Skip to contentSkip to conversation

hmm_range_fault related fixes and legacy API removal v2

PATCH In review Christoph Hellwig 5 participants 24 messages Jul 22, 2019 – Jul 23, 2019

Subsystems (6)
lore
Share & tools
lore
#0 lore raw

Hi Jérôme, Ben and Jason,

below is a series against the hmm tree which fixes up the mmap_sem locking in nouveau and while at it also removes leftover legacy HMM APIs only used by nouveau.

The first 4 patches are a bug fix for nouveau, which I suspect should go into this merge window even if the code is marked as staging, just to avoid people copying the breakage.

Changes since v1:
- don't return the valid state from hmm_range_unregister
- additional nouveau cleanups

#1 lore raw

PATCH 1/6 mm: always return EBUSY for invalid ranges in hmm_range_{fault,snapshot}

We should not have two different error codes for the same condition. In addition this really complicates the code due to the special handling of EAGAIN that drops the mmap_sem due to the FAULT_FLAG_ALLOW_RETRY logic in the core vm.

---
 Documentation/vm/hmm.rst |  2 +-
 mm/hmm.c                 | 10 ++++------
 2 files changed, 5 insertions(+), 7 deletions(-)
Documentation/vm/hmm.rst +1 -1
index 7d90964abbb0..710ce1c701bf 100644
@@ -237,7 +237,7 @@ The usage pattern is::
       ret = hmm_range_snapshot(&range);
       if (ret) {
           up_read(&mm->mmap_sem);
-          if (ret == -EAGAIN) {
+          if (ret == -EBUSY) {
             /*
              * No need to check hmm_range_wait_until_valid() return value
              * on retry we will get proper error with hmm_range_snapshot()
mm/hmm.c +4 -6
index e1eedef129cf..16b6731a34db 100644
@@ -946,7 +946,7 @@ EXPORT_SYMBOL(hmm_range_unregister);
  * @range: range
  * Return: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
  *          permission (for instance asking for write and range is read only),
- *          -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid
+ *          -EBUSY if you need to retry, -EFAULT invalid (ie either no valid
  *          vma or it is illegal to access that range), number of valid pages
  *          in range->pfns[] (from range start address).
  *
@@ -967,7 +967,7 @@ long hmm_range_snapshot(struct hmm_range *range)
 	do {
 		/* If range is no longer valid force retry. */
 		if (!range->valid)
-			return -EAGAIN;
+			return -EBUSY;
 
 		vma = find_vma(hmm->mm, start);
 		if (vma == NULL || (vma->vm_flags & device_vma))
@@ -1062,10 +1062,8 @@ long hmm_range_fault(struct hmm_range *range, bool block)
 
 	do {
 		/* If range is no longer valid force retry. */
-		if (!range->valid) {
-			up_read(&hmm->mm->mmap_sem);
-			return -EAGAIN;
-		}
+		if (!range->valid)
+			return -EBUSY;
 
 		vma = find_vma(hmm->mm, start);
 		if (vma == NULL || (vma->vm_flags & device_vma))
Signature · 6 lines
2.20.1

_______________________________________________
Nouveau mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/nouveau
#2 lore raw

Re: [PATCH 1/6] mm: always return EBUSY for invalid ranges in hmm_range_{fault,snapshot}

Quoted 50 lines · Christoph Hellwig

We should not have two different error codes for the same condition. In addition this really complicates the code due to the special handling of EAGAIN that drops the mmap_sem due to the FAULT_FLAG_ALLOW_RETRY logic in the core vm.

---
 Documentation/vm/hmm.rst |  2 +-
 mm/hmm.c                 | 10 ++++------
 2 files changed, 5 insertions(+), 7 deletions(-)
Documentation/vm/hmm.rst +1 -1
index 7d90964abbb0..710ce1c701bf 100644
@@ -237,7 +237,7 @@ The usage pattern is::
       ret = hmm_range_snapshot(&range);
       if (ret) {
           up_read(&mm->mmap_sem);
-          if (ret == -EAGAIN) {
+          if (ret == -EBUSY) {
             /*
              * No need to check hmm_range_wait_until_valid() return value
              * on retry we will get proper error with hmm_range_snapshot()
mm/hmm.c +2 -2
index e1eedef129cf..16b6731a34db 100644
@@ -946,7 +946,7 @@ EXPORT_SYMBOL(hmm_range_unregister);
  * @range: range
  * Return: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
  *          permission (for instance asking for write and range is read only),
- *          -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid
+ *          -EBUSY if you need to retry, -EFAULT invalid (ie either no valid
  *          vma or it is illegal to access that range), number of valid pages
  *          in range->pfns[] (from range start address).
  *
@@ -967,7 +967,7 @@ long hmm_range_snapshot(struct hmm_range *range)
        do {
                /* If range is no longer valid force retry. */
                if (!range->valid)
-                       return -EAGAIN;
+                       return -EBUSY;
 
