1
00:00:00,000 --> 00:00:08,469
foreign

2
00:00:00,500 --> 00:00:08,469
[Music]

3
00:00:12,790 --> 00:00:16,020
[Music]

4
00:00:14,160 --> 00:00:18,539
so many people here

5
00:00:16,020 --> 00:00:19,800
okay so let me introduce

6
00:00:18,539 --> 00:00:24,240
our speaker

7
00:00:19,800 --> 00:00:28,080
Rohan McClure so Rohan is a first year

8
00:00:24,240 --> 00:00:29,160
grad at IBM mostly hacking on the Linux

9
00:00:28,080 --> 00:00:32,160
kernel

10
00:00:29,160 --> 00:00:35,040
the Linux kernel strived to be bad work

11
00:00:32,160 --> 00:00:36,420
compatible for user space

12
00:00:35,040 --> 00:00:40,440
today

13
00:00:36,420 --> 00:00:43,260
Rohan will talk about the time that he

14
00:00:40,440 --> 00:00:45,780
accidentally broke that promise

15
00:00:43,260 --> 00:00:49,909
please welcome Rohan

16
00:00:45,780 --> 00:00:49,909
[Applause]

17
00:00:50,280 --> 00:00:55,620
two apologies before I begin I was awake

18
00:00:53,399 --> 00:00:58,199
in Canberra at four at 4am this morning

19
00:00:55,620 --> 00:00:59,879
and so I may be a bit loopy the other

20
00:00:58,199 --> 00:01:01,860
apology is for presenting a talk about

21
00:00:59,879 --> 00:01:04,379
the Linux kernel from an Apple laptop

22
00:01:01,860 --> 00:01:06,420
but um thankfully for your sake I save

23
00:01:04,379 --> 00:01:09,420
you the second hand stress of watching

24
00:01:06,420 --> 00:01:11,520
me X rander so um

25
00:01:09,420 --> 00:01:13,680
I'd like to begin today with an

26
00:01:11,520 --> 00:01:15,780
acknowledgment of country I begin today

27
00:01:13,680 --> 00:01:17,720
by acknowledging the Warren Jerry Roy

28
00:01:15,780 --> 00:01:20,100
warrang people of the colon Nation

29
00:01:17,720 --> 00:01:22,439
traditional custodians of the land on

30
00:01:20,100 --> 00:01:23,820
which We Gather today and I pay my

31
00:01:22,439 --> 00:01:26,159
respects to their Elders past and

32
00:01:23,820 --> 00:01:27,720
present I extend that respect to

33
00:01:26,159 --> 00:01:29,939
Aboriginal and Torres Strait Islander

34
00:01:27,720 --> 00:01:32,159
peoples here today

35
00:01:29,939 --> 00:01:34,500
hi my name is Rohan thanks for coming to

36
00:01:32,159 --> 00:01:36,180
my talk so for the last year in a bit

37
00:01:34,500 --> 00:01:40,259
I've been hacking on the colonel as well

38
00:01:36,180 --> 00:01:42,780
as openssl and uh just in my first patch

39
00:01:40,259 --> 00:01:45,119
series I managed to really max out on

40
00:01:42,780 --> 00:01:48,799
the number of people I could irritate

41
00:01:45,119 --> 00:01:51,420
um so uh I've made my first patch series

42
00:01:48,799 --> 00:01:52,799
and a month after it's upstreamed I get

43
00:01:51,420 --> 00:01:55,740
this sort of feedback

44
00:01:52,799 --> 00:01:57,899
this breaks power PC 32-bit

45
00:01:55,740 --> 00:01:59,640
uh

46
00:01:57,899 --> 00:02:03,360
probably breaks every Sysco with a

47
00:01:59,640 --> 00:02:04,320
64-bit argument all right great great

48
00:02:03,360 --> 00:02:06,540
um

49
00:02:04,320 --> 00:02:08,220
sorts of damage exactly well a bit of an

50
00:02:06,540 --> 00:02:10,440
introduction to the kernel perhaps it

51
00:02:08,220 --> 00:02:13,140
needs no product no introduction but

52
00:02:10,440 --> 00:02:14,480
I'll give a a brief one at a thousand

53
00:02:13,140 --> 00:02:16,980
feet

54
00:02:14,480 --> 00:02:18,840
