Make individual print calls thread safe #1290
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Thomas1664
commented
Nov 28, 2023
BillyONeal
reviewed
Nov 28, 2023
There was a problem hiding this comment.
This is, at a minimum, missing documentation on what the thread safety guarantees are to callers. In the header there needs to be a statement like "This function is safe to call from multiple threads. When called from multiple threads, calls are atomic with respect to other callers, but are not atomic with respect to the underlying console due to buffering."
That said, I'm not sure this is the correct approach. It doesn't matter how thread safe print is if callers are doing
print("the answer is ");
print(std::to_string(42));
For example, println is still broken:
vcpkg-tool/include/vcpkg/base/messages.h
Lines 192 to 196 in 1767aae
My intent of pointing out the printing issue in the previous PR was that it indicated to me that the careful analysis necessary to safely parallelize had not happened, not that I believed the fix was to make that API thread safe.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
BillyONeal
added a commit
to BillyONeal/vcpkg-tool
that referenced
this pull request
Dec 21, 2023
…ome clear that we need a thread safe channel through which errors and warnings can both be reasonably reported. Now that microsoft#1279 is landed and functionally everything in the codebase uses ExpectedL, we can look at what the new thing that fixes issues is. Consider the following: ```c++ ExpectedL<T> example_api(int a); ExpectedL<std::unique_ptr<SourceControlFile>> try_load_port_manifest_text(StringView text, StringView control_path, MessageSink& warning_sink); ``` The reason this can't return the warnings through the ExpectedL channel is that we don't want the 'error' state to be engaged when there are merely warnings. Moreover, that these channels are different channels means that situations that might want to return errors and warnings together, as happens when parsing files, means that order relationships between errors and warnings is lost. It is probably a good idea in general to put warnings and errors about the same location next to each other in the output, but that's hard to do with this interface. Rather than multiplexing everything through the return value, this proposal is to multiplex only the success or failure through the return value, and report any specific error information through an out parameter. 1. Distinguish whether an overall operation succeeded or failed in the return value, but 2. record any errors or warnings via an out parameter. Applying this to the above gives: ```c++ Optional<T> example_api(MessageContext& context, int a); // unique_ptr is already 'optional' std::unique_ptr<SourceControlFile> try_load_port_manifest_text(MessageContext& context, StringView text, StringView control_path); ``` Issues this new mechanism fixes: * Errors and warnings can share the same channel and thus be printed together * The interface between code wanting to report events and the code wanting to consume them is a natural thread synchronization boundary. Other attempts to fix this have been incorrect by synchronizing individual print calls ( microsoft#1290 ) or complex enough that we are not sure they are correct by trying to recover boundaries by reparsing our own error output ( microsoft#908 ) * This shuts down the "error: error:" and similar bugs where it isn't clear who is formatting the overall error message vs. talking about individual components Known issues that are not fixed by this change: * This still doesn't make it easy for callers to programmatically handle specific types of errors. Currently, we have some APIs that still use explicit `std::error_code` because they want to do different things for 'file does not exist' vs. 'there was an I/O error'. Given that this condition isn't well served by the ExpectedL mechanism I don't want to wait until we have a better solution to it to proceed. * Because we aren't making the context parameter the 'success carrier' it's more complex to implement 'warnings as errors' or similar functionality where the caller decides how 'important' something is. I would be in favor of moving all success tests to the context parameter but I'm not proposing that because the other vcpkg maintainers do not like it. * Contextual information / stack problems aren't solved. However, the context parameter might be extended in the future to help with this.
