Skip to contentSkip to conversation

x86/sgx: Fix deadlock and race conditions between fork() and EPC reclaim

PATCH Superseded Jarkko Sakkinen 2 participants 13 messages Apr 3, 2020 – Apr 10, 2020

Subsystems (2)

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 x86/sgx: Fix deadlock and race conditions between fork() and EPC reclaim

From: Sean Christopherson <[email protected]>

Drop the synchronize_srcu() from sgx_encl_mm_add() and replace it with a mm_list versioning concept to avoid deadlock when adding a mm during dup_mmap()/fork(), and to ensure copied PTEs are zapped.

When dup_mmap() runs, it holds mmap_sem for write in both the old mm and new mm. Invoking synchronize_srcu() while holding mmap_sem of a mm that is already attached to the enclave will deadlock if the reclaimer is in the process of walking mm_list, as the reclaimer will try to acquire mmap_sem (of the old mm) while holding encl->srcu for read.

 INFO: task ksgxswapd:181 blocked for more than 120 seconds.
 ksgxswapd       D    0   181      2 0x80004000
 Call Trace:
  __schedule+0x2db/0x700
  schedule+0x44/0xb0
  rwsem_down_read_slowpath+0x370/0x470
  down_read+0x95/0xa0
  sgx_reclaim_pages+0x1d2/0x7d0
  ksgxswapd+0x151/0x2e0
  kthread+0x120/0x140
  ret_from_fork+0x35/0x40

 INFO: task fork_consistenc:18824 blocked for more than 120 seconds.
 fork_consistenc D    0 18824  18786 0x00004320
 Call Trace:
  __schedule+0x2db/0x700
  schedule+0x44/0xb0
  schedule_timeout+0x205/0x300
  wait_for_completion+0xb7/0x140
  __synchronize_srcu.part.22+0x81/0xb0
  synchronize_srcu_expedited+0x27/0x30
  synchronize_srcu+0x57/0xe0
  sgx_encl_mm_add+0x12b/0x160
  sgx_vma_open+0x22/0x40
  dup_mm+0x521/0x580
  copy_process+0x1a56/0x1b50
  _do_fork+0x85/0x3a0
  __x64_sys_clone+0x8e/0xb0
  do_syscall_64+0x57/0x1b0
  entry_SYSCALL_64_after_hwframe+0x44/0xa9

Furthermore, doing synchronize_srcu() in sgx_encl_mm_add() does not prevent the new mm from having stale PTEs pointing at the EPC page to be reclaimed. dup_mmap() calls vm_ops->open()/sgx_encl_mm_add() _after_ PTEs are copied to the new mm, i.e. blocking fork() until reclaim zaps the old mm is pointless as the stale PTEs have already been created in the new mm.

All other flows that walk mm_list can safely race with dup_mmap() or are protected by a different mechanism. Add comments to all srcu readers that don't check the list version to document why its ok for the flow to ignore the version.

Note, synchronize_srcu() is still needed when removing a mm from an enclave, as the srcu readers must complete their walk before the mm can be freed. Removing a mm is never done while holding mmap_sem.

---
v2:
* Remove smp_wmb() as x86 does not reorder writes in the pipeline.
* Refine comments to be more to the point and more maintainable when things might change.
* Replace the ad hoc (goto-based) loop construct with a proper loop construct.

 arch/x86/kernel/cpu/sgx/encl.c    | 11 +++++++--
 arch/x86/kernel/cpu/sgx/encl.h    |  1 +
 arch/x86/kernel/cpu/sgx/reclaim.c | 39 +++++++++++++++++++++----------
 3 files changed, 37 insertions(+), 14 deletions(-)
arch/x86/kernel/cpu/sgx/encl.c +9 -2
index e0124a2f22d5..5b15352b3d4f 100644
@@ -196,6 +196,9 @@ int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm)
 	struct sgx_encl_mm *encl_mm;
 	int ret;
 
+	/* mm_list can be accessed only by a single thread at a time. */
+	lockdep_assert_held_write(&mm->mmap_sem);
+
 	if (atomic_read(&encl->flags) & SGX_ENCL_DEAD)
 		return -EINVAL;
 
@@ -221,12 +224,16 @@ int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm)
 		return ret;
 	}
 
+	/*
+	 * The page reclaimer uses list version for synchronization instead of
+	 * synchronize_scru() because otherwise we could conflict with
+	 * dup_mmap().
+	 */
 	spin_lock(&encl->mm_lock);
 	list_add_rcu(&encl_mm->list, &encl->mm_list);
+	encl->mm_list_version++;
 	spin_unlock(&encl->mm_lock);
 
-	synchronize_srcu(&encl->srcu);
-
 	return 0;
 }
 
arch/x86/kernel/cpu/sgx/encl.h +1 -0
index 44b353aa8866..f0f72e591244 100644
@@ -74,6 +74,7 @@ struct sgx_encl {
 	struct mutex lock;
 	struct list_head mm_list;
 	spinlock_t mm_lock;
+	unsigned long mm_list_version;
 	struct file *backing;
 	struct kref refcount;
 	struct srcu_struct srcu;
arch/x86/kernel/cpu/sgx/reclaim.c +27 -12
index 39f0ddefbb79..5fb8bdfa6a1f 100644
@@ -184,28 +184,38 @@ static void sgx_reclaimer_block(struct sgx_epc_page *epc_page)
 	struct sgx_encl_page *page = epc_page->owner;
 	unsigned long addr = SGX_ENCL_PAGE_ADDR(page);
 	struct sgx_encl *encl = page->encl;
+	unsigned long mm_list_version;
 	struct sgx_encl_mm *encl_mm;
 	struct vm_area_struct *vma;
 	int idx, ret;
 
