-
Notifications
You must be signed in to change notification settings - Fork 29
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[DRAFT] Enable IPC tests with the proxy library #868
base: main
Are you sure you want to change the base?
[DRAFT] Enable IPC tests with the proxy library #868
Conversation
b26565d
to
9c8c488
Compare
5842644
to
0ed63c4
Compare
trackingFree()
for providers not supporting free()
op and enable all IPC tests
0ed63c4
to
057c819
Compare
@vinser52 please review |
src/provider/provider_tracking.c
Outdated
@@ -382,6 +382,12 @@ static umf_result_t trackingFree(void *hProvider, void *ptr, size_t size) { | |||
ret = umfMemoryProviderFree(p->hUpstream, ptr, size); | |||
if (ret != UMF_RESULT_SUCCESS) { | |||
LOG_ERR("upstream provider failed to free the memory"); | |||
// Do not add memory back to the tracker, | |||
// if the provider does not support the free() op. | |||
if (ret == UMF_RESULT_ERROR_NOT_SUPPORTED) { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I agree that we should not add memory back to the tracker in case of memory provider does not support free
.
But why is pool implementation tries to call umfMemoryProviderFree()
at all?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Do you mean umfMemoryProviderFree()
above ? - inside trackingFree()
in line no. 382 ?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
With such fix we get a memory leak.
Imagine the pool calls umfMemoryProviderFree()
for the address range [a. b)
. After that pool calls umfMemoryProviderAlloc()
my understanding is that memory provider will never return the range [a. b)
as a result of malloc, but it will return the next chunk [b,c)
. If the pool will call umfMemoryProviderFree()
for the [b,c)
range it also will be lost and never be used again.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, the file provider and the devdax provider work exactly as as you wrote - they were designed to work in this way.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I agree with @ldorau - this is how these providers work. And because of this, in the real-life scenarios they should be used either with a pool that never calls free() or with CoarseProvider
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
There is also an additional way fix this bug - free() in devdax/file could do the unmap but we could keep the offset unchanged - this way the free() would be supported and entries would be correctly removed from the tracker while we still keep file/devdax provider logic simple (so the memory would be still unavailable after free).
I agree that at least we are not wasting VA space (but it is not a problem), but since we never decrement the offset in devdax and file providers by design we still have a kind of memory leak. Because if the umfMemoryProviderAlloc
and umfMemoryProviderFree
are called multiple times, after some time umfMemoryProviderAlloc
will return OOM because offset reached the end of file/devdax.
We discussed this with @ldorau offline. My main comment that currently this PR eliminates the issue:
- today we have a memory leak if
umfMemoryProviderFree
is called. If we do not remove the range from the tracker our diagnostic in DEBUG mode will detect it and report. - with this PR the memory leak still there but our diagnostic is not able to detect it and report.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I agree with @ldorau - this is how these providers work. And because of this, in the real-life scenarios they should be used either with a pool that never calls free() or with CoarseProvider
Just to clarify, I am not challenging how devdax or file providers work. we agreed on that. My concern is about this particular flow when some pool calls umfmemoryProviderFree
for devdax or file providers. This situation should not happens or should be handled correctly. Today looks like there is a memory leak in that case and this PR does not fix the leak.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Removed (code change). It will be fixed in another way.
Done.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I was saying that providers without free should have a null ptr to function, and we should have an API that pool can check if free function is supported if not - if it's not supported it should never call providerFree().
Or as another option - pool informs umf if it can support provider without free - if pool cannot handle it, we always wrap it with coarse provider.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Code changed.
Done.
src/provider/provider_tracking.c
Outdated
@@ -382,6 +382,12 @@ static umf_result_t trackingFree(void *hProvider, void *ptr, size_t size) { | |||
ret = umfMemoryProviderFree(p->hUpstream, ptr, size); | |||
if (ret != UMF_RESULT_SUCCESS) { | |||
LOG_ERR("upstream provider failed to free the memory"); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
in case the free() is not supported we should call LOG_ERR. Please do not nest ifs here and use switch or if/else if
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Removed (code change). It will be fixed in another way.
Done.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Code changed.
Done.
057c819
to
2058774
Compare
trackingFree()
for providers not supporting free()
op and enable all IPC testsefef7d0
to
3805ad8
Compare
fc98d1d
to
30fc257
Compare
f49d417
to
e9102ad
Compare
9f3d43b
to
42d02ec
Compare
Add the coarse library that will replace the coarse provider. Signed-off-by: Lukasz Dorau <[email protected]>
Signed-off-by: Lukasz Dorau <[email protected]>
Signed-off-by: Lukasz Dorau <[email protected]>
Signed-off-by: Lukasz Dorau <[email protected]>
Signed-off-by: Lukasz Dorau <[email protected]>
Remove free_not_supp from the ipcTestParams tuple. It is not needed any more. Signed-off-by: Lukasz Dorau <[email protected]>
Fix two error messages and add some new ones. Signed-off-by: Lukasz Dorau <[email protected]>
Signed-off-by: Lukasz Dorau <[email protected]>
42d02ec
to
a25b910
Compare
Description
Enable IPC tests with the proxy library.
Fixes: #864
Requires:
Checklist