                vma = find_vma(hmm->mm, start);
                if (vma == NULL || (vma->vm_flags & device_vma))

mm/hmm.c long hmm_range_fault(struct hmm_range *range, bool block) · line 1066

Quoted 9 lines
mm/hmm.c +2 -4
@@ -1062,10 +1062,8 @@ long hmm_range_fault(struct hmm_range *range, bool block)
 
        do {
                /* If range is no longer valid force retry. */
-               if (!range->valid) {
-                       up_read(&hmm->mm->mmap_sem);
-                       return -EAGAIN;
-               }
+               if (!range->valid)
+                       return -EBUSY;

Is it fine to remove up_read(&hmm->mm->mmap_sem) ?

Quoted 6 lines
mm/hmm.c +0 -1
                vma = find_vma(hmm->mm, start);
                if (vma == NULL || (vma->vm_flags & device_vma))
--

2.20.1

_______________________________________________
Nouveau mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/nouveau

Reviewed-by
#3 lore raw

Re: [PATCH 1/6] mm: always return EBUSY for invalid ranges in hmm_range_{fault,snapshot}

Quoted 49 lines · Souptick Joarder
Quoted 48 lines · Christoph Hellwig

We should not have two different error codes for the same condition. In addition this really complicates the code due to the special handling of EAGAIN that drops the mmap_sem due to the FAULT_FLAG_ALLOW_RETRY logic in the core vm.

 Documentation/vm/hmm.rst |  2 +-
 mm/hmm.c                 | 10 ++++------
 2 files changed, 5 insertions(+), 7 deletions(-)
Documentation/vm/hmm.rst +2 -1
index 7d90964abbb0..710ce1c701bf 100644
+++ b/Documentation/vm/hmm.rst
@@ -237,7 +237,7 @@ The usage pattern is::
       ret = hmm_range_snapshot(&range);
       if (ret) {
           up_read(&mm->mmap_sem);
-          if (ret == -EAGAIN) {
+          if (ret == -EBUSY) {
             /*
              * No need to check hmm_range_wait_until_valid() return value
              * on retry we will get proper error with hmm_range_snapshot()
mm/hmm.c +3 -2
index e1eedef129cf..16b6731a34db 100644
+++ b/mm/hmm.c
@@ -946,7 +946,7 @@ EXPORT_SYMBOL(hmm_range_unregister);
  * @range: range
  * Return: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
  *          permission (for instance asking for write and range is read only),
- *          -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid
+ *          -EBUSY if you need to retry, -EFAULT invalid (ie either no valid
  *          vma or it is illegal to access that range), number of valid pages
  *          in range->pfns[] (from range start address).
  *
@@ -967,7 +967,7 @@ long hmm_range_snapshot(struct hmm_range *range)
        do {
                /* If range is no longer valid force retry. */
                if (!range->valid)
-                       return -EAGAIN;
+                       return -EBUSY;
 
                vma = find_vma(hmm->mm, start);
                if (vma == NULL || (vma->vm_flags & device_vma))
@@ -1062,10 +1062,8 @@ long hmm_range_fault(struct hmm_range *range, bool block)
 
Quoted 10 lines

mm/hmm.c long hmm_range_fault(struct hmm_range *range, bool block) · line 1066

Quoted 8 lines
mm/hmm.c +2 -4
        do {
                /* If range is no longer valid force retry. */
-               if (!range->valid) {
-                       up_read(&hmm->mm->mmap_sem);
-                       return -EAGAIN;
-               }
+               if (!range->valid)
+                       return -EBUSY;

Is it fine to remove up_read(&hmm->mm->mmap_sem) ?

It seems very subtle, but under the covers this calls handle_mm_fault() with FAULT_FLAG_ALLOW_RETRY which will cause the mmap sem to become unlocked along the -EAGAIN return path.