-	idx = srcu_read_lock(&encl->srcu);
+	do {
+		mm_list_version = encl->mm_list_version;
 
-	list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
-		if (!mmget_not_zero(encl_mm->mm))
-			continue;
+		/* Fence reads as the CPU can reorder them. This guarantees
+		 * that we don't access old list with a new version.
+		 */
+		smp_rmb();
 
-		down_read(&encl_mm->mm->mmap_sem);
+		idx = srcu_read_lock(&encl->srcu);
 
-		ret = sgx_encl_find(encl_mm->mm, addr, &vma);
-		if (!ret && encl == vma->vm_private_data)
-			zap_vma_ptes(vma, addr, PAGE_SIZE);
+		list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
+			if (!mmget_not_zero(encl_mm->mm))
+				continue;
 
-		up_read(&encl_mm->mm->mmap_sem);
+			down_read(&encl_mm->mm->mmap_sem);
 
-		mmput_async(encl_mm->mm);
-	}
+			ret = sgx_encl_find(encl_mm->mm, addr, &vma);
+			if (!ret && encl == vma->vm_private_data)
+				zap_vma_ptes(vma, addr, PAGE_SIZE);
 
-	srcu_read_unlock(&encl->srcu, idx);
+			up_read(&encl_mm->mm->mmap_sem);
+
+			mmput_async(encl_mm->mm);
+		}
+
+		srcu_read_unlock(&encl->srcu, idx);
+	} while (unlikely(encl->mm_list_version != mm_list_version));
 
 	mutex_lock(&encl->lock);
 
@@ -250,6 +260,11 @@ static const cpumask_t *sgx_encl_ewb_cpumask(struct sgx_encl *encl)
 	struct sgx_encl_mm *encl_mm;
 	int idx;
 
+	/*
+	 * Can race with sgx_encl_mm_add(), but ETRACK has already been
+	 * executed, which means that the CPUs running in the new mm will enter
+	 * into the enclave with a fresh epoch.
+	 */
 	cpumask_clear(cpumask);
 
 	idx = srcu_read_lock(&encl->srcu);
Signature · 1 line
2.25.1
#1 lore raw
Quoted 162 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

Drop the synchronize_srcu() from sgx_encl_mm_add() and replace it with a mm_list versioning concept to avoid deadlock when adding a mm during dup_mmap()/fork(), and to ensure copied PTEs are zapped.

When dup_mmap() runs, it holds mmap_sem for write in both the old mm and new mm. Invoking synchronize_srcu() while holding mmap_sem of a mm that is already attached to the enclave will deadlock if the reclaimer is in the process of walking mm_list, as the reclaimer will try to acquire mmap_sem (of the old mm) while holding encl->srcu for read.

 INFO: task ksgxswapd:181 blocked for more than 120 seconds.
 ksgxswapd       D    0   181      2 0x80004000
 Call Trace:
  __schedule+0x2db/0x700
  schedule+0x44/0xb0
  rwsem_down_read_slowpath+0x370/0x470
  down_read+0x95/0xa0
  sgx_reclaim_pages+0x1d2/0x7d0
  ksgxswapd+0x151/0x2e0
  kthread+0x120/0x140
  ret_from_fork+0x35/0x40

 INFO: task fork_consistenc:18824 blocked for more than 120 seconds.
 fork_consistenc D    0 18824  18786 0x00004320
 Call Trace:
  __schedule+0x2db/0x700
  schedule+0x44/0xb0
  schedule_timeout+0x205/0x300
  wait_for_completion+0xb7/0x140
  __synchronize_srcu.part.22+0x81/0xb0
  synchronize_srcu_expedited+0x27/0x30
  synchronize_srcu+0x57/0xe0
  sgx_encl_mm_add+0x12b/0x160
  sgx_vma_open+0x22/0x40
  dup_mm+0x521/0x580
  copy_process+0x1a56/0x1b50
  _do_fork+0x85/0x3a0
  __x64_sys_clone+0x8e/0xb0
  do_syscall_64+0x57/0x1b0
  entry_SYSCALL_64_after_hwframe+0x44/0xa9

Furthermore, doing synchronize_srcu() in sgx_encl_mm_add() does not prevent the new mm from having stale PTEs pointing at the EPC page to be reclaimed. dup_mmap() calls vm_ops->open()/sgx_encl_mm_add() _after_ PTEs are copied to the new mm, i.e. blocking fork() until reclaim zaps the old mm is pointless as the stale PTEs have already been created in the new mm.

All other flows that walk mm_list can safely race with dup_mmap() or are protected by a different mechanism. Add comments to all srcu readers that don't check the list version to document why its ok for the flow to ignore the version.

Note, synchronize_srcu() is still needed when removing a mm from an enclave, as the srcu readers must complete their walk before the mm can be freed. Removing a mm is never done while holding mmap_sem.

