bug#44760: Closure copy in ‘guix system init’ is inefficient

[raingloom]

On Fri, 20 Nov 2020 12:02:25 +0100
Ludovic Courtès <ludo@gnu.org> wrote:

> ‘guix system init’ ends by copying the system’s closure from the
> “host” store to the target store; it also initializes the database of
> that target store.
> 
> That copy is inefficient for several reasons.  Let’s pick one file,
> shred.1.gz, that ends up being copied, and let’s look at its
> occurrences in the strace log of ‘guix system init config.scm
> /tmp/os’:
> 
> --8<---------------cut here---------------start------------->8---
> $ grep -A2 '/shred.1.gz' ,,s
> lstat("/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shred.1.gz",
> {st_mode=S_IFREG|0444, st_size=1490, ...}) = 0 openat(AT_FDCWD,
> "/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shred.1.gz",
> O_RDONLY) = 15 fstat(15, {st_mode=S_IFREG|0444, st_size=1490, ...}) =
> 0 openat(AT_FDCWD,
> "/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shred.1.gz",
> O_WRONLY|O_CREAT|O_TRUNC, 0444) = 16 read(15,
> "\37\213\10\0\0\0\0\0\2\3\215VMs\3336\20\275\363Wluh\354\251L%vg\322:M"...,
> 8192) = 1490 write(16,
> "\37\213\10\0\0\0\0\0\2\3\215VMs\3336\20\275\363Wluh\354\251L%vg\322:M"...,
> 1490) = 1490 -- utimensat(AT_FDCWD,
> "/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shred.1.gz",
> [{tv_sec=1605721025, tv_nsec=616985411} /*
> 2020-11-18T18:37:05.616985411+0100 */, {tv_sec=1, tv_nsec=0} /*
> 1970-01-01T01:00:01+0100 */], 0) = 0
> lstat("/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/sleep.1.gz",
> {st_mode=S_IFREG|0444, st_size=813, ...}) = 0 openat(AT_FDCWD,
> "/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/sleep.1.gz",
> O_RDONLY) = 15 --
> lstat("/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shred.1.gz",
> {st_mode=S_IFREG|0444, st_size=1490, ...}) = 0
> lstat("/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shuf.1.gz",
> {st_mode=S_IFREG|0444, st_size=972, ...}) = 0
> lstat("/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/sleep.1.gz",
> {st_mode=S_IFREG|0444, st_size=813, ...}) = 0 --
> lstat("/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shred.1.gz",
> {st_mode=S_IFREG|0444, st_size=1490, ...}) = 0 openat(AT_FDCWD,
> "/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shred.1.gz",
> O_RDONLY) = 17 lseek(17, 0, SEEK_CUR)                  = 0 read(17,
> "\37\213\10\0\0\0\0\0\2\3\215VMs\3336\20\275\363Wluh\354\251L%vg\322:M"...,
> 1490) = 1490 --
> lstat("/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shred.1.gz",
> {st_mode=S_IFREG|0444, st_size=1490, ...}) = 0 openat(AT_FDCWD,
> "/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shred.1.gz",
> O_RDONLY) = 17 lseek(17, 0, SEEK_CUR)                  = 0 read(17,
> "\37\213\10\0\0\0\0\0\2\3\215VMs\3336\20\275\363Wluh\354\251L%vg\322:M"...,
> 1490) = 1490 --
> link("/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shred.1.gz",
> "/tmp/os/gnu/store/.links/0w0qcs5lp36i89yry91r2ixlghihzf0vc56bpd9yylj342gv82xl")
> = 0
> lstat("/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shuf.1.gz",
> {st_mode=S_IFREG|0444, st_size=972, ...}) = 0 openat(AT_FDCWD,
> "/tmp/os/gnu/store/57xj5gcy1jbl9ai2lnrqnpr0dald9i65-coreutils-8.32/share/man/man1/shuf.1.gz",
> O_RDONLY) = 17 --8<---------------cut
> here---------------end--------------->8---
> 
> First, /tmp/os/…/shred.1.gz is read entirely twice: once in
> ‘register-items’ (in the ‘nar-sha256’ call) to compute its hash, and a
> second time for deduplication (the ‘deduplicate’ call in there.)
> 
> The ‘nar-sha256’ call could be avoided because the database of
> /gnu/store contains that value.  As for deduplication, we could
> perhaps create those ‘.links’ entries as we copy files instead of
> re-traversing the whole thing afterwards.
> 
> Second, all of /tmp/os is traversed to reset timestamps, although we
> could have cleared those timestamps when we created those files in the
> first place (<https://issues.guix.gnu.org/44741> prevents that though,
> unless we keep a bug-fixed copy of ‘copy-recursively’ in there.)
> 
> Third, in the case of the installer, we’re really copying from
> /mnt/guix-inst/store to /mnt/gnu/store, which is likely the same
> device.  In this case we could create hard links instead of actually
> copying files.
> 
> Fourth, we’re adding items one by one in the target store database,
> but it may be more efficient to more or less dump the subset of the
> source database in bulk.
> 
> Surely we can do better.
> 
> Ludo’.
> 
> 
> 

Also, if a store is already present (eg.: because of a partial
install), it could make sense to (optionally) keep its contents. AFAIK
this is still not possible. It was one the bigger time sinks while I
was working on the F2FS support.