I think without the commit message I wouldn't have been able to understand that, so Christoph, could you also add the comment below please?

Otherwise

Frankly, I find the 'int *locked' scheme of GUP easier to understand..

mm/hmm.c +3 -1
index 16b6731a34db79..54b3a4162ae949 100644
@@ -301,8 +301,10 @@ static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr,
 	flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 	flags |= write_fault ? FAULT_FLAG_WRITE : 0;
 	ret = handle_mm_fault(vma, addr, flags);
-	if (ret & VM_FAULT_RETRY)
+	if (ret & VM_FAULT_RETRY) {
+		/* Note, handle_mm_fault did up_read(&mm->mmap_sem)) */
 		return -EAGAIN;
+	}
 	if (ret & VM_FAULT_ERROR) {
 		*pfn = range->values[HMM_PFN_ERROR];
 		return -EFAULT;
#4 lore raw

Re: [PATCH 1/6] mm: always return EBUSY for invalid ranges in hmm_range_{fault,snapshot}

Quoted 3 lines · Jason Gunthorpe

I think without the commit message I wouldn't have been able to understand that, so Christoph, could you also add the comment below please?

I don't think this belongs into this patch. I can add it as a separate patch under your name and with your signoff if you are ok with that.
_______________________________________________
Nouveau mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/nouveau

#5 lore raw

Re: [PATCH 1/6] mm: always return EBUSY for invalid ranges in hmm_range_{fault,snapshot}

Quoted 7 lines · Christoph Hellwig
Quoted 3 lines · Jason Gunthorpe

I think without the commit message I wouldn't have been able to understand that, so Christoph, could you also add the comment below please?

I don't think this belongs into this patch. I can add it as a separate patch under your name and with your signoff if you are ok with that.

Yep, thanks

Jason

#6 lore raw

PATCH 2/6 mm: move hmm_vma_range_done and hmm_vma_fault to nouveau

In reply to Christoph Hellwig

These two functions are marked as a legacy APIs to get rid of, but seem to suit the current nouveau flow. Move it to the only user in preparation for fixing a locking bug involving caller and callee. All comments referring to the old API have been removed as this now is a driver private helper.

---
 drivers/gpu/drm/nouveau/nouveau_svm.c | 45 +++++++++++++++++++++-
 include/linux/hmm.h                   | 54 ---------------------------
 2 files changed, 43 insertions(+), 56 deletions(-)
drivers/gpu/drm/nouveau/nouveau_svm.c +43 -2
index 8c92374afcf2..cde09003c06b 100644
@@ -475,6 +475,47 @@ nouveau_svm_fault_cache(struct nouveau_svm *svm,
 		fault->inst, fault->addr, fault->access);
 }
 
+static inline bool nouveau_range_done(struct hmm_range *range)
+{
+	bool ret = hmm_range_valid(range);
+
+	hmm_range_unregister(range);
+	return ret;
+}
+
+static int
+nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range,
+		    bool block)
+{
+	long ret;
+
+	range->default_flags = 0;
+	range->pfn_flags_mask = -1UL;
+
+	ret = hmm_range_register(range, mirror,
+				 range->start, range->end,
+				 PAGE_SHIFT);
+	if (ret)
+		return (int)ret;
+
+	if (!hmm_range_wait_until_valid(range, HMM_RANGE_DEFAULT_TIMEOUT)) {
+		up_read(&range->vma->vm_mm->mmap_sem);
+		return -EAGAIN;
+	}
+
+	ret = hmm_range_fault(range, block);
+	if (ret <= 0) {
+		if (ret == -EBUSY || !ret) {
+			up_read(&range->vma->vm_mm->mmap_sem);
+			ret = -EBUSY;
+		} else if (ret == -EAGAIN)
+			ret = -EBUSY;
+		hmm_range_unregister(range);
+		return ret;
+	}
+	return 0;
+}
+
 static int
 nouveau_svm_fault(struct nvif_notify *notify)
 {
@@ -649,10 +690,10 @@ nouveau_svm_fault(struct nvif_notify *notify)
 		range.values = nouveau_svm_pfn_values;
 		range.pfn_shift = NVIF_VMM_PFNMAP_V0_ADDR_SHIFT;
 again:
-		ret = hmm_vma_fault(&svmm->mirror, &range, true);
+		ret = nouveau_range_fault(&svmm->mirror, &range, true);
 		if (ret == 0) {
 			mutex_lock(&svmm->mutex);
-			if (!hmm_vma_range_done(&range)) {
+			if (!nouveau_range_done(&range)) {
 				mutex_unlock(&svmm->mutex);
 				goto again;
 			}
include/linux/hmm.h +0 -54