---
v2:
* Remove smp_wmb() as x86 does not reorder writes in the pipeline.
* Refine comments to be more to the point and more maintainable when things might change.
* Replace the ad hoc (goto-based) loop construct with a proper loop construct.

 arch/x86/kernel/cpu/sgx/encl.c    | 11 +++++++--
 arch/x86/kernel/cpu/sgx/encl.h    |  1 +
 arch/x86/kernel/cpu/sgx/reclaim.c | 39 +++++++++++++++++++++----------
 3 files changed, 37 insertions(+), 14 deletions(-)
arch/x86/kernel/cpu/sgx/encl.c +9 -2
index e0124a2f22d5..5b15352b3d4f 100644
@@ -196,6 +196,9 @@ int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm)
 	struct sgx_encl_mm *encl_mm;
 	int ret;
 
+	/* mm_list can be accessed only by a single thread at a time. */
+	lockdep_assert_held_write(&mm->mmap_sem);
+
 	if (atomic_read(&encl->flags) & SGX_ENCL_DEAD)
 		return -EINVAL;
 
@@ -221,12 +224,16 @@ int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm)
 		return ret;
 	}
 
+	/*
+	 * The page reclaimer uses list version for synchronization instead of
+	 * synchronize_scru() because otherwise we could conflict with
+	 * dup_mmap().
+	 */
 	spin_lock(&encl->mm_lock);
 	list_add_rcu(&encl_mm->list, &encl->mm_list);
+	encl->mm_list_version++;
 	spin_unlock(&encl->mm_lock);
 
-	synchronize_srcu(&encl->srcu);
-
 	return 0;
 }
 
arch/x86/kernel/cpu/sgx/encl.h +1 -0
index 44b353aa8866..f0f72e591244 100644
@@ -74,6 +74,7 @@ struct sgx_encl {
 	struct mutex lock;
 	struct list_head mm_list;
 	spinlock_t mm_lock;
+	unsigned long mm_list_version;
 	struct file *backing;
 	struct kref refcount;
 	struct srcu_struct srcu;
arch/x86/kernel/cpu/sgx/reclaim.c +14 -11
index 39f0ddefbb79..5fb8bdfa6a1f 100644
@@ -184,28 +184,38 @@ static void sgx_reclaimer_block(struct sgx_epc_page *epc_page)
 	struct sgx_encl_page *page = epc_page->owner;
 	unsigned long addr = SGX_ENCL_PAGE_ADDR(page);
 	struct sgx_encl *encl = page->encl;
+	unsigned long mm_list_version;
 	struct sgx_encl_mm *encl_mm;
 	struct vm_area_struct *vma;
 	int idx, ret;
 
-	idx = srcu_read_lock(&encl->srcu);
+	do {
+		mm_list_version = encl->mm_list_version;
 
-	list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
-		if (!mmget_not_zero(encl_mm->mm))
-			continue;
+		/* Fence reads as the CPU can reorder them. This guarantees
+		 * that we don't access old list with a new version.
+		 */
+		smp_rmb();
 
-		down_read(&encl_mm->mm->mmap_sem);
+		idx = srcu_read_lock(&encl->srcu);
 
-		ret = sgx_encl_find(encl_mm->mm, addr, &vma);
-		if (!ret && encl == vma->vm_private_data)
-			zap_vma_ptes(vma, addr, PAGE_SIZE);
+		list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
+			if (!mmget_not_zero(encl_mm->mm))
+				continue;
 
-		up_read(&encl_mm->mm->mmap_sem);
+			down_read(&encl_mm->mm->mmap_sem);
 
-		mmput_async(encl_mm->mm);
-	}
+			ret = sgx_encl_find(encl_mm->mm, addr, &vma);
+			if (!ret && encl == vma->vm_private_data)

arch/x86/kernel/cpu/sgx/reclaim.c static void sgx_reclaimer_block(struct sgx_epc_page *epc_page) · line 218

Quoted 10 lines
arch/x86/kernel/cpu/sgx/reclaim.c +8 -1
+				zap_vma_ptes(vma, addr, PAGE_SIZE);
 
-	srcu_read_unlock(&encl->srcu, idx);
+			up_read(&encl_mm->mm->mmap_sem);
+
+			mmput_async(encl_mm->mm);
+		}
+
+		srcu_read_unlock(&encl->srcu, idx);
+	} while (unlikely(encl->mm_list_version != mm_list_version));

This is bad, or at least makes the code unnecessarily hard to understand given the use of the fences: encl->mm_list_version is read twice per iteration.

I'll change the code to use "for ( ; ; )" and have exactly one read per iteration.

