Re: [PATCH 3/3] builtin/repack.c: don't move existing packs out of the way

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Mon, Nov 16, 2020 at 03:29:05PM -0800, Junio C Hamano wrote:
> Taylor Blau <me@xxxxxxxxxxxx> writes:
>
> > This behavior dates all the way back to 2ad47d6 (git-repack: Be
> > careful when updating the same pack as an existing one., 2006-06-25).
> > 2ad47d6 is mainly concerned about a case where a newly written pack
> > would have a different structure than its index. This used to be
> > possible when the pack name was a hash of the set of objects. Under this
> > naming scheme, two packs that store the same set of objects could differ
> > in delta selection, object positioning, or both. If this happened, then
> > any such packs would be unreadable in the instant between copying the
> > new pack and new index (i.e., either the index or pack will be stale
> > depending on the order that they were copied).
>
> True.  So the idea is that we can now pretend that we never wrote
> the new packfile and leave the existing one as-is?

True for packs, which are guaranteed to be unchanged by having the same
checksum in their filename. For other files that use their pack's
checksum in their filename, we want to overwrite the existing copy with
the one we just wrote.

> > This patch is mostly limited to removing code paths that deal with the
> > 'old' prefixing, with the exception of pack metadata.
>
> ... "pack metadata" meaning?  We do not remove and replace the file,
> but we update their mtime to keep these packfiles more fresh than
> other packfiles, or something?

Meaning: .idx, .bitmap, .promisor files. I probably ought to be more
clear in what "metadata" means, since it could easily be confused with
mtime and others.

> > @@ -463,109 +463,34 @@ int cmd_repack(int argc, const char **argv, const char *prefix)
> >
> >  	/*
> >  	 * Ok we have prepared all new packfiles.
> >  	 */
> >  	for_each_string_list_item(item, &names) {
> >  		for (ext = 0; ext < ARRAY_SIZE(exts); ext++) {
> >  			char *fname, *fname_old;
> >
> >  			fname = mkpathdup("%s/pack-%s%s",
> >  					packdir, item->string, exts[ext].name);
> >  			fname_old = mkpathdup("%s-%s%s",
> >  					packtmp, item->string, exts[ext].name);
> > +
> > +			if (((uintptr_t)item->util) & (1 << ext)) {
> > +				struct stat statbuffer;
> > +				if (!stat(fname_old, &statbuffer)) {
> > +					statbuffer.st_mode &= ~(S_IWUSR | S_IWGRP | S_IWOTH);
> > +					chmod(fname_old, statbuffer.st_mode);
> > +				}
> > +
> >  				if (rename(fname_old, fname))
> >  					die_errno(_("renaming '%s' failed"), fname_old);
>
> OK, so this is where the previous step matters.  We do the same as
> before (i.e. stat+chmod and rename) only for paths we have created.
>
> We don't reuse the old one because we have already written a new
> file so we won't save anything by doing so, and checking if the
> target of rename exists already before deciding not to rename cannot
> be made atomic, so just relying on rename() to do the right thing is
> a good idea anyway.
>
> > +			} else if (!exts[ext].optional)
> > +				die(_("missing required file: %s"), fname_old);
>
> If one branch of if/else if.../else requires multi-statement block,
> use {} on all of them, for consistency.

Thanks.

> So, if we wrote a few .$ext for a packfile but not some .$ext, if
> the one we didn't write is among the necessary one, we are in
> trouble?  OK.

Yes, exactly.

> > +			else if (unlink(fname) < 0 && errno != ENOENT)
> > +				die_errno(_("could not unlink: %s"), fname);
>
> And if we wrote .pack and .idx but not .bitmap, the old .bitmpa that
> has the same pack hash may be stale and we discard it for safety?
> That sounds "prudent" but it is not immdiately clear from what
> danger we are protecting ourselves.
>
> In any case, much of what I speculated while reading the proposed
> log message turned out to be false, which may be a sign that the log
> message did not explain the approach clearly enough.  I thought that
> a newly created file that happened to be identical to existing ones
> would be discarded without getting renamed to their final location,
> but the code does not do such special casing.  I thought the
> 'metadata' it talks about were to compensate for side effects of
> reusing the old files, but that was not what the 'metadata' was even
> about.

It's more about: we trust what pack-objects wrote to be the current
state of things. So, if a bitmap already exists, but the caller did a
back-to-back repack and this time generates their bitmap with the
hash cache extension, we prefer the newer bitmap to the one that already
existed.

That goes both ways: if pack-objects _didn't_ write a file that already
exists, we want to drop the existing file. Peff seems to talk more about
this in his email, which I'll respond to now...

> Other than that, the change in [2/3] and [3/3] look quite sensible
> (I am not saying [1/3] is bad---I haven't looked at it yet).
>
> Thanks.

Thanks,
Taylor



[Index of Archives]     [Linux Kernel Development]     [Gcc Help]     [IETF Annouce]     [DCCP]     [Netdev]     [Networking]     [Security]     [V4L]     [Bugtraq]     [Yosemite]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux SCSI]     [Fedora Users]

  Powered by Linux