60 lines not on this page · raw message

Signature · 1 line
2.20.1
Reviewed-by on v1 2/6
#7 lore raw

Re: [PATCH 2/6] mm: move hmm_vma_range_done and hmm_vma_fault to nouveau

Quoted 11 lines · Christoph Hellwig

These two functions are marked as a legacy APIs to get rid of, but seem to suit the current nouveau flow. Move it to the only user in preparation for fixing a locking bug involving caller and callee. All comments referring to the old API have been removed as this now is a driver private helper.

---
 drivers/gpu/drm/nouveau/nouveau_svm.c | 45 +++++++++++++++++++++-
 include/linux/hmm.h                   | 54 ---------------------------
 2 files changed, 43 insertions(+), 56 deletions(-)

Jason

#8 lore raw

PATCH 3/6 nouveau: remove the block parameter to nouveau_range_fault

In reply to Christoph Hellwig

The parameter is always false, so remove it as well as the -EAGAIN handling that can only happen for the non-blocking case.

---
 drivers/gpu/drm/nouveau/nouveau_svm.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)
drivers/gpu/drm/nouveau/nouveau_svm.c +3 -4
index cde09003c06b..5dd83a46578f 100644
@@ -484,8 +484,7 @@ static inline bool nouveau_range_done(struct hmm_range *range)
 }
 
 static int
-nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range,
-		    bool block)
+nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range)
 {
 	long ret;
 
@@ -503,7 +502,7 @@ nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range,
 		return -EAGAIN;
 	}
 
-	ret = hmm_range_fault(range, block);
+	ret = hmm_range_fault(range, true);
 	if (ret <= 0) {
 		if (ret == -EBUSY || !ret) {
 			up_read(&range->vma->vm_mm->mmap_sem);
@@ -690,7 +689,7 @@ nouveau_svm_fault(struct nvif_notify *notify)
 		range.values = nouveau_svm_pfn_values;
 		range.pfn_shift = NVIF_VMM_PFNMAP_V0_ADDR_SHIFT;
 again:
-		ret = nouveau_range_fault(&svmm->mirror, &range, true);
+		ret = nouveau_range_fault(&svmm->mirror, &range);
 		if (ret == 0) {
 			mutex_lock(&svmm->mutex);
 			if (!nouveau_range_done(&range)) {
Signature · 1 line
2.20.1
Reviewed-by on v1 3/6
#9 lore raw

Re: [PATCH 3/6] nouveau: remove the block parameter to nouveau_range_fault

Quoted 2 lines · Christoph Hellwig

The parameter is always false, so remove it as well as the -EAGAIN handling that can only happen for the non-blocking case.

? Did the EAGAIN handling get removed in this patch?

Quoted 3 lines
 drivers/gpu/drm/nouveau/nouveau_svm.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

Code seems fine

Jason
_______________________________________________
Nouveau mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/nouveau

#10 lore raw

Re: [PATCH 3/6] nouveau: remove the block parameter to nouveau_range_fault

Quoted 5 lines · Jason Gunthorpe
Quoted 2 lines · Christoph Hellwig

The parameter is always false, so remove it as well as the -EAGAIN handling that can only happen for the non-blocking case.

? Did the EAGAIN handling get removed in this patch?

No. The next revision will remove it.

#11 lore raw

PATCH 4/6 nouveau: unlock mmap_sem on all errors from nouveau_range_fault

In reply to Christoph Hellwig

Currently nouveau_svm_fault expects nouveau_range_fault to never unlock mmap_sem, but the latter unlocks it for a random selection of error codes. Fix this up by always unlocking mmap_sem for non-zero return values in nouveau_range_fault, and only unlocking it in the caller for successful returns.

