diff mbox series

[bug#42023] database: register-items: reduce transaction scope.

Message ID 87bljka234.fsf@cune.org
State Accepted
Headers show
Series [bug#42023] database: register-items: reduce transaction scope. | expand

Checks

Context Check Description
cbaines/comparison success View comparision
cbaines/git branch success View Git branch
cbaines/applying patch fail View Laminar job

Commit Message

Caleb Ristvedt Aug. 9, 2020, 4:13 a.m. UTC 
Ludovic Courtès <ludo@gnu.org> writes:

> Hi reepca,
>
> Did you have time to look into this (see below)?  I still see a lot of
> contention on /var/guix/db/db.sqlite-shm on berlin and I feel like these
> changes would be welcome.  :-)

Apologies for the long delay, I think I'm caught up on this issue
now. Patch series detailed below.

>
> Thanks,
> Ludo’.

>>> So general "fiddling" can't happen, but GC can. The responsibility for
>>> both locking and registering as temporary roots falls on the caller,
>>> currently, as I believe it should. We may want to document this
>>> responsibility in the docstring for register-items, though.
>>
>> Yes, it would be good to add a sentence to document it.

Patch attached.

>>
>>> So currently finalize-store-file is safe from clobbering by another
>>> process attempting to finalize to the same path, but not safe from
>>> garbage collection between when the temporary store file is renamed and
>>> when it is registered. It needs an add-temp-root prior to renaming.
>>
>> Ah, so we’d need to do that before applying the patch that reduces the
>> scope of the transaction, right?

AFAIK the only code path that actually uses finalize-store-file
currently is from the build hook / offload thingy, and it turns out that
the outputs of derivations should already be registered as temp roots by
the daemon process that launched the offload process (specifically
registered in haveDerivation() in nix/libstore/build.cc). So this is
technically not currently necessary before or after the scope-reducing
patch. But it makes sense to guarantee that the register-items
invocation will only happen on items that are protected from garbage
collection, so I've put a patch, right before the documenting of that
requirement, that modifies finalize-store-file to always add the file
being finalized as a temp root for the extent of the register-items
invocation.

>>> Also, replace-with-link doesn't check whether the "containing directory"
>>> is the store like optimisePath_() does, so in theory if another process
>>> tried to make a permanent change to the store's permissions it could be
>>> clobbered when replace-with-link restores the permissions. I don't know
>>> of any instance where this could happen currently, but it's something to
>>> keep in mind.
>>
>> I guess we should also avoid changing permissions on /gnu/store, it
>> would be wiser.

While rebasing my changes I noticed that the current implementation of
this uses (%store-directory) from (guix build utils), which may not
correspond to the #:store keyword argument of 'deduplicate'. So I added
a patch that propagates the store through to replace-with-link and from
there to with-writable-file.

>>> diff --git a/guix/store/deduplication.scm b/guix/store/deduplication.scm
>>> index 6784ee0b92..b6d94e49c2 100644
>>> --- a/guix/store/deduplication.scm
>>> +++ b/guix/store/deduplication.scm
>>> @@ -161,26 +161,44 @@ under STORE."
>>>                    (scandir* path))
>>>          (let ((link-file (string-append links-directory "/"
>>>                                          (bytevector->nix-base32-string hash))))
>>> -          (if (file-exists? link-file)
>>> -              (replace-with-link link-file path
>>> -                                 #:swap-directory links-directory)
>>> -              (catch 'system-error
>>> -                (lambda ()
>>> -                  (link path link-file))
>>> -                (lambda args
>>> -                  (let ((errno (system-error-errno args)))
>>> -                    (cond ((= errno EEXIST)
>>> -                           ;; Someone else put an entry for PATH in
>>> -                           ;; LINKS-DIRECTORY before we could.  Let's use it.
>>> -                           (replace-with-link path link-file
>>> -                                              #:swap-directory links-directory))
>>> -                          ((= errno ENOSPC)
>>> -                           ;; There's not enough room in the directory index for
>>> -                           ;; more entries in .links, but that's fine: we can
>>> -                           ;; just stop.
>>> -                           #f)
>>> -                          ((= errno EMLINK)
>>> -                           ;; PATH has reached the maximum number of links, but
>>> -                           ;; that's OK: we just can't deduplicate it more.
>>> -                           #f)
>>> -                          (else (apply throw args)))))))))))
>>> +          (let retry ()
>>> +            (if (file-exists? link-file)
>>> +                (catch 'system-error
>>> +                  (lambda ()
>>> +                    (replace-with-link link-file path
>>> +                                       #:swap-directory links-directory))
>>> +                  (lambda args
>>> +                    (if (and (= (system-error-errno args) ENOENT)
>>> +                             ;; ENOENT can be produced because:
>>> +                             ;; - LINK-FILE has missing directory components
>>> +                             ;; - LINKS-DIRECTORY has missing directory
>>> +                             ;;   components
>>> +                             ;; - PATH has missing directory components
>>> +                             ;;
>>> +                             ;; the last two are errors, the first just
>>> +                             ;; requires a retry.
>>> +                             (file-exists? (dirname path))
>>> +                             (file-exists? links-directory))
>>> +                        (retry)
>>> +                        (apply throw args))))
>>
>> I feel like there are TOCTTOU issues here and redundant ‘stat’ calls,
>> plus the risk of catching a ‘system-error’ coming from “somewhere else.”
>>
>> What about baking this logic directly in ‘replace-with-link’, and
>> replacing ‘file-exists?’ calls by 'system-error handling?