/Jarko

#2 lore raw
Quoted 46 lines · Jarkko Sakkinen
Quoted 30 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

arch/x86/kernel/cpu/sgx/reclaim.c +13 -11
-	idx = srcu_read_lock(&encl->srcu);
+	do {
+		mm_list_version = encl->mm_list_version;
 
-	list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
-		if (!mmget_not_zero(encl_mm->mm))
-			continue;
+		/* Fence reads as the CPU can reorder them. This guarantees
+		 * that we don't access old list with a new version.
+		 */
+		smp_rmb();
 
-		down_read(&encl_mm->mm->mmap_sem);
+		idx = srcu_read_lock(&encl->srcu);
 
-		ret = sgx_encl_find(encl_mm->mm, addr, &vma);
-		if (!ret && encl == vma->vm_private_data)
-			zap_vma_ptes(vma, addr, PAGE_SIZE);
+		list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
+			if (!mmget_not_zero(encl_mm->mm))
+				continue;
 
-		up_read(&encl_mm->mm->mmap_sem);
+			down_read(&encl_mm->mm->mmap_sem);
 
-		mmput_async(encl_mm->mm);
-	}
+			ret = sgx_encl_find(encl_mm->mm, addr, &vma);
+			if (!ret && encl == vma->vm_private_data)

arch/x86/kernel/cpu/sgx/reclaim.c static void sgx_reclaimer_block(struct sgx_epc_page *epc_page) · line 218

Quoted 10 lines
arch/x86/kernel/cpu/sgx/reclaim.c +8 -1
+				zap_vma_ptes(vma, addr, PAGE_SIZE);
 
-	srcu_read_unlock(&encl->srcu, idx);
+			up_read(&encl_mm->mm->mmap_sem);
+
+			mmput_async(encl_mm->mm);
+		}
+
+		srcu_read_unlock(&encl->srcu, idx);
+	} while (unlikely(encl->mm_list_version != mm_list_version));

This is bad, or at least makes the code unnecessarily hard to understand given the use of the fences: encl->mm_list_version is read twice per iteration.

Quoted 2 lines

I'll change the code to use "for ( ; ; )" and have exactly one read per iteration.

It has to be read twice per iteration, the whole point is to see if the version at the start of the iteration matches the version at the end.

#3 lore raw
Quoted 47 lines · Sean Christopherson
Quoted 46 lines · Jarkko Sakkinen
Quoted 30 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

arch/x86/kernel/cpu/sgx/reclaim.c +13 -11
-	idx = srcu_read_lock(&encl->srcu);
+	do {
+		mm_list_version = encl->mm_list_version;
 
-	list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
-		if (!mmget_not_zero(encl_mm->mm))
-			continue;
+		/* Fence reads as the CPU can reorder them. This guarantees
+		 * that we don't access old list with a new version.
+		 */
+		smp_rmb();
 
-		down_read(&encl_mm->mm->mmap_sem);
+		idx = srcu_read_lock(&encl->srcu);
 
-		ret = sgx_encl_find(encl_mm->mm, addr, &vma);
-		if (!ret && encl == vma->vm_private_data)
-			zap_vma_ptes(vma, addr, PAGE_SIZE);
+		list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
+			if (!mmget_not_zero(encl_mm->mm))
+				continue;
 
-		up_read(&encl_mm->mm->mmap_sem);
+			down_read(&encl_mm->mm->mmap_sem);
 
-		mmput_async(encl_mm->mm);
-	}
+			ret = sgx_encl_find(encl_mm->mm, addr, &vma);
+			if (!ret && encl == vma->vm_private_data)

arch/x86/kernel/cpu/sgx/reclaim.c static void sgx_reclaimer_block(struct sgx_epc_page *epc_page) · line 218

Quoted 10 lines
arch/x86/kernel/cpu/sgx/reclaim.c +8 -1
+				zap_vma_ptes(vma, addr, PAGE_SIZE);
 
-	srcu_read_unlock(&encl->srcu, idx);
+			up_read(&encl_mm->mm->mmap_sem);
+
+			mmput_async(encl_mm->mm);
+		}
+
+		srcu_read_unlock(&encl->srcu, idx);
+	} while (unlikely(encl->mm_list_version != mm_list_version));

This is bad, or at least makes the code unnecessarily hard to understand given the use of the fences: encl->mm_list_version is read twice per iteration.

Quoted 5 lines
Quoted 2 lines

I'll change the code to use "for ( ; ; )" and have exactly one read per iteration.

It has to be read twice per iteration, the whole point is to see if the version at the start of the iteration matches the version at the end.

Nope if you use zero version as invalid version.

/Jarkko

#4 lore raw
Quoted 48 lines · Jarkko Sakkinen
Quoted 47 lines · Sean Christopherson
Quoted 46 lines · Jarkko Sakkinen
Quoted 30 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

arch/x86/kernel/cpu/sgx/reclaim.c +13 -11
-	idx = srcu_read_lock(&encl->srcu);
+	do {
+		mm_list_version = encl->mm_list_version;
 
-	list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
-		if (!mmget_not_zero(encl_mm->mm))
-			continue;
+		/* Fence reads as the CPU can reorder them. This guarantees
+		 * that we don't access old list with a new version.
+		 */
+		smp_rmb();
 
-		down_read(&encl_mm->mm->mmap_sem);
+		idx = srcu_read_lock(&encl->srcu);
 
-		ret = sgx_encl_find(encl_mm->mm, addr, &vma);
-		if (!ret && encl == vma->vm_private_data)
-			zap_vma_ptes(vma, addr, PAGE_SIZE);
+		list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
+			if (!mmget_not_zero(encl_mm->mm))
+				continue;
 
-		up_read(&encl_mm->mm->mmap_sem);
+			down_read(&encl_mm->mm->mmap_sem);
 