---
 drivers/gpu/drm/nouveau/nouveau_svm.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/nouveau/nouveau_svm.c +6 -6
index 5dd83a46578f..5de2d54b9782 100644
@@ -494,8 +494,10 @@ nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range)
 	ret = hmm_range_register(range, mirror,
 				 range->start, range->end,
 				 PAGE_SHIFT);
-	if (ret)
+	if (ret) {
+		up_read(&range->vma->vm_mm->mmap_sem);
 		return (int)ret;
+	}
 
 	if (!hmm_range_wait_until_valid(range, HMM_RANGE_DEFAULT_TIMEOUT)) {
 		up_read(&range->vma->vm_mm->mmap_sem);
@@ -504,11 +506,9 @@ nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range)
 
 	ret = hmm_range_fault(range, true);
 	if (ret <= 0) {
-		if (ret == -EBUSY || !ret) {
-			up_read(&range->vma->vm_mm->mmap_sem);
-			ret = -EBUSY;
-		} else if (ret == -EAGAIN)
+		if (ret == 0)
 			ret = -EBUSY;
+		up_read(&range->vma->vm_mm->mmap_sem);
 		hmm_range_unregister(range);
 		return ret;
 	}
@@ -706,8 +706,8 @@ nouveau_svm_fault(struct nvif_notify *notify)
 						NULL);
 			svmm->vmm->vmm.object.client->super = false;
 			mutex_unlock(&svmm->mutex);
+			up_read(&svmm->mm->mmap_sem);
 		}
-		up_read(&svmm->mm->mmap_sem);
 
 		/* Cancel any faults in the window whose pages didn't manage
 		 * to keep their valid bit, or stay writeable when required.
Signature · 1 line
2.20.1
#12 lore raw

Re: [PATCH 4/6] nouveau: unlock mmap_sem on all errors from nouveau_range_fault

Quoted 28 lines · Christoph Hellwig

Currently nouveau_svm_fault expects nouveau_range_fault to never unlock mmap_sem, but the latter unlocks it for a random selection of error codes. Fix this up by always unlocking mmap_sem for non-zero return values in nouveau_range_fault, and only unlocking it in the caller for successful returns.

 drivers/gpu/drm/nouveau/nouveau_svm.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/nouveau/nouveau_svm.c +4 -1
index 5dd83a46578f..5de2d54b9782 100644
+++ b/drivers/gpu/drm/nouveau/nouveau_svm.c
@@ -494,8 +494,10 @@ nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range)
 	ret = hmm_range_register(range, mirror,
 				 range->start, range->end,
 				 PAGE_SHIFT);
-	if (ret)
+	if (ret) {
+		up_read(&range->vma->vm_mm->mmap_sem);
 		return (int)ret;
+	}
 
 	if (!hmm_range_wait_until_valid(range, HMM_RANGE_DEFAULT_TIMEOUT)) {
 		up_read(&range->vma->vm_mm->mmap_sem);
@@ -504,11 +506,9 @@ nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range)
 
 	ret = hmm_range_fault(range, true);

drivers/gpu/drm/nouveau/nouveau_svm.c nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range) · line 513

Quoted 10 lines
drivers/gpu/drm/nouveau/nouveau_svm.c +2 -4
 	if (ret <= 0) {
-		if (ret == -EBUSY || !ret) {
-			up_read(&range->vma->vm_mm->mmap_sem);
-			ret = -EBUSY;
-		} else if (ret == -EAGAIN)
+		if (ret == 0)
 			ret = -EBUSY;
+		up_read(&range->vma->vm_mm->mmap_sem);
 		hmm_range_unregister(range);
 		return ret;

Hum..

The caller does this:

again:

		ret = nouveau_range_fault(&svmm->mirror, &range);
		if (ret == 0) {
			mutex_lock(&svmm->mutex);
			if (!nouveau_range_done(&range)) {
				mutex_unlock(&svmm->mutex);
				goto again;

And we can't call nouveau_range_fault() -> hmm_range_fault() without holding the mmap_sem, so we can't allow nouveau_range_fault to unlock it.