Patch attached. I've renamed replace-with-link to canonicalize-with-link
since it may now have to create the target link. Unfortunately there are
places where 'file-exists?' error handling is necessary, simply because
of ambiguity in errnos. For example, link(oldpath, newpath) can return
ENOENT, but there's no way of knowing from that alone whether this is
because oldpath has missing directories or newpath has missing
directories or the directory components of oldpath are all present but
the file itself is missing (the case we need to be able to recognize and
retry in case of).

Also, I tried removing the first check for whether the canonical link
exists in favor of error handling like you suggested, but this actually
messes up the hard-coded number of link-invocations expected in
tests/store-deduplication.scm - it expects a single link invocation when
the canonical link already exists, but the error-handling approach uses
two - one to discover it exists, and another to create the temporary
link. I didn't know how to reconcile the testing strategy with this
behavior, so I kept the behavior of first using a 'file-exists?' call to
test for the existence of the canonical link.

All tests pass except for tests/packages.scm, which failed even without
the patches.

- reepca

Comments

Ludovic Courtès Sept. 14, 2020, 8:58 a.m. UTC | #1 
Hi Caleb,

And apologies for the delay.

I think we’ve drifted from the original patch and that’s become a tricky
7-patch series, which partly explains the delay—not that I’m looking for
an excuse.  ;-)

I decided to go ahead and apply some of these on your behalf.  Comments
below.

Caleb Ristvedt <caleb.ristvedt@cune.org> skribis:

> From 4c8f0cc50e2a1a33d9ce2f8e58cc426872676a7f Mon Sep 17 00:00:00 2001
> From: Caleb Ristvedt <caleb.ristvedt@cune.org>
> Date: Wed, 24 Jun 2020 01:00:40 -0500
> Subject: [PATCH 1/6] .dir-locals.el: fix call-with-{retrying-}transaction
>  indenting.
>
> * .dir-locals.el (call-with-transaction, call-with-retrying-transaction):
>   change scheme-indent-function property from 2 to 1.

Applied.

> From 9717568f922e0921e5fdc320cbe6689768d29a29 Mon Sep 17 00:00:00 2001
> From: Caleb Ristvedt <caleb.ristvedt@cune.org>
> Date: Sat, 8 Aug 2020 10:05:22 -0500
> Subject: [PATCH 2/6] deduplication: pass store directory to replace-with-link.
>
> This causes with-writable-file to take into consideration the actual store
> being used, as passed to 'deduplicate', rather than
> whatever (%store-directory) may return.
>
> * guix/store/deduplication.scm (replace-with-link): new keyword argument
>   'store'.  Pass to with-writable-file.
>   (with-writable-file, call-with-writable-file): new store argument.
>   (deduplicate): pass store to replace-with-link.

Applied.

> From b992a3aaac7e3b30222e0bf1df09093f18e25e6a Mon Sep 17 00:00:00 2001
> From: Caleb Ristvedt <caleb.ristvedt@cune.org>
> Date: Sat, 8 Aug 2020 11:25:57 -0500
> Subject: [PATCH 3/6] deduplication: retry on ENOENT.
>
> It's possible for the garbage collector to remove the "canonical" link after
> it's been detected as existing by 'deduplicate'.  This would cause an ENOENT
> error when replace-with-link attempts to create the temporary link.  This
> changes it so that it will properly handle that by retrying.

