a v2, this is *NOT* the mmap_sem. It's the i_mmap_rwsem which protects a very different data structure from the mmap_sem.

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

These random delays, however, are deemed unacceptable. The cause of that delay is the long wait time in acquiring the mmap_sem when trying to share the huge PMDs.

To remove the unacceptable delays, we have to limit the amount of wait time on the mmap_sem. So the new down_write_timedlock() function is used to acquire the write lock on the mmap_sem with a timeout value of 10ms which should not cause a perceivable delay. If timeout happens, the task will abandon its effort to share the PMD and allocate its own copy instead.

mm/hugetlb.c +10 -2

index 6d7296dd11b8..445af661ae29 100644
@@ -4750,6 +4750,8 @@ void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
 	}
 }
 
+#define PMD_SHARE_DISABLE_THRESHOLD	(1 << 8)
+
 /*
  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
  * and returns the corresponding pte. While this is not necessary for the

@@ -4770,11 +4772,24 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	pte_t *spte = NULL;
 	pte_t *pte;
 	spinlock_t *ptl;
+	static atomic_t timeout_cnt;
 
-	if (!vma_shareable(vma, addr))
-		return (pte_t *)pmd_alloc(mm, pud, addr);
+	/*
+	 * Don't share if it is not sharable or locking attempt timed out
+	 * after 10ms. After 256 timeouts, PMD sharing will be permanently
+	 * disabled as it is just too slow.
+	 */
+	if (!vma_shareable(vma, addr) ||
+	   (atomic_read(&timeout_cnt) >= PMD_SHARE_DISABLE_THRESHOLD))

All this got me wondering if we really need to take i_mmap_rwsem in write mode here. We are not changing the tree, only traversing it looking for a suitable vma.

Unless I am missing something, the hugetlb code only ever takes the semaphore in write mode; never read. Could this have been the result of changing the tree semaphore to read/write? Instead of analyzing all the code, the easiest and safest thing would have been to take all accesses in write mode.

Quoted 8 lines · Mike Kravetz

All this got me wondering if we really need to take i_mmap_rwsem in write mode here. We are not changing the tree, only traversing it looking for a suitable vma.

Unless I am missing something, the hugetlb code only ever takes the semaphore in write mode; never read. Could this have been the result of changing the tree semaphore to read/write? Instead of analyzing all the code, the easiest and safest thing would have been to take all accesses in write mode.

I was wondering the same thing. It was changed here:

commit 83cde9e8ba95d180eaefefe834958fbf7008cf39

Author: Davidlohr Bueso <[email protected]>
Date:   Fri Dec 12 16:54:21 2014 -0800

mm: use new helper functions around the i_mmap_mutex

Convert all open coded mutex_lock/unlock calls to the i_mmap_[lock/unlock]_write() helpers.

and a subsequent patch said:

This conversion is straightforward. For now, all users take the write lock.

There were subsequent patches which changed a few places c8475d144abb1e62958cc5ec281d2a9e161c1946 1acf2e040721564d579297646862b8ea3dd4511b d28eb9c861f41aa2af4cfcc5eeeddff42b13d31e 874bfcaf79e39135cd31e1cfc9265cf5222d1ec3 3dec0ba0be6a532cac949e02b853021bf6d57dad

but I don't know why this one wasn't changed.

(I was also wondering about caching a potentially sharable page table in the address_space to avoid having to walk the VMA tree at all if that one happened to be sharable).

I also think that the right solution is within the mm instead of adding a new api to rwsem and the extra complexity/overhead to osq _just_ for this
case. We've managed to not need timeout extensions in our locking primitives
thus far, which is a good thing imo.

Quoted 19 lines · Waiman Long

When allocating a large amount of static hugepages (~500-1500GB) on a system with large number of CPUs (4, 8 or even 16 sockets), performance degradation (random multi-second delays) was observed when thousands of processes are trying to fault in the data into the huge pages. The likelihood of the delay increases with the number of sockets and hence the CPUs a system has. This only happens in the initial setup phase and will be gone after all the necessary data are faulted in.

These random delays, however, are deemed unacceptable. The cause of that delay is the long wait time in acquiring the mmap_sem when trying to share the huge PMDs.

To remove the unacceptable delays, we have to limit the amount of wait time on the mmap_sem. So the new down_write_timedlock() function is used to acquire the write lock on the mmap_sem with a timeout value of 10ms which should not cause a perceivable delay. If timeout happens, the task will abandon its effort to share the PMD and allocate its own copy instead.