Maybe this instead?

drivers/gpu/drm/nouveau/nouveau_svm.c +3 -8
index a9c5c58d425b3d..92cf760a9bcc5d 100644
@@ -494,21 +494,16 @@ nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range)
 	ret = hmm_range_register(range, mirror,
 				 range->start, range->end,
 				 PAGE_SHIFT);
-	if (ret) {
-		up_read(&range->vma->vm_mm->mmap_sem);
+	if (ret)
 		return (int)ret;
-	}
 
-	if (!hmm_range_wait_until_valid(range, HMM_RANGE_DEFAULT_TIMEOUT)) {
-		up_read(&range->vma->vm_mm->mmap_sem);
+	if (!hmm_range_wait_until_valid(range, HMM_RANGE_DEFAULT_TIMEOUT))
 		return -EBUSY;
-	}
 
 	ret = hmm_range_fault(range, true);
 	if (ret <= 0) {
 		if (ret == 0)
 			ret = -EBUSY;
-		up_read(&range->vma->vm_mm->mmap_sem);
 		hmm_range_unregister(range);
 		return ret;
 	}
@@ -706,8 +701,8 @@ nouveau_svm_fault(struct nvif_notify *notify)
 						NULL);
 			svmm->vmm->vmm.object.client->super = false;
 			mutex_unlock(&svmm->mutex);
-			up_read(&svmm->mm->mmap_sem);
 		}
+		up_read(&svmm->mm->mmap_sem);
 
 		/* Cancel any faults in the window whose pages didn't manage
 		 * to keep their valid bit, or stay writeable when required.
#13 lore raw

Re: [PATCH 4/6] nouveau: unlock mmap_sem on all errors from nouveau_range_fault

Quoted 12 lines · Jason Gunthorpe

Hum..

The caller does this:

again:

		ret = nouveau_range_fault(&svmm->mirror, &range);
		if (ret == 0) {
			mutex_lock(&svmm->mutex);
			if (!nouveau_range_done(&range)) {
				mutex_unlock(&svmm->mutex);
				goto again;
Quoted 3 lines

And we can't call nouveau_range_fault() -> hmm_range_fault() without holding the mmap_sem, so we can't allow nouveau_range_fault to unlock it.

Goto again can only happen if nouveau_range_fault was successful, in which case we did not drop mmap_sem.

Also:

drivers/gpu/drm/nouveau/nouveau_svm.c nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range) · line 507

Quoted 6 lines
drivers/gpu/drm/nouveau/nouveau_svm.c +0 -1
 	ret = hmm_range_fault(range, true);
 	if (ret <= 0) {
 		if (ret == 0)
 			ret = -EBUSY;
-		up_read(&range->vma->vm_mm->mmap_sem);
 		hmm_range_unregister(range);

This would hold mmap_sem over hmm_range_unregister, which can lead to deadlock if we call exit_mmap and then acquire mmap_sem again.

#14 lore raw

Re: [PATCH 4/6] nouveau: unlock mmap_sem on all errors from nouveau_range_fault

Quoted 13 lines · Christoph Hellwig
Quoted 12 lines · Jason Gunthorpe

Hum..

The caller does this:

again:

		ret = nouveau_range_fault(&svmm->mirror, &range);
		if (ret == 0) {
			mutex_lock(&svmm->mutex);
			if (!nouveau_range_done(&range)) {
				mutex_unlock(&svmm->mutex);
				goto again;
Quoted 6 lines
Quoted 3 lines

And we can't call nouveau_range_fault() -> hmm_range_fault() without holding the mmap_sem, so we can't allow nouveau_range_fault to unlock it.

Goto again can only happen if nouveau_range_fault was successful, in which case we did not drop mmap_sem.

Oh, right we switch from success = number of pages to success =0..

However the reason this looks so weird to me is that the locking pattern isn't being followed, any result of hmm_range_fault should be ignored until we enter the svmm->mutex and check if there was a colliding invalidation.

So the 'goto again' *should* be possible even if range_fault failed.

But that is not for this patch..

Quoted 9 lines

drivers/gpu/drm/nouveau/nouveau_svm.c nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range) · line 507

Quoted 6 lines
drivers/gpu/drm/nouveau/nouveau_svm.c +0 -1
 	ret = hmm_range_fault(range, true);
 	if (ret <= 0) {
 		if (ret == 0)
 			ret = -EBUSY;
-		up_read(&range->vma->vm_mm->mmap_sem);
 		hmm_range_unregister(range);

This would hold mmap_sem over hmm_range_unregister, which can lead to deadlock if we call exit_mmap and then acquire mmap_sem again.

That reminds me, this code is also leaking hmm_range_unregister() in the success path, right?

I think the right way to structure this is to move the goto again and related into the nouveau_range_fault() so the whole retry algorithm is sensibly self contained.

Jason

#15 lore raw

Re: [PATCH 4/6] nouveau: unlock mmap_sem on all errors from nouveau_range_fault

Quoted 2 lines · Jason Gunthorpe

That reminds me, this code is also leaking hmm_range_unregister() in the success path, right?