Would that ENOENT cause an error, or just a missed deduplication opportunity?

> * guix/store/deduplication.scm (replace-with-link): renamed to
>   canonicalize-with-link, now also handles the case where the target link
>   doesn't exist yet, and retries on ENOENT.
>   (deduplicate): modified to use canonicalize-with-link.

There’s an issue with this patch.  I gave it a spin (offloading a few
builds) and it got stuck in a infinite loop:

--8<---------------cut here---------------start------------->8---
stat("/gnu/store/.links/0k63r0n3681r2gqd00blq4z5xd7cw1knv0x049p99f0pq31brhrk", 0x7ffe43898cd0) = -1 ENOENT (Dosiero aŭ dosierujo ne ekzistas)
link("/gnu/store/83jy739bn644w3pnwgb5kwjig0kzs92f-libwps-0.4.12/lib/libwps-0.4.so", "/gnu/store/.links/0k63r0n3681r2gqd00blq4z5xd7cw1knv0x049p99f0pq31brhrk") = -1 EEXIST (Dosiero jam ekzistas)
stat("/gnu/store/.links/0k63r0n3681r2gqd00blq4z5xd7cw1knv0x049p99f0pq31brhrk", 0x7ffe43898cd0) = -1 ENOENT (Dosiero aŭ dosierujo ne ekzistas)
link("/gnu/store/83jy739bn644w3pnwgb5kwjig0kzs92f-libwps-0.4.12/lib/libwps-0.4.so", "/gnu/store/.links/0k63r0n3681r2gqd00blq4z5xd7cw1knv0x049p99f0pq31brhrk") = -1 EEXIST (Dosiero jam ekzistas)
stat("/gnu/store/.links/0k63r0n3681r2gqd00blq4z5xd7cw1knv0x049p99f0pq31brhrk", 0x7ffe43898cd0) = -1 ENOENT (Dosiero aŭ dosierujo ne ekzistas)
link("/gnu/store/83jy739bn644w3pnwgb5kwjig0kzs92f-libwps-0.4.12/lib/libwps-0.4.so", "/gnu/store/.links/0k63r0n3681r2gqd00blq4z5xd7cw1knv0x049p99f0pq31brhrk") = -1 EEXIST (Dosiero jam ekzistas)
--8<---------------cut here---------------end--------------->8---

I think we should work on reducing the complexity of that code (e.g.,
there are several layers of system-error handling).

I’ve omitted it and I propose discussing it in a separate issue if we
need to.

> From 6b7d011680c77642f24396be0eb0015c20413d1a Mon Sep 17 00:00:00 2001
> From: Caleb Ristvedt <caleb.ristvedt@cune.org>
> Date: Sat, 8 Aug 2020 08:31:38 -0500
> Subject: [PATCH 4/6] nar: ensure finalization target is a temp root during
>  registration.
>
> Note that this is currently unnecessary, as finalize-store-file is only used
> from the offload hook, and the daemon process that spawned the offload hook
> will have already registered the derivation outputs as temp roots prior to
> attempting to offload (see haveDerivation() in nix/libstore/build.cc).  But
> it's necessary to ensure that the register-items invocation works properly
> when finalize-store-file is used in other contexts.
>
> * guix/nar.scm (finalize-store-file): make target a temp root during extent of
>   register-items invocation.

[...]

> +          ;; TODO: don't use an RPC for this, add it to *this process's* temp
> +          ;; roots file.
> +          (with-store store
> +            (add-temp-root store target)

I agree that this is the right thing but as you note, it’s currently
unnecessary in the context of ‘guix offload’, and I’d rather avoid
opening more connections to the daemon from ‘guix offload’ and this
increases load, pressure on the store database, etc.

> From 55dd48e88d641bbc17b4d9484d6ee84acfb29766 Mon Sep 17 00:00:00 2001
> From: Caleb Ristvedt <caleb.ristvedt@cune.org>
> Date: Wed, 8 Jul 2020 11:33:23 -0500
> Subject: [PATCH 5/6] database: document extra registration requirements.
>
> It's necessary that store items be locked and protected from garbage
> collection while they are being registered.  This documents that.
>
> * guix/store/database.scm (register-path, register-items): document GC
>   protection and locking requirements.

Applied.