-		mmput_async(encl_mm->mm);
-	}
+			ret = sgx_encl_find(encl_mm->mm, addr, &vma);
+			if (!ret && encl == vma->vm_private_data)

arch/x86/kernel/cpu/sgx/reclaim.c static void sgx_reclaimer_block(struct sgx_epc_page *epc_page) · line 218

Quoted 10 lines
arch/x86/kernel/cpu/sgx/reclaim.c +8 -1
+				zap_vma_ptes(vma, addr, PAGE_SIZE);
 
-	srcu_read_unlock(&encl->srcu, idx);
+			up_read(&encl_mm->mm->mmap_sem);
+
+			mmput_async(encl_mm->mm);
+		}
+
+		srcu_read_unlock(&encl->srcu, idx);
+	} while (unlikely(encl->mm_list_version != mm_list_version));

This is bad, or at least makes the code unnecessarily hard to understand given the use of the fences: encl->mm_list_version is read twice per iteration.

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines

I'll change the code to use "for ( ; ; )" and have exactly one read per iteration.

It has to be read twice per iteration, the whole point is to see if the version at the start of the iteration matches the version at the end.

Nope if you use zero version as invalid version.

Which would break if more than one mm is added to an enclave.

#5 lore raw
Quoted 52 lines · Sean Christopherson
Quoted 51 lines · Jarkko Sakkinen
Quoted 50 lines · Sean Christopherson
Quoted 48 lines · Jarkko Sakkinen
Quoted 30 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

arch/x86/kernel/cpu/sgx/reclaim.c +13 -11
-	idx = srcu_read_lock(&encl->srcu);
+	do {
+		mm_list_version = encl->mm_list_version;
 
-	list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
-		if (!mmget_not_zero(encl_mm->mm))
-			continue;
+		/* Fence reads as the CPU can reorder them. This guarantees
+		 * that we don't access old list with a new version.
+		 */
+		smp_rmb();
 
-		down_read(&encl_mm->mm->mmap_sem);
+		idx = srcu_read_lock(&encl->srcu);
 
-		ret = sgx_encl_find(encl_mm->mm, addr, &vma);
-		if (!ret && encl == vma->vm_private_data)
-			zap_vma_ptes(vma, addr, PAGE_SIZE);
+		list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
+			if (!mmget_not_zero(encl_mm->mm))
+				continue;
 
-		up_read(&encl_mm->mm->mmap_sem);
+			down_read(&encl_mm->mm->mmap_sem);
 
-		mmput_async(encl_mm->mm);
-	}
+			ret = sgx_encl_find(encl_mm->mm, addr, &vma);
+			if (!ret && encl == vma->vm_private_data)

arch/x86/kernel/cpu/sgx/reclaim.c static void sgx_reclaimer_block(struct sgx_epc_page *epc_page) · line 218

Quoted 10 lines
arch/x86/kernel/cpu/sgx/reclaim.c +8 -1
+				zap_vma_ptes(vma, addr, PAGE_SIZE);
 
-	srcu_read_unlock(&encl->srcu, idx);
+			up_read(&encl_mm->mm->mmap_sem);
+
+			mmput_async(encl_mm->mm);
+		}
+
+		srcu_read_unlock(&encl->srcu, idx);
+	} while (unlikely(encl->mm_list_version != mm_list_version));

This is bad, or at least makes the code unnecessarily hard to understand given the use of the fences: encl->mm_list_version is read twice per iteration.

I'll change the code to use "for ( ; ; )" and have exactly one read per iteration.

Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

It has to be read twice per iteration, the whole point is to see if the version at the start of the iteration matches the version at the end.

Nope if you use zero version as invalid version.

Which would break if more than one mm is added to an enclave.

Ah, you mean within the function (just saw v3).

#6 lore raw
In reply to Jarkko Sakkinen

arch/x86/kernel/cpu/sgx/encl.c int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm) · line 233

Quoted 12 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

arch/x86/kernel/cpu/sgx/encl.c +5 -0
@@ -221,12 +224,16 @@ int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm)
 		return ret;
 	}
 
+	/*
+	 * The page reclaimer uses list version for synchronization instead of
+	 * synchronize_scru() because otherwise we could conflict with
+	 * dup_mmap().
+	 */
 	spin_lock(&encl->mm_lock);
 	list_add_rcu(&encl_mm->list, &encl->mm_list);

You dropped the smp_wmb().

Quoted 32 lines
arch/x86/kernel/cpu/sgx/encl.c +1 -2
+	encl->mm_list_version++;
 	spin_unlock(&encl->mm_lock);
 
-	synchronize_srcu(&encl->srcu);
-
 	return 0;
 }
 
arch/x86/kernel/cpu/sgx/encl.h +1 -0
index 44b353aa8866..f0f72e591244 100644
@@ -74,6 +74,7 @@ struct sgx_encl {
 	struct mutex lock;
 	struct list_head mm_list;
 	spinlock_t mm_lock;
+	unsigned long mm_list_version;
 	struct file *backing;
 	struct kref refcount;
 	struct srcu_struct srcu;
arch/x86/kernel/cpu/sgx/reclaim.c +1 -0
index 39f0ddefbb79..5fb8bdfa6a1f 100644
@@ -184,28 +184,38 @@ static void sgx_reclaimer_block(struct sgx_epc_page *epc_page)
 	struct sgx_encl_page *page = epc_page->owner;
 	unsigned long addr = SGX_ENCL_PAGE_ADDR(page);
 	struct sgx_encl *encl = page->encl;
+	unsigned long mm_list_version;
 	struct sgx_encl_mm *encl_mm;
 	struct vm_area_struct *vma;
 	int idx, ret;

arch/x86/kernel/cpu/sgx/reclaim.c static void sgx_reclaimer_block(struct sgx_epc_page *epc_page) · line 196