No, that is done by hmm_vma_range_done / nouveau_range_done for the success path.

Quoted 4 lines

I think the right way to structure this is to move the goto again and related into the nouveau_range_fault() so the whole retry algorithm is sensibly self contained.

Then we'd take svmm->mutex inside the helper and let the caller unlock that. Either way it is a bit of a mess, and I'd rather prefer if someone has the hardware would do a grand rewrite of this path eventually. Alternatively if no one signs up to mainain this code we should eventually drop it given the staging status.

#16 lore raw

Re: [PATCH 4/6] nouveau: unlock mmap_sem on all errors from nouveau_range_fault

Quoted 6 lines · Christoph Hellwig
Quoted 2 lines · Jason Gunthorpe

That reminds me, this code is also leaking hmm_range_unregister() in the success path, right?

No, that is done by hmm_vma_range_done / nouveau_range_done for the success path.

.. which is done with the mmap_sem held :(

Quoted 9 lines
Quoted 3 lines

I think the right way to structure this is to move the goto again and related into the nouveau_range_fault() so the whole retry algorithm is sensibly self contained.

Then we'd take svmm->mutex inside the helper and let the caller unlock that. Either way it is a bit of a mess, and I'd rather prefer if someone has the hardware would do a grand rewrite of this path eventually. Alternatively if no one signs up to mainain this code we should eventually drop it given the staging status.

I tend to agree with the sentiment, it just makes me sad that all the examples we have of these APIs are so troubled.

Jason

#17 lore raw

PATCH 5/6 nouveau: return -EBUSY when hmm_range_wait_until_valid fails

In reply to Christoph Hellwig

-EAGAIN has a magic meaning for non-blocking faults, so don't overload it. Given that the caller doesn't check for specific error codes this change is purely cosmetic.

---
 drivers/gpu/drm/nouveau/nouveau_svm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/nouveau/nouveau_svm.c +1 -1
index 5de2d54b9782..a9c5c58d425b 100644
@@ -501,7 +501,7 @@ nouveau_range_fault(struct hmm_mirror *mirror, struct hmm_range *range)
 
 	if (!hmm_range_wait_until_valid(range, HMM_RANGE_DEFAULT_TIMEOUT)) {
 		up_read(&range->vma->vm_mm->mmap_sem);
-		return -EAGAIN;
+		return -EBUSY;
 	}
 
 	ret = hmm_range_fault(range, true);
Signature · 1 line
2.20.1
Reviewed-by on v1 5/6
#18 lore raw

Re: [PATCH 5/6] nouveau: return -EBUSY when hmm_range_wait_until_valid fails

Quoted 8 lines · Christoph Hellwig

-EAGAIN has a magic meaning for non-blocking faults, so don't overload it. Given that the caller doesn't check for specific error codes this change is purely cosmetic.

---
 drivers/gpu/drm/nouveau/nouveau_svm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Agree

Jason

#19 lore raw

PATCH 6/6 mm: remove the legacy hmm_pfn_* APIs

In reply to Christoph Hellwig

Switch the one remaining user in nouveau over to its replacement, and remove all the wrappers.

---
 drivers/gpu/drm/nouveau/nouveau_dmem.c |  2 +-
 include/linux/hmm.h                    | 34 --------------------------
 2 files changed, 1 insertion(+), 35 deletions(-)
drivers/gpu/drm/nouveau/nouveau_dmem.c +1 -1
index 1333220787a1..345c63cb752a 100644
@@ -845,7 +845,7 @@ nouveau_dmem_convert_pfn(struct nouveau_drm *drm,
 		struct page *page;
 		uint64_t addr;
 
-		page = hmm_pfn_to_page(range, range->pfns[i]);
+		page = hmm_device_entry_to_page(range, range->pfns[i]);
 		if (page == NULL)
 			continue;
 
include/linux/hmm.h +0 -34
index 7ef56dc18050..9f32586684c9 100644
@@ -290,40 +290,6 @@ static inline uint64_t hmm_device_entry_from_pfn(const struct hmm_range *range,
 		range->flags[HMM_PFN_VALID];
 }
 