> From 30afb453ce4eb161bb87645a0e6314e6af82a61f Mon Sep 17 00:00:00 2001
> From: Christopher Baines <mail@cbaines.net>
> Date: Tue, 23 Jun 2020 17:36:49 +0100
> Subject: [PATCH 6/6] database: register-items: reduce transaction scope.
>
> It was made transactional in a4678c6ba18d8dbd79d931f80426eebf61be7ebe, with
> the reasoning to prevent broken intermediate states from being visible. I
> think this means something like an entry being in ValidPaths, but the Refs not
> being inserted.
>
> Using a transaction for this makes sense, but I think using one single
> transaction for the whole register-items call is unnecessary to avoid broken
> states from being visible, and could block other writes to the store database
> while register-items is running. Because the deduplication and resetting
> timestamps happens within the transaction as well, even though these things
> don't involve the database, writes to the database will still be blocked while
> this is happening.
>
> To reduce the potential for register-items to block other writers to the
> database for extended periods, this commit moves the transaction to just wrap
> the call to sqlite-register. This is the one place where writes occur, so that
> should prevent the broken intermediate states issue above. The one difference
> this will make is some of the registered items will be visible to other
> connections while others may be still being added. I think this is OK, as it's
> equivalent to just registering different items.
>
> * guix/store/database.scm (register-items): Reduce transaction scope.

Applied.

Thanks Caleb & Chris!

Ludo’.
diff mbox series

Patch

From 30afb453ce4eb161bb87645a0e6314e6af82a61f Mon Sep 17 00:00:00 2001
From: Christopher Baines <mail@cbaines.net>
Date: Tue, 23 Jun 2020 17:36:49 +0100
Subject: [PATCH 6/6] database: register-items: reduce transaction scope.

It was made transactional in a4678c6ba18d8dbd79d931f80426eebf61be7ebe, with
the reasoning to prevent broken intermediate states from being visible. I
think this means something like an entry being in ValidPaths, but the Refs not
being inserted.

Using a transaction for this makes sense, but I think using one single
transaction for the whole register-items call is unnecessary to avoid broken
states from being visible, and could block other writes to the store database
while register-items is running. Because the deduplication and resetting
timestamps happens within the transaction as well, even though these things
don't involve the database, writes to the database will still be blocked while
this is happening.

To reduce the potential for register-items to block other writers to the
database for extended periods, this commit moves the transaction to just wrap
the call to sqlite-register. This is the one place where writes occur, so that
should prevent the broken intermediate states issue above. The one difference
this will make is some of the registered items will be visible to other
connections while others may be still being added. I think this is OK, as it's
equivalent to just registering different items.

* guix/store/database.scm (register-items): Reduce transaction scope.
---
 guix/store/database.scm | 37 +++++++++++++++++++------------------
 1 file changed, 19 insertions(+), 18 deletions(-)

diff --git a/guix/store/database.scm b/guix/store/database.scm
index e39a1603af..2ea63b17aa 100644
--- a/guix/store/database.scm
+++ b/guix/store/database.scm
@@ -457,24 +457,25 @@  typically by adding them as temp-roots."
       (when reset-timestamps?
         (reset-timestamps real-file-name))
       (let-values (((hash nar-size) (nar-sha256 real-file-name)))
-        (sqlite-register db #:path to-register
-                         #:references (store-info-references item)
-                         #:deriver (store-info-deriver item)
-                         #:hash (string-append "sha256:"
-                                               (bytevector->base16-string hash))
-                         #:nar-size nar-size
-                         #:time registration-time)
+        (call-with-retrying-transaction db
+          (lambda ()
+            (sqlite-register db #:path to-register
+                             #:references (store-info-references item)
+                             #:deriver (store-info-deriver item)
+                             #:hash (string-append
+                                     "sha256:"
+                                     (bytevector->base16-string hash))
+                             #:nar-size nar-size
+                             #:time registration-time)))
         (when deduplicate?
           (deduplicate real-file-name hash #:store store-dir)))))
 
-  (call-with-retrying-transaction db
-      (lambda ()
-        (let* ((prefix   (format #f "registering ~a items" (length items)))
-               (progress (progress-reporter/bar (length items)
-                                                prefix log-port)))
-          (call-with-progress-reporter progress
-            (lambda (report)
-              (for-each (lambda (item)
-                          (register db item)
-                          (report))
-                        items)))))))
+  (let* ((prefix   (format #f "registering ~a items" (length items)))
+         (progress (progress-reporter/bar (length items)
+                                          prefix log-port)))
+    (call-with-progress-reporter progress
+      (lambda (report)
+        (for-each (lambda (item)
+                    (register db item)
+                    (report))
+                  items)))))
-- 
2.28.0