Quoted 10 lines
arch/x86/kernel/cpu/sgx/reclaim.c +4 -4
 
-	idx = srcu_read_lock(&encl->srcu);
+	do {
+		mm_list_version = encl->mm_list_version;
 
-	list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
-		if (!mmget_not_zero(encl_mm->mm))
-			continue;
+		/* Fence reads as the CPU can reorder them. This guarantees
+		 * that we don't access old list with a new version.

This comment is flat out wrong. This has nothing to do the CPU reordering things. The smp_{r,w}mb() are nothing more than compiler barriers, and even those go away when the kernel is built with SMP=0.

I don't mind gutting the other comments, but there is a well established pattern form smb_wmb()/smp_rmb() pairs, I would strongly prefer to keep the exact comment I submitted.

Quoted 50 lines
arch/x86/kernel/cpu/sgx/reclaim.c +17 -8
+		 */
+		smp_rmb();
 
 
 
-		down_read(&encl_mm->mm->mmap_sem);
+		idx = srcu_read_lock(&encl->srcu);
 
-		ret = sgx_encl_find(encl_mm->mm, addr, &vma);
-		if (!ret && encl == vma->vm_private_data)
-			zap_vma_ptes(vma, addr, PAGE_SIZE);
+		list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
+			if (!mmget_not_zero(encl_mm->mm))
+				continue;
 
-		up_read(&encl_mm->mm->mmap_sem);
+			down_read(&encl_mm->mm->mmap_sem);
 
-		mmput_async(encl_mm->mm);
-	}
+			ret = sgx_encl_find(encl_mm->mm, addr, &vma);
+			if (!ret && encl == vma->vm_private_data)
+				zap_vma_ptes(vma, addr, PAGE_SIZE);
 
-	srcu_read_unlock(&encl->srcu, idx);
+			up_read(&encl_mm->mm->mmap_sem);
+
+			mmput_async(encl_mm->mm);
+		}
+
+		srcu_read_unlock(&encl->srcu, idx);
+	} while (unlikely(encl->mm_list_version != mm_list_version));
 
arch/x86/kernel/cpu/sgx/reclaim.c +5 -0
 	mutex_lock(&encl->lock);
@@ -250,6 +260,11 @@ static const cpumask_t *sgx_encl_ewb_cpumask(struct sgx_encl *encl)
 	struct sgx_encl_mm *encl_mm;
 	int idx;
 
+	/*
+	 * Can race with sgx_encl_mm_add(), but ETRACK has already been
+	 * executed, which means that the CPUs running in the new mm will enter
+	 * into the enclave with a fresh epoch.
+	 */
 	cpumask_clear(cpumask);
 
 	idx = srcu_read_lock(&encl->srcu);
Signature · 1 line
2.25.1
#7 lore raw
Quoted 6 lines · Sean Christopherson
Quoted 5 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

arch/x86/kernel/cpu/sgx/encl.c +0 -0
@@ -221,12 +224,16 @@ int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm)
 		return ret;
 	}
 
Quoted 9 lines

arch/x86/kernel/cpu/sgx/encl.c int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm) · line 233

Quoted 7 lines
arch/x86/kernel/cpu/sgx/encl.c +5 -0
+	/*
+	 * The page reclaimer uses list version for synchronization instead of
+	 * synchronize_scru() because otherwise we could conflict with
+	 * dup_mmap().
+	 */
 	spin_lock(&encl->mm_lock);
 	list_add_rcu(&encl_mm->list, &encl->mm_list);

You dropped the smp_wmb().

As I said to you in my review x86 pipeline does not reorder writes.

/Jarkko

#8 lore raw
Quoted 17 lines · Jarkko Sakkinen
Quoted 6 lines · Sean Christopherson
Quoted 5 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

arch/x86/kernel/cpu/sgx/encl.c +0 -0
@@ -221,12 +224,16 @@ int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm)
 		return ret;
 	}
 
Quoted 9 lines

arch/x86/kernel/cpu/sgx/encl.c int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm) · line 233

Quoted 7 lines
arch/x86/kernel/cpu/sgx/encl.c +5 -0
+	/*
+	 * The page reclaimer uses list version for synchronization instead of
+	 * synchronize_scru() because otherwise we could conflict with
+	 * dup_mmap().
+	 */
 	spin_lock(&encl->mm_lock);
 	list_add_rcu(&encl_mm->list, &encl->mm_list);

You dropped the smp_wmb().

Quoted 1 line

As I said to you in my review x86 pipeline does not reorder writes.

And as I pointed out in this thread, smp_wmb() is a _compiler_ barrier if and only if CONFIG_SMP=y. The compiler can reorder list_add_rcu() and mm_list_version++ because from it's perspective there is no dependency between the two. And that's entirely true except for the SMP case where the consumer of mm_list_version is relying on the list to be updated before the version changes.