-/*
- * Old API:
- * hmm_pfn_to_page()
- * hmm_pfn_to_pfn()
- * hmm_pfn_from_page()
- * hmm_pfn_from_pfn()
- *
- * This are the OLD API please use new API, it is here to avoid cross-tree
- * merge painfullness ie we convert things to new API in stages.
- */
-static inline struct page *hmm_pfn_to_page(const struct hmm_range *range,
-					   uint64_t pfn)
-{
-	return hmm_device_entry_to_page(range, pfn);
-}
-
-static inline unsigned long hmm_pfn_to_pfn(const struct hmm_range *range,
-					   uint64_t pfn)
-{
-	return hmm_device_entry_to_pfn(range, pfn);
-}
-
-static inline uint64_t hmm_pfn_from_page(const struct hmm_range *range,
-					 struct page *page)
-{
-	return hmm_device_entry_from_page(range, page);
-}
-
-static inline uint64_t hmm_pfn_from_pfn(const struct hmm_range *range,
-					unsigned long pfn)
-{
-	return hmm_device_entry_from_pfn(range, pfn);
-}
-
 /*
  * Mirroring: how to synchronize device page table with CPU page table.
  *
Signature · 1 line
2.20.1
Tested-by
#20 lore raw
In reply to Christoph Hellwig
Quoted 6 lines · Christoph Hellwig

Hi Jérôme, Ben and Jason,

below is a series against the hmm tree which fixes up the mmap_sem locking in nouveau and while at it also removes leftover legacy HMM APIs only used by nouveau.

Quoted 8 lines

The first 4 patches are a bug fix for nouveau, which I suspect should go into this merge window even if the code is marked as staging, just to avoid people copying the breakage.

Changes since v1:
- don't return the valid state from hmm_range_unregister
- additional nouveau cleanups

I ran some OpenCL tests from Jerome with nouveau and this series, 5.3.0-rc1, and my two HMM fixes:
("mm/hmm: fix ZONE_DEVICE anon page mapping reuse") ("mm/hmm: Fix bad subpage pointer in try_to_unmap_one")

You can add for the series:

_______________________________________________
Nouveau mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/nouveau

#21 lore raw
Quoted 17 lines · Ralph Campbell
Quoted 6 lines · Christoph Hellwig

Hi Jérôme, Ben and Jason,

below is a series against the hmm tree which fixes up the mmap_sem locking in nouveau and while at it also removes leftover legacy HMM APIs only used by nouveau.

Quoted 8 lines

The first 4 patches are a bug fix for nouveau, which I suspect should go into this merge window even if the code is marked as staging, just to avoid people copying the breakage.

Changes since v1:
- don't return the valid state from hmm_range_unregister
- additional nouveau cleanups

Quoted 7 lines

I ran some OpenCL tests from Jerome with nouveau and this series, 5.3.0-rc1, and my two HMM fixes:
("mm/hmm: fix ZONE_DEVICE anon page mapping reuse") ("mm/hmm: Fix bad subpage pointer in try_to_unmap_one")

You can add for the series:

Thanks, would you also rebase and resend the patch deleting struct hmm_update ?

Jason
_______________________________________________
Nouveau mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/nouveau

#22 lore raw
In reply to Christoph Hellwig
Quoted 9 lines · Christoph Hellwig

Hi Jérôme, Ben and Jason,

below is a series against the hmm tree which fixes up the mmap_sem locking in nouveau and while at it also removes leftover legacy HMM APIs only used by nouveau.

The first 4 patches are a bug fix for nouveau, which I suspect should go into this merge window even if the code is marked as staging, just to avoid people copying the breakage.

Ignoring the STAGING issue I've tried to use the same guideline as for -stable for -rc ..

So this is a real problem, we definitely hit the locking bugs if we retry/etc under stress, so I would be OK to send it to Linus for early-rc.

However, it doesn't look like the 1st patch is fixing a current bug though, the only callers uses blocking = true, so just the middle three are -rc?

Thanks,
Jason

#23 lore raw
Quoted 10 lines · Jason Gunthorpe

Ignoring the STAGING issue I've tried to use the same guideline as for -stable for -rc ..

So this is a real problem, we definitely hit the locking bugs if we retry/etc under stress, so I would be OK to send it to Linus for early-rc.

However, it doesn't look like the 1st patch is fixing a current bug though, the only callers uses blocking = true, so just the middle three are -rc?

nonblocking isn't used anywher, but it is a major, major API bug. Your call, but if it was my tree I'd probably send it to Linus.