<snip>

Quoted 28 lines

mm/hugetlb.c +10 -2

index 6d7296dd11b8..445af661ae29 100644
@@ -4750,6 +4750,8 @@ void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
 	}
 }
 
+#define PMD_SHARE_DISABLE_THRESHOLD	(1 << 8)
+
 /*
  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
  * and returns the corresponding pte. While this is not necessary for the

@@ -4770,11 +4772,24 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	pte_t *spte = NULL;
 	pte_t *pte;
 	spinlock_t *ptl;
+	static atomic_t timeout_cnt;
 
-	if (!vma_shareable(vma, addr))
-		return (pte_t *)pmd_alloc(mm, pud, addr);
+	/*
+	 * Don't share if it is not sharable or locking attempt timed out
+	 * after 10ms. After 256 timeouts, PMD sharing will be permanently
+	 * disabled as it is just too slow.
+	 */
+	if (!vma_shareable(vma, addr) ||
+	   (atomic_read(&timeout_cnt) >= PMD_SHARE_DISABLE_THRESHOLD))

mm/hugetlb.c pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) · line 4777

Quoted 10 lines

mm/hugetlb.c +8 -1

+		goto out_no_share;
+
+	if (!i_mmap_timedlock_write(mapping, ms_to_ktime(10))) {
+		if (atomic_inc_return(&timeout_cnt) ==
+		    PMD_SHARE_DISABLE_THRESHOLD)
+			pr_info("Hugetlbfs PMD sharing disabled because of timeouts!\n");
+		goto out_no_share;
+	}
 
-	i_mmap_lock_write(mapping);

All this got me wondering if we really need to take i_mmap_rwsem in write mode here. We are not changing the tree, only traversing it looking for a suitable vma.

Unless I am missing something, the hugetlb code only ever takes the semaphore in write mode; never read. Could this have been the result of changing the tree semaphore to read/write? Instead of analyzing all the code, the easiest and safest thing would have been to take all accesses in write mode.

I can investigate more, but wanted to ask the question in case someone already knows.

At one time, I thought it was safe to acquire the semaphore in read mode for huge_pmd_share, but write mode for huge_pmd_unshare. See commit b43a99900559. This was reverted along with another patch for other reasons.

If we change change from write to read mode, this may have significant impact on the stalls.

If we can take the rwsem in read mode, that should solve the problem AFAICS. As I don't have a full understanding of the history of that code, I didn't try to do that in my patch.

Quoted 3 lines · Waiman Long

If we can take the rwsem in read mode, that should solve the problem AFAICS. As I don't have a full understanding of the history of that code, I didn't try to do that in my patch.

Do you still have access to an environment that creates the long stalls? If so, can you try the simple change of taking the semaphore in read mode in huge_pmd_share.

(quoted diff) +8 -2

+#define PMD_SHARE_DISABLE_THRESHOLD	(1 << 8)
+
 /*
  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
  * and returns the corresponding pte. While this is not necessary for the

@@ -4770,11 +4772,24 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	pte_t *spte = NULL;
 	pte_t *pte;
 	spinlock_t *ptl;
+	static atomic_t timeout_cnt;
 
-	if (!vma_shareable(vma, addr))
-		return (pte_t *)pmd_alloc(mm, pud, addr);
+	/*
+	 * Don't share if it is not sharable or locking attempt timed out
+	 * after 10ms. After 256 timeouts, PMD sharing will be permanently
+	 * disabled as it is just too slow.
+	 */

(quoted diff) +3 -1

-	i_mmap_lock_write(mapping);
 	vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
 		if (svma == vma)
 			continue;

@@ -4806,6 +4821,9 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	pte = (pte_t *)pmd_alloc(mm, pud, addr);
 	i_mmap_unlock_write(mapping);
 	return pte;
+
+out_no_share:
+	return (pte_t *)pmd_alloc(mm, pud, addr);
 }

A customer with large SMP systems (up to 16 sockets) with application that uses large amount of static hugepages (~500-1500GB) are experiencing random multisecond delays. These delays was caused by the long time it took to scan the VMA interval tree with mmap_sem held.

To fix this problem while perserving existing behavior as much as possible, we need to allow timeout in down_write() and disabling PMD sharing when it is taking too long to do so. Since a transaction can involving touching multiple huge pages, timing out for each of the huge page interactions does not completely solve the problem. So a threshold is set to completely disable PMD sharing if too many timeouts happen.

The first 4 patches of this 5-patch series adds a new down_write_timedlock() API which accepts a timeout argument and return true is locking is successful or false otherwise. It works more or less than a down_write_trylock() but the calling thread may sleep.

Quoted 12 lines · Waiman Long

A customer with large SMP systems (up to 16 sockets) with application that uses large amount of static hugepages (~500-1500GB) are experiencing random multisecond delays. These delays was caused by the long time it took to scan the VMA interval tree with mmap_sem held.

To fix this problem while perserving existing behavior as much as possible, we need to allow timeout in down_write() and disabling PMD sharing when it is taking too long to do so. Since a transaction can involving touching multiple huge pages, timing out for each of the huge page interactions does not completely solve the problem. So a threshold is set to completely disable PMD sharing if too many timeouts happen.

Quoted 4 lines

The first 4 patches of this 5-patch series adds a new down_write_timedlock() API which accepts a timeout argument and return true is locking is successful or false otherwise. It works more or less than a down_write_trylock() but the calling thread may sleep.

Just on general principle, this is a non-starter. If a lock is being held too long, then whatever the lock is protecting needs fixing. Adding timeouts to locks and sysctls to tune them is not a viable solution to address latencies caused by algorithm scalability issues.