#9 lore raw
Quoted 18 lines · Sean Christopherson
Quoted 17 lines · Jarkko Sakkinen
Quoted 15 lines · Sean Christopherson

arch/x86/kernel/cpu/sgx/encl.c int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm) · line 233

Quoted 12 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

arch/x86/kernel/cpu/sgx/encl.c +5 -0
@@ -221,12 +224,16 @@ int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm)
 		return ret;
 	}
 
+	/*
+	 * The page reclaimer uses list version for synchronization instead of
+	 * synchronize_scru() because otherwise we could conflict with
+	 * dup_mmap().
+	 */
 	spin_lock(&encl->mm_lock);
 	list_add_rcu(&encl_mm->list, &encl->mm_list);

You dropped the smp_wmb().

Quoted 8 lines
Quoted 1 line

As I said to you in my review x86 pipeline does not reorder writes.

And as I pointed out in this thread, smp_wmb() is a _compiler_ barrier if and only if CONFIG_SMP=y. The compiler can reorder list_add_rcu() and mm_list_version++ because from it's perspective there is no dependency between the two. And that's entirely true except for the SMP case where the consumer of mm_list_version is relying on the list to be updated before the version changes.

I see.

So why not change the variable volatile given that x86 is the only arch that this code gets used?

/Jarkko

#10 lore raw
Quoted 28 lines · Jarkko Sakkinen
Quoted 18 lines · Sean Christopherson
Quoted 17 lines · Jarkko Sakkinen
Quoted 15 lines · Sean Christopherson

arch/x86/kernel/cpu/sgx/encl.c int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm) · line 233

Quoted 12 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

arch/x86/kernel/cpu/sgx/encl.c +5 -0
@@ -221,12 +224,16 @@ int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm)
 		return ret;
 	}
 
+	/*
+	 * The page reclaimer uses list version for synchronization instead of
+	 * synchronize_scru() because otherwise we could conflict with
+	 * dup_mmap().
+	 */
 	spin_lock(&encl->mm_lock);
 	list_add_rcu(&encl_mm->list, &encl->mm_list);

You dropped the smp_wmb().

Quoted 8 lines
Quoted 1 line

As I said to you in my review x86 pipeline does not reorder writes.

And as I pointed out in this thread, smp_wmb() is a _compiler_ barrier if and only if CONFIG_SMP=y. The compiler can reorder list_add_rcu() and mm_list_version++ because from it's perspective there is no dependency between the two. And that's entirely true except for the SMP case where the consumer of mm_list_version is relying on the list to be updated before the version changes.

Quoted 4 lines

I see.

So why not change the variable volatile given that x86 is the only arch that this code gets used?

Please note that I'm fully aware of

https://www.kernel.org/doc/html/latest/process/volatile-considered-harmful.html

Just wondering. Anyway, I'll add smp_wmb() back since it is safe play in terms of acceptance.

/Jarkko

#11 lore raw
Quoted 34 lines · Jarkko Sakkinen
Quoted 28 lines · Jarkko Sakkinen
Quoted 18 lines · Sean Christopherson
Quoted 17 lines · Jarkko Sakkinen
Quoted 15 lines · Sean Christopherson

arch/x86/kernel/cpu/sgx/encl.c int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm) · line 233

Quoted 12 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

arch/x86/kernel/cpu/sgx/encl.c +5 -0
@@ -221,12 +224,16 @@ int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm)
 		return ret;
 	}
 
+	/*
+	 * The page reclaimer uses list version for synchronization instead of
+	 * synchronize_scru() because otherwise we could conflict with
+	 * dup_mmap().
+	 */
 	spin_lock(&encl->mm_lock);
 	list_add_rcu(&encl_mm->list, &encl->mm_list);

You dropped the smp_wmb().

Quoted 8 lines
Quoted 1 line

As I said to you in my review x86 pipeline does not reorder writes.

And as I pointed out in this thread, smp_wmb() is a _compiler_ barrier if and only if CONFIG_SMP=y. The compiler can reorder list_add_rcu() and mm_list_version++ because from it's perspective there is no dependency between the two. And that's entirely true except for the SMP case where the consumer of mm_list_version is relying on the list to be updated before the version changes.

Quoted 4 lines

I see.

So why not change the variable volatile given that x86 is the only arch that this code gets used?

Quoted 6 lines

Please note that I'm fully aware of

https://www.kernel.org/doc/html/latest/process/volatile-considered-harmful.html

Just wondering. Anyway, I'll add smp_wmb() back since it is safe play in terms of acceptance.

Because volatile is overkill and too heavy-handed for what is needed here. E.g. if SMP=n, then there are no barriers needed whatsover, as a CPU can't race with itself.

Even for SMP=y, volatile is too much as it prevents optimization that would otherwise be allowed. For the smp_wmb() it doesn't matter much, if at all, since the list and version are updated inside of a critical section, i.e. barriered on both sides so the compiler can't do much optimization anyways.

But for the smp_rmb() case, the compiler is free to cache the version in a register early on in the function, it just needs to make sure that the access is done before starting the list walk. If encl->mm_list_version were volatile, the compiler would not be allowed to do such an optimization as it'd be required to access memory exactly when it's referenced in code.