there sets of sort of dichotomy between

55
00:02:16,980 --> 00:02:20,640
kernel space which is privileged and

56
00:02:18,840 --> 00:02:23,580
user space which is basically

57
00:02:20,640 --> 00:02:25,020
unprivileged and essentially two

58
00:02:23,580 --> 00:02:26,940
security rings in this sort of

59
00:02:25,020 --> 00:02:28,739
Monolithic kernel

60
00:02:26,940 --> 00:02:30,720
um and so user space relies heavily on

61
00:02:28,739 --> 00:02:32,640
the Kernel then for its Process

62
00:02:30,720 --> 00:02:36,360
Management for its file system access

63
00:02:32,640 --> 00:02:39,420
for just general access to resources

64
00:02:36,360 --> 00:02:41,940
Etc and so with that comes a number of

65
00:02:39,420 --> 00:02:44,459
software contracts and so if you were

66
00:02:41,940 --> 00:02:45,800
let's say trying to break all of user

67
00:02:44,459 --> 00:02:48,239
space that is all

68
00:02:45,800 --> 00:02:49,860
unprivileged programs on your system

69
00:02:48,239 --> 00:02:53,280
what might you do well you might

70
00:02:49,860 --> 00:02:55,620
actually violate these contracts

71
00:02:53,280 --> 00:02:57,959
um for those of us unfamiliar with the

72
00:02:55,620 --> 00:03:00,300
idea of a software contract essentially

73
00:02:57,959 --> 00:03:03,420
if you've programmed it all you've used

74
00:03:00,300 --> 00:03:05,700
some sort of API an API stands for an

75
00:03:03,420 --> 00:03:07,620
application programming interface and

76
00:03:05,700 --> 00:03:08,879
critically I'm going to draw on this

77
00:03:07,620 --> 00:03:12,060
distinction later

78
00:03:08,879 --> 00:03:14,760
it's a sort of contract on source code

79
00:03:12,060 --> 00:03:16,680
it's usually not difficult to consume

80
00:03:14,760 --> 00:03:19,379
someone else's Library when you have

81
00:03:16,680 --> 00:03:22,080
tooling which supports you doing so

82
00:03:19,379 --> 00:03:25,019
that is there's a common ABI which I'll

83
00:03:22,080 --> 00:03:26,940
get to in a moment so that you can call

84
00:03:25,019 --> 00:03:28,920
library code provided the parameters

85
00:03:26,940 --> 00:03:31,200
which the documentation if there is

86
00:03:28,920 --> 00:03:34,620
documentation or examples if there are

87
00:03:31,200 --> 00:03:38,280
examples or code if there are is code

88
00:03:34,620 --> 00:03:41,459
um you can usually call that Library

89
00:03:38,280 --> 00:03:43,080
um subject to its specification and an

90
00:03:41,459 --> 00:03:45,659
API implies that that library is

91
00:03:43,080 --> 00:03:48,840
intended to be stable to some extent and

92
00:03:45,659 --> 00:03:51,080
it has exported functions and you can

93
00:03:48,840 --> 00:03:54,120
call it without risk of personal injury

94
00:03:51,080 --> 00:03:57,000
at some rate but uh contrast that

95
00:03:54,120 --> 00:03:58,920
however to an ABI an application binary

96
00:03:57,000 --> 00:04:01,260
interface

97
00:03:58,920 --> 00:04:03,959
what keeps programs which were compiled

98
00:04:01,260 --> 00:04:06,239
yesterday or just as easily 20 years ago

99
00:04:03,959 --> 00:04:07,980
what keeps them running on Modern

100
00:04:06,239 --> 00:04:10,080
systems well that's that's an

101
00:04:07,980 --> 00:04:12,780
application binary interface and these

102
00:04:10,080 --> 00:04:15,720
are extremely prescriptive this covers

103
00:04:12,780 --> 00:04:17,699