BillyONeal
added a commit
to BillyONeal/vcpkg-tool
that referenced
this pull request
Dec 22, 2023
…ome clear that we need a thread safe channel through which errors and warnings can both be reasonably reported. Now that microsoft#1279 is landed and functionally everything in the codebase uses ExpectedL, we can look at what the new thing that fixes issues is. Consider the following: ```c++ ExpectedL<T> example_api(int a); ExpectedL<std::unique_ptr<SourceControlFile>> try_load_port_manifest_text(StringView text, StringView control_path, MessageSink& warning_sink); ``` The reason this can't return the warnings through the ExpectedL channel is that we don't want the 'error' state to be engaged when there are merely warnings. Moreover, that these channels are different channels means that situations that might want to return errors and warnings together, as happens when parsing files, means that order relationships between errors and warnings is lost. It is probably a good idea in general to put warnings and errors about the same location next to each other in the output, but that's hard to do with this interface. Rather than multiplexing everything through the return value, this proposal is to multiplex only the success or failure through the return value, and report any specific error information through an out parameter. 1. Distinguish whether an overall operation succeeded or failed in the return value, but 2. record any errors or warnings via an out parameter. Applying this to the above gives: ```c++ Optional<T> example_api(MessageContext& context, int a); // unique_ptr is already 'optional' std::unique_ptr<SourceControlFile> try_load_port_manifest_text(MessageContext& context, StringView text, StringView control_path); ``` Issues this new mechanism fixes: * Errors and warnings can share the same channel and thus be printed together * The interface between code wanting to report events and the code wanting to consume them is a natural thread synchronization boundary. Other attempts to fix this have been incorrect by synchronizing individual print calls ( microsoft#1290 ) or complex enough that we are not sure they are correct by trying to recover boundaries by reparsing our own error output ( microsoft#908 ) * This shuts down the "error: error:" and similar bugs where it isn't clear who is formatting the overall error message vs. talking about individual components Known issues that are not fixed by this change: * This still doesn't make it easy for callers to programmatically handle specific types of errors. Currently, we have some APIs that still use explicit `std::error_code` because they want to do different things for 'file does not exist' vs. 'there was an I/O error'. Given that this condition isn't well served by the ExpectedL mechanism I don't want to wait until we have a better solution to it to proceed. * Because we aren't making the context parameter the 'success carrier' it's more complex to implement 'warnings as errors' or similar functionality where the caller decides how 'important' something is. I would be in favor of moving all success tests to the context parameter but I'm not proposing that because the other vcpkg maintainers do not like it. * Contextual information / stack problems aren't solved. However, the context parameter might be extended in the future to help with this.
BillyONeal
added a commit
to BillyONeal/vcpkg-tool
that referenced
this pull request
Dec 29, 2023
…ome clear that we need a thread safe channel through which errors and warnings can both be reasonably reported. Now that microsoft#1279 is landed and functionally everything in the codebase uses ExpectedL, we can look at what the new thing that fixes issues is. Consider the following: ```c++ ExpectedL<T> example_api(int a); ExpectedL<std::unique_ptr<SourceControlFile>> try_load_port_manifest_text(StringView text, StringView control_path, MessageSink& warning_sink); ``` The reason this can't return the warnings through the ExpectedL channel is that we don't want the 'error' state to be engaged when there are merely warnings. Moreover, that these channels are different channels means that situations that might want to return errors and warnings together, as happens when parsing files, means that order relationships between errors and warnings is lost. It is probably a good idea in general to put warnings and errors about the same location next to each other in the output, but that's hard to do with this interface. Rather than multiplexing everything through the return value, this proposal is to multiplex only the success or failure through the return value, and report any specific error information through an out parameter. 1. Distinguish whether an overall operation succeeded or failed in the return value, but 2. record any errors or warnings via an out parameter. Applying this to the above gives: ```c++ Optional<T> example_api(MessageContext& context, int a); // unique_ptr is already 'optional' std::unique_ptr<SourceControlFile> try_load_port_manifest_text(MessageContext& context, StringView text, StringView control_path); ``` Issues this new mechanism fixes: * Errors and warnings can share the same channel and thus be printed together * The interface between code wanting to report events and the code wanting to consume them is a natural thread synchronization boundary. Other attempts to fix this have been incorrect by synchronizing individual print calls ( microsoft#1290 ) or complex enough that we are not sure they are correct by trying to recover boundaries by reparsing our own error output ( microsoft#908 ) * This shuts down the "error: error:" and similar bugs where it isn't clear who is formatting the overall error message vs. talking about individual components Known issues that are not fixed by this change: * This still doesn't make it easy for callers to programmatically handle specific types of errors. Currently, we have some APIs that still use explicit `std::error_code` because they want to do different things for 'file does not exist' vs. 'there was an I/O error'. Given that this condition isn't well served by the ExpectedL mechanism I don't want to wait until we have a better solution to it to proceed. * Because we aren't making the context parameter the 'success carrier' it's more complex to implement 'warnings as errors' or similar functionality where the caller decides how 'important' something is. I would be in favor of moving all success tests to the context parameter but I'm not proposing that because the other vcpkg maintainers do not like it. * Contextual information / stack problems aren't solved. However, the context parameter might be extended in the future to help with this.