This is easily visible in the compiled code, as encl->mm_list_version is only read from memory once per iteration (in the unlikely case that there are multiple iterations). It takes the do-while loop, which appears to read encl->mm_list_version twice per iteration:

  do {
        mm_list_version = encl->mm_list_version;

<walk list>

  } while (unlikely(encl->mm_list_version != mm_list_version));

And turns it into an optimized loop that loads encl->mm_list_version the minimum number of times. If encl->mm_list_version list were volatile, the compiler would not be allowed to cache it in %rax.

        mov    0x58(%r12),%rax          // Load from encl->mm_list_version
        lea    0x40(%r12),%rbx          // Interleaved to optimize ALU vs LD
        and    $0xfffffffffffff000,%rbp // Interleaved to optimize ALU vs LD
        mov    %rax,0x8(%rsp)           // mm_list_version = encl->mm_list_version

walk_mm_list:

        <blah blah blah>

        mov    0x58(%r12),%rax          // Load from encl->mm_list_version
        cmp    0x8(%rsp),%rax           // Compare against mm_list_version
        jne    update_mm_list_version

        <happy path>
        ret

update_mm_list_version:

        mov    %rax,0x8(%rsp)           // mm_list_version = encl->mm_list_version
        jmpq   0xffffffff8102e460 <walk_mm_list>
#12 lore raw
Quoted 90 lines · Sean Christopherson
Quoted 34 lines · Jarkko Sakkinen
Quoted 28 lines · Jarkko Sakkinen
Quoted 18 lines · Sean Christopherson
Quoted 17 lines · Jarkko Sakkinen
Quoted 15 lines · Sean Christopherson

Review int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm) · line 233

Quoted 12 lines · Jarkko Sakkinen

From: Sean Christopherson <[email protected]>

(quoted diff) +5 -0
@@ -221,12 +224,16 @@ int sgx_encl_mm_add(struct sgx_encl *encl, struct mm_struct *mm)
 		return ret;
 	}
 
+	/*
+	 * The page reclaimer uses list version for synchronization instead of
+	 * synchronize_scru() because otherwise we could conflict with
+	 * dup_mmap().
+	 */
 	spin_lock(&encl->mm_lock);
 	list_add_rcu(&encl_mm->list, &encl->mm_list);

You dropped the smp_wmb().

Quoted 8 lines
Quoted 1 line

As I said to you in my review x86 pipeline does not reorder writes.

And as I pointed out in this thread, smp_wmb() is a _compiler_ barrier if and only if CONFIG_SMP=y. The compiler can reorder list_add_rcu() and mm_list_version++ because from it's perspective there is no dependency between the two. And that's entirely true except for the SMP case where the consumer of mm_list_version is relying on the list to be updated before the version changes.

Quoted 4 lines

I see.

So why not change the variable volatile given that x86 is the only arch that this code gets used?

Quoted 6 lines

Please note that I'm fully aware of

https://www.kernel.org/doc/html/latest/process/volatile-considered-harmful.html

Just wondering. Anyway, I'll add smp_wmb() back since it is safe play in terms of acceptance.

Because volatile is overkill and too heavy-handed for what is needed here. E.g. if SMP=n, then there are no barriers needed whatsover, as a CPU can't race with itself.

Even for SMP=y, volatile is too much as it prevents optimization that would otherwise be allowed. For the smp_wmb() it doesn't matter much, if at all, since the list and version are updated inside of a critical section, i.e. barriered on both sides so the compiler can't do much optimization anyways.

But for the smp_rmb() case, the compiler is free to cache the version in a register early on in the function, it just needs to make sure that the access is done before starting the list walk. If encl->mm_list_version were volatile, the compiler would not be allowed to do such an optimization as it'd be required to access memory exactly when it's referenced in code.

This is easily visible in the compiled code, as encl->mm_list_version is only read from memory once per iteration (in the unlikely case that there are multiple iterations). It takes the do-while loop, which appears to read encl->mm_list_version twice per iteration:

  do {
        mm_list_version = encl->mm_list_version;

<walk list>

  } while (unlikely(encl->mm_list_version != mm_list_version));

And turns it into an optimized loop that loads encl->mm_list_version the minimum number of times. If encl->mm_list_version list were volatile, the compiler would not be allowed to cache it in %rax.

        mov    0x58(%r12),%rax          // Load from encl->mm_list_version
        lea    0x40(%r12),%rbx          // Interleaved to optimize ALU vs LD
        and    $0xfffffffffffff000,%rbp // Interleaved to optimize ALU vs LD
        mov    %rax,0x8(%rsp)           // mm_list_version = encl->mm_list_version

walk_mm_list:

        <blah blah blah>

        mov    0x58(%r12),%rax          // Load from encl->mm_list_version
        cmp    0x8(%rsp),%rax           // Compare against mm_list_version
        jne    update_mm_list_version

        <happy path>
        ret
Quoted 3 lines

update_mm_list_version:

        mov    %rax,0x8(%rsp)           // mm_list_version = encl->mm_list_version
        jmpq   0xffffffff8102e460 <walk_mm_list>

Thanks and check v4 so that I can squash it :-)

/Jarkko