stuff like calling Convention as well as

104
00:04:15,720 --> 00:04:19,260
the focus of this talk is on the syscall

105
00:04:17,699 --> 00:04:22,199
interface so I'll get to that in a

106
00:04:19,260 --> 00:04:23,759
moment but um ABI is extremely

107
00:04:22,199 --> 00:04:26,580
prescriptive is what's worth knowing

108
00:04:23,759 --> 00:04:29,040
it's down to exact register allocation

109
00:04:26,580 --> 00:04:29,639
and stack layout

110
00:04:29,040 --> 00:04:32,460
um

111
00:04:29,639 --> 00:04:34,620
uh with an API you can call your library

112
00:04:32,460 --> 00:04:36,600
and as much as your tool chain has

113
00:04:34,620 --> 00:04:38,820
support for that perhaps a foreign

114
00:04:36,600 --> 00:04:41,340
function interface with a bi it

115
00:04:38,820 --> 00:04:43,979
prescribes exactly how binaries look and

116
00:04:41,340 --> 00:04:45,540
how the kernel must respond so there are

117
00:04:43,979 --> 00:04:48,000
actually a number of apis it's not

118
00:04:45,540 --> 00:04:51,060
always just user space to Kernel

119
00:04:48,000 --> 00:04:53,100
um usually when it comes to API the base

120
00:04:51,060 --> 00:04:55,500
right is a what's your object format

121
00:04:53,100 --> 00:04:57,780
which you're compiling to

122
00:04:55,500 --> 00:05:00,360
um if you're compiling code for modern

123
00:04:57,780 --> 00:05:03,360
Linux that's almost always elf the

124
00:05:00,360 --> 00:05:08,460
executable and linkable format

125
00:05:03,360 --> 00:05:10,860
on top of Elf you get further uh uh

126
00:05:08,460 --> 00:05:13,139
further specifications which come from

127
00:05:10,860 --> 00:05:15,360
your choice of architecture the machine

128
00:05:13,139 --> 00:05:18,900
which you're compiling for architecture

129
00:05:15,360 --> 00:05:20,360
is sorry ABI by the way is not ISO

130
00:05:18,900 --> 00:05:23,220
instruction set

131
00:05:20,360 --> 00:05:25,919
ABI defines things like calling

132
00:05:23,220 --> 00:05:28,259
convention for example and then even

133
00:05:25,919 --> 00:05:30,900
operating system can further

134
00:05:28,259 --> 00:05:34,620
constrain

135
00:05:30,900 --> 00:05:38,340
ABI that is how libraries and their

136
00:05:34,620 --> 00:05:40,560
consumers operate but um

137
00:05:38,340 --> 00:05:42,720
the particular level of ABI we're going

138
00:05:40,560 --> 00:05:46,139
to talk about is What's called the

139
00:05:42,720 --> 00:05:49,800
kernel to user space API and that one is

140
00:05:46,139 --> 00:05:51,780
stable intended to be stable there's

141
00:05:49,800 --> 00:05:53,820
another ABI in the kernel which is the

142
00:05:51,780 --> 00:05:56,160
kernel to modules ABI but that's not the

143
00:05:53,820 --> 00:05:58,580
one I broke so we're going to talk about

144
00:05:56,160 --> 00:05:58,580
this one

145
00:05:58,860 --> 00:06:02,759
here

146
00:06:00,000 --> 00:06:05,580
so about the Linux kernel to user space

147
00:06:02,759 --> 00:06:07,620
API this particular bit of the API is

148
00:06:05,580 --> 00:06:09,780
about the sort of tools which arbitrary

149
00:06:07,620 --> 00:06:12,180
binaries on a Linux system should be

150
00:06:09,780 --> 00:06:13,320
able to rely upon existing on their

151
00:06:12,180 --> 00:06:15,780
system

152
00:06:13,320 --> 00:06:18,360
so they include things like the vdso

153
00:06:15,780 --> 00:06:21,600
that's the virtual Dynamic shared object