BillyONeal
added a commit
to BillyONeal/vcpkg-tool
that referenced
this pull request
Jan 26, 2024
…ome clear that we need a thread safe channel through which errors and warnings can both be reasonably reported. Now that microsoft#1279 is landed and functionally everything in the codebase uses ExpectedL, we can look at what the new thing that fixes issues is. Consider the following: ```c++ ExpectedL<T> example_api(int a); ExpectedL<std::unique_ptr<SourceControlFile>> try_load_port_manifest_text(StringView text, StringView control_path, MessageSink& warning_sink); ``` The reason this can't return the warnings through the ExpectedL channel is that we don't want the 'error' state to be engaged when there are merely warnings. Moreover, that these channels are different channels means that situations that might want to return errors and warnings together, as happens when parsing files, means that order relationships between errors and warnings is lost. It is probably a good idea in general to put warnings and errors about the same location next to each other in the output, but that's hard to do with this interface. Rather than multiplexing everything through the return value, this proposal is to multiplex only the success or failure through the return value, and report any specific error information through an out parameter. 1. Distinguish whether an overall operation succeeded or failed in the return value, but 2. record any errors or warnings via an out parameter. Applying this to the above gives: ```c++ Optional<T> example_api(MessageContext& context, int a); // unique_ptr is already 'optional' std::unique_ptr<SourceControlFile> try_load_port_manifest_text(MessageContext& context, StringView text, StringView control_path); ``` Issues this new mechanism fixes: * Errors and warnings can share the same channel and thus be printed together * The interface between code wanting to report events and the code wanting to consume them is a natural thread synchronization boundary. Other attempts to fix this have been incorrect by synchronizing individual print calls ( microsoft#1290 ) or complex enough that we are not sure they are correct by trying to recover boundaries by reparsing our own error output ( microsoft#908 ) * This shuts down the "error: error:" and similar bugs where it isn't clear who is formatting the overall error message vs. talking about individual components Known issues that are not fixed by this change: * This still doesn't make it easy for callers to programmatically handle specific types of errors. Currently, we have some APIs that still use explicit `std::error_code` because they want to do different things for 'file does not exist' vs. 'there was an I/O error'. Given that this condition isn't well served by the ExpectedL mechanism I don't want to wait until we have a better solution to it to proceed. * Because we aren't making the context parameter the 'success carrier' it's more complex to implement 'warnings as errors' or similar functionality where the caller decides how 'important' something is. I would be in favor of moving all success tests to the context parameter but I'm not proposing that because the other vcpkg maintainers do not like it. * Contextual information / stack problems aren't solved. However, the context parameter might be extended in the future to help with this.
Thomas1664
changed the title
I know but this isn't the point of this PR. Here, the goal really is to make single prints thread safe. However, the |
Motivation, benefits? |
Anything that prints something can't be parallelized right now. Protection with mutexes on the calling site doesn't work because prints may hide behind function calls. |
I don't like these approaches because you would have to pass this down the call stack and potentially down to things like |
2 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Currently, printing isn't thread safe because colorization requires multiple calls to the system API. Furthermore, functions like println wrapped multiple prints inside which may lead to \n being printed at the wrong place.
This PR fixes this by synchronizing access to the system API. With this PR, all calls to the Console's print functions are atomic with respect to each other.
However, there is no synchronization mechanism for printing multiple things after one another. Therefore, the next step is to delete
println().