Skip to content
Snippets Groups Projects
  • Matt Harbison's avatar
    092eff6833a7
    lfs: infer the blob store URL from paths.default · 092eff6833a7
    Matt Harbison authored
    If `lfs.url` is specified, it takes precedence.  However, now that we support
    serving blobs via hgweb, we shouldn't *require* this setting.  Less
    configuration is better (things will work out of the box once this is sorted
    out), and git has similar functionality.
    
    This is not a complete solution- it isn't able to infer the blob store from an
    explicitly supplied path, and it should consider `paths.default-push` for push.
    The pull solution for that is a bit hacky, and this alone is an improvement for
    the vast majority of cases.
    
    Even though there are only a handful of references to the saved remote store,
    the location of them makes things complicated.
    
      1) downloading files on demand in the revlog flag processor
      2) copying to readonlyvfs with bundlerepo
      3) downloading in the file prefetch hook
      4) the canupload()/skipdownload() checks
      5) uploading blobs
    
    Since revlog doesn't have a repo or ui reference, we can't avoid creating a
    remote store when the extension is loaded.  While the long term goal is to make
    sure the prefetch hook is invoked early for every command for efficiency, this
    handling in the flag processor is needed as a last ditch fetch.
    
    In order to support the clone command, the remote store needs to be created
    later than when the extension loads, since `paths.default` isn't set until just
    before the files are checked out.  Therefore, this patch changes the prefetch
    hook to ignore the saved reference, and build a new one.
    
    The canupload()/skipdownload() checks simply check if the stored instance is a
    `_nullremote`.  Since this can only be set via `lfs.url` (which is reflected in
    the saved reference), checking only the instance created when the extension
    loaded is fine.
    
    The blob uploading function is called from several places:
    
      1) a prepush hook
      2) when writing a new bundle
      3) from infinitepush
    
    The prepush hook gets an exchange.pushop, so it has a path to where the push is
    going.  The bundle writer and infinitepush don't.  Further, bundle creation for
    things like strip and amend are causing blobs to be uploaded.  This seems wrong,
    but I don't want to side track this sorting that out, so punt on trying to
    handle explicit push paths or `paths.default-push`.
    
    I also think that sending blobs to a remote store when pushing to a local repo
    is wrong.  This functionality predates the usercache, so perhaps that's the
    reason for it.  I've got some patches floating around to stop sending blobs
    remotely in this case, and instead write directly to the other repo's blob
    store.  But the tests for corruption handling weren't happy with this change,
    and I don't have time to rewrite them.  So exclude filesystem based paths from
    this for now.
    
    I don't think there's much of a chance to implement `paths.remote:lfsurl` style
    configs, given how early these are resolved vs how late the remote store is
    created.  But git has it, so I threw a TODO in there, in case anyone has ideas.
    
    I have no idea why this is now doing http auth twice when it wasn't before.  I
    don't think the original blobstore's url is ever being used in these cases.
    092eff6833a7
    History
    lfs: infer the blob store URL from paths.default
    Matt Harbison authored
    If `lfs.url` is specified, it takes precedence.  However, now that we support
    serving blobs via hgweb, we shouldn't *require* this setting.  Less
    configuration is better (things will work out of the box once this is sorted
    out), and git has similar functionality.
    
    This is not a complete solution- it isn't able to infer the blob store from an
    explicitly supplied path, and it should consider `paths.default-push` for push.
    The pull solution for that is a bit hacky, and this alone is an improvement for
    the vast majority of cases.
    
    Even though there are only a handful of references to the saved remote store,
    the location of them makes things complicated.
    
      1) downloading files on demand in the revlog flag processor
      2) copying to readonlyvfs with bundlerepo
      3) downloading in the file prefetch hook
      4) the canupload()/skipdownload() checks
      5) uploading blobs
    
    Since revlog doesn't have a repo or ui reference, we can't avoid creating a
    remote store when the extension is loaded.  While the long term goal is to make
    sure the prefetch hook is invoked early for every command for efficiency, this
    handling in the flag processor is needed as a last ditch fetch.
    
    In order to support the clone command, the remote store needs to be created
    later than when the extension loads, since `paths.default` isn't set until just
    before the files are checked out.  Therefore, this patch changes the prefetch
    hook to ignore the saved reference, and build a new one.
    
    The canupload()/skipdownload() checks simply check if the stored instance is a
    `_nullremote`.  Since this can only be set via `lfs.url` (which is reflected in
    the saved reference), checking only the instance created when the extension
    loaded is fine.
    
    The blob uploading function is called from several places:
    
      1) a prepush hook
      2) when writing a new bundle
      3) from infinitepush
    
    The prepush hook gets an exchange.pushop, so it has a path to where the push is
    going.  The bundle writer and infinitepush don't.  Further, bundle creation for
    things like strip and amend are causing blobs to be uploaded.  This seems wrong,
    but I don't want to side track this sorting that out, so punt on trying to
    handle explicit push paths or `paths.default-push`.
    
    I also think that sending blobs to a remote store when pushing to a local repo
    is wrong.  This functionality predates the usercache, so perhaps that's the
    reason for it.  I've got some patches floating around to stop sending blobs
    remotely in this case, and instead write directly to the other repo's blob
    store.  But the tests for corruption handling weren't happy with this change,
    and I don't have time to rewrite them.  So exclude filesystem based paths from
    this for now.
    
    I don't think there's much of a chance to implement `paths.remote:lfsurl` style
    configs, given how early these are resolved vs how late the remote store is
    created.  But git has it, so I threw a TODO in there, in case anyone has ideas.
    
    I have no idea why this is now doing http auth twice when it wasn't before.  I
    don't think the original blobstore's url is ever being used in these cases.