154
00:06:18,360 --> 00:06:25,020
it's a library which is always mapped

155
00:06:21,600 --> 00:06:27,360
into process memory whenever you load

156
00:06:25,020 --> 00:06:30,539
any sort of Elf file

157
00:06:27,360 --> 00:06:32,280
as well as that this Cisco interface

158
00:06:30,539 --> 00:06:36,180
I'm going to talk about more about those

159
00:06:32,280 --> 00:06:39,919
in a moment the video so actually uh

160
00:06:36,180 --> 00:06:42,419
exists to circumvent some of the

161
00:06:39,919 --> 00:06:44,039
performance ramifications of CIS calls

162
00:06:42,419 --> 00:06:45,240
which almost always imply a sort of

163
00:06:44,039 --> 00:06:49,560
context switch

164
00:06:45,240 --> 00:06:52,199
and so Cisco system calls have been

165
00:06:49,560 --> 00:06:54,539
opted out so that your Lipsy can can

166
00:06:52,199 --> 00:06:56,460
choose to call things via the video so

167
00:06:54,539 --> 00:06:57,720
if appropriate

168
00:06:56,460 --> 00:07:00,060
um

169
00:06:57,720 --> 00:07:02,880
as well as that there are also file

170
00:07:00,060 --> 00:07:04,680
systems so Unix participates in a sort

171
00:07:02,880 --> 00:07:09,080
of sorry Linux participates in a sort of

172
00:07:04,680 --> 00:07:12,180
Unix file based ideology and so

173
00:07:09,080 --> 00:07:15,000
access to your csfs procfs and the

174
00:07:12,180 --> 00:07:17,759
little the little known config FS are

175
00:07:15,000 --> 00:07:20,759
also deemed to be Abi this talk is going

176
00:07:17,759 --> 00:07:24,060
to focus on ciscalls however so

177
00:07:20,759 --> 00:07:25,680
um what actually is a Cisco I think at a

178
00:07:24,060 --> 00:07:27,599
thousand feet again the way I describe

179
00:07:25,680 --> 00:07:30,419
it is a sort of pseudo function call

180
00:07:27,599 --> 00:07:33,660
that is uh when it comes to control flow

181
00:07:30,419 --> 00:07:35,580
the the same pattern exists so your your

182
00:07:33,660 --> 00:07:37,400
code will eventually yield when you

183
00:07:35,580 --> 00:07:40,380
initiate assist call

184
00:07:37,400 --> 00:07:41,639
when the syscall should terminate

185
00:07:40,380 --> 00:07:44,160
hopefully

186
00:07:41,639 --> 00:07:48,419
um and then uh execution should be

187
00:07:44,160 --> 00:07:50,280
resumed at your calling process and so

188
00:07:48,419 --> 00:07:52,440
Linux is a bit interesting I'll talk

189
00:07:50,280 --> 00:07:53,160
more about this in a moment but this is

190
00:07:52,440 --> 00:07:55,319
um

191
00:07:53,160 --> 00:07:58,199
this is the sort of code which the

192
00:07:55,319 --> 00:07:59,580
kernel is expected to support right so I

193
00:07:58,199 --> 00:08:03,060
don't know how familiar you are with

194
00:07:59,580 --> 00:08:05,099
power assembly I walk you walk uh this

195
00:08:03,060 --> 00:08:09,000
through with you

196
00:08:05,099 --> 00:08:11,520
um so uh as per the calling convention

197
00:08:09,000 --> 00:08:13,680
we're going to specify in r0 what sort

198
00:08:11,520 --> 00:08:16,319
of Cisco we want to do clearly we want

199
00:08:13,680 --> 00:08:19,860
to do an exit first parameter is zero

200
00:08:16,319 --> 00:08:21,479
exit with status zero says call

201
00:08:19,860 --> 00:08:22,740
um we have two Cisco instructions on

202
00:08:21,479 --> 00:08:24,960
power

203
00:08:22,740 --> 00:08:26,520
um your Lipsy will choose the most

204
