1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
|
Towards replacing rpm + yum (0.1):
- drop the filelists from the main package set file, split out to a
secondary file. for package sets that depend on other package sets,
we need to be able to generate properties with owning packages that
are in another set. this way, a package that requires a file, will
look up the provides in the set and find the package that owns it
and then try mark that for update.
- installer part:
- pre install check; check that dirs can be created (no files where
want to create dirs), move config files according to file
flags. (.rpmnew etc)
- store rpm headers for installed packages.
- implement rpm uninstall and update.
- triggers? just say no?
- rpm seems to consider glibc > 2.6.90 to mean greater than
2.6.90-anything. That is, a comparison that doesn't mention the
release field, shouldn't regard the release field of pkgs it
compares against. glibc-common-2.6.90 has
conflicts: glibc < 2.6.90, glibc > 2.6.90
since rpm doesn't let you do glibc != 2.6.90, and
requires: glibc = 2.6.90
will always pull in glibc. But even with a != relation, would
glibc-2.6.90-16 be equal to 2.6.90? glibc 2.7.90-8 dropped it in
favor of requires = 2.7.90-8 (#225806).
- signed packages
- space calculation before transaction, but ideally, do a number of
smaller transactions.
- pre-link changing binaries and libs on disk screwing up checksum?
- pipelined download and install; topo-sort packages in update set,
pick one with all deps in the current set, add it to the current set
and satisfy deps against update set => result: minimal update
transaction. Queue download and install/update transaction for the
packages in the minimal set, start over. This also makes the
installation phase much more interruptible, basically just stop
after a sub-transaction finishes. As we keep the update set around
as a target, we can restart if needed. Probably don't need to, can
just do a new update. During a sub-transaction we should keep the
target set (i.e. the current set to be) around as a lock file
(system.repo.lock or so, see git) so that razor updates are
prevented if the systems crashes during an update.
- implement depsolving between multiple package sets by creating an
iterator that has a sorted list of all installed pkgs from all sets,
all installed requires from all sets, all installed provides from
all sets etc. could be a list of tuples (pkgs index, set index).
should simplify even the two-set depsolving a bit since we can
pretend there's just one set. this should also be useful for the
'overlay set' idea where the system set is actually made up of a
number of sets, but typically a read-only set from a read-only fs
and a read-write set from a r/w fs.
- locking: we use advisory file locking on the system set
(/var/lib/razor/system.repo) to indicate a transaction is in
progress. The locking algorithm is as follows:
1. obtain advisory lock on system set. if this is already taken,
we know that a process is actively modifying the system set and
we have to wait. there's a fcntl that lets you block for the
lock to go away.
2. if a system-next.repo file already exists an earlier razor
process was interrupted or crashed and we may want to clean
that up. the system-next.repo file will record what the
previous instance was trying to do and we can just replay that
to clean up.
3. create the new package set whichever way and write it to
system-next.repo, then start installing/removing rpms.
4. When the update is complete, rename system-next.repo to
system.repo and remove the advisory lock.
we should probably introduce a new object that encapsulates this
sequence, the filename conventions, rpm cache, e.g. struct
razor_image, with operations such as
#define RAZOR_IMAGE_READ 0x01
#define RAZOR_IMAGE_WRITE 0x02
struct razor_image *
razor_image_open(const char *root, unsigned int flags);
int
razor_image_begin_transaction(struct razor_image *image,
struct razor_set *target);
int
razor_image_finish_transaction(struct razor_image *image);
the transaction pipelineing described above sits on top of this,
since each step there needs to complete a full transaction that
writes out a new package set.
for overlay package sets we could do something like
struct razor_image *
razor_image_open_with_base(const char *root, unsigned int flags,
struct razor_image *base);
where base specifies the r/o package set it's layered on. this
allows for stacking several layers of images.
Package set file format items:
- drop the 4k section alignment
- just use strings for header identifiers, make the string pool
section have a fixed string (maybe make "strings" always the first
string so its index is 0), or maybe just require that it's the first
section in the file.
- nail down byte-order of repo file.
- version the sections in the file, put the element size in the header
so we can add stuff to elements in a backwards compatible way.
maybe not necessary, we can just add sections that augment the
sections we want to add to (similar to how rpm has add versioned
deps).
Misc ideas:
- keep history of installed packages/journal of package transaction,
so we can roll back to yesterday, or see what got installed in the
latest yum update.
- transactions, proper recovery, make sure we don't poop our package
database (no more rm /var/lib/rpm/__cache*).
- use hash table for package and property lists so we only store
unique lists (like for string pool).
- use existing, running system as repo; eg
razor update razor://other-box.local evince
to pull eg the latest evince and dependencies from another box. We
should be able to regenerate a rzr pkg from the system so we can
reuse the signature from the originating repo.
- Ok, maybe the fastest package set merge method in the end is to use
the razor_importer, but use a hash table for the properties. This
way we can assign them unique IDs immediately (like tokenizing
strings).
- test suite should be easy, just keep .repo files around and test
different type of upgrades that way (obsoletes, conflicts, file
conflicts, file/dir problems etc). Or maybe just keep a simple file
format ad use a custom importer to create the .repo files.
- try to clean up the
do { ... } while (((e++)->name & RAZOR_ENTRY_LAST) == 0);
idiom for iteration of directories.
- overlay package sets? mount a read-only /usr over nfs or from the
virt-host and have a local package set overlaid over the read-only
one. shouldn't need new features in the core package set data
structure, but should be just conventions on top. we have the base
package set from the r/o system, the overlay set from the local
system and we can have an effective package set which is the merge
of everything from the overlay into the base set. the effective set
is easy to compute and we could do it on the fly or cache it. or
maybe the effective set is the on-disk representation and the
overlay can be computed when needed, we just keep a link back to the
base.
- incremental rawhide repo updates? instead of downloading 10MB zipped
repo every time, download a diff repo? Should be pretty small,
especially if we don't have file checksums in metadata. Filenames
and properties are for the most part already present, typically just
a version bump plus maybe tweaking a couple requires. The upstream
repo can store multiple incremental updates in one big file and
provide an index file that maps updates for a given date (we should
use repo-file checksums though) to a range in the file: Download the
index file, search for a match for your latest rawhide.repo file,
download range of updates that brings it up to date.
- use hash tables for dirs when importing files to avoid qsorting all
files in rawhide.
Bugs:
- eliminate duplicate entries in package property lists.
|