00:08:24,960 --> 00:08:28,080
recent one

205
00:08:26,520 --> 00:08:31,319
uh

206
00:08:28,080 --> 00:08:33,659
in short this Cisco instruction is a way

207
00:08:31,319 --> 00:08:38,339
for unprivileged code

208
00:08:33,659 --> 00:08:40,339
to cause the kernel to facilitate some

209
00:08:38,339 --> 00:08:44,880
sort of critical function call

210
00:08:40,339 --> 00:08:48,000
and so nearly always the way this is

211
00:08:44,880 --> 00:08:51,180
implemented in Hardware is by means of

212
00:08:48,000 --> 00:08:52,200
an interrupt this Cisco is a synchronous

213
00:08:51,180 --> 00:08:54,480
interrupt

214
00:08:52,200 --> 00:08:56,880
in general your interrupt Hardware works

215
00:08:54,480 --> 00:08:59,100
for events which can fire at any time

216
00:08:56,880 --> 00:09:00,120
but this is an interrupt instigated by

217
00:08:59,100 --> 00:09:02,399
code

218
00:09:00,120 --> 00:09:04,019
and that interrupt Handler is expected

219
00:09:02,399 --> 00:09:07,399
to fulfill the contract of whatever

220
00:09:04,019 --> 00:09:07,399
ciscal you just called

221
00:09:07,980 --> 00:09:13,320
a really critical Cisco is fork for

222
00:09:11,220 --> 00:09:16,200
example and this is just to motivate how

223
00:09:13,320 --> 00:09:19,459
how privileged how much reflection on

224
00:09:16,200 --> 00:09:22,680
program data the assist calls end up

225
00:09:19,459 --> 00:09:24,660
invoking so this is a process in

226
00:09:22,680 --> 00:09:26,519
threading controls this call and it's

227
00:09:24,660 --> 00:09:30,000
being abstracted out by libsee so I can

228
00:09:26,519 --> 00:09:32,040
just call the fork Fork function

229
00:09:30,000 --> 00:09:34,680
um and so Fork has to look at the

230
00:09:32,040 --> 00:09:36,779
entirety of your process state

231
00:09:34,680 --> 00:09:38,580
and it what it does is it it duplicates

232
00:09:36,779 --> 00:09:40,200
that thread

233
00:09:38,580 --> 00:09:41,700
um that is it duplicates its memory

234
00:09:40,200 --> 00:09:43,500
mappings and it keeps the same

235
00:09:41,700 --> 00:09:47,399
instruction pointer in the resultant

236
00:09:43,500 --> 00:09:49,620
thread and so when the threads wake up

237
00:09:47,399 --> 00:09:51,720
they both resume execution after the

238
00:09:49,620 --> 00:09:54,060
call site and you'll notice that we take

239
00:09:51,720 --> 00:09:56,640
a return value bid

240
00:09:54,060 --> 00:09:58,800
um that's our process ID it's zero if

241
00:09:56,640 --> 00:10:01,220
we're Frankenstein's monster it's the

242
00:09:58,800 --> 00:10:03,420
name of Frankenstein's monster otherwise

243
00:10:01,220 --> 00:10:05,880
and you you branch on that condition

244
00:10:03,420 --> 00:10:08,760
figure out whether am I the child thread

245
00:10:05,880 --> 00:10:10,860
or am I or am I the parent

246
00:10:08,760 --> 00:10:12,720
um and so ciscals yeah they're a sort of

247
00:10:10,860 --> 00:10:15,300
function they or rather they have a

248
00:10:12,720 --> 00:10:17,640
function like kind of interface but

249
00:10:15,300 --> 00:10:19,680
clearly but interestingly the the

250
00:10:17,640 --> 00:10:22,260
function calling ABI and elf is not

251
00:10:19,680 --> 00:10:25,200
actually doesn't 100 line up with um

252
00:10:22,260 --> 00:10:26,880
with uh Cisco calling conventions but

253
00:10:25,200 --> 00:10:29,040
clearly this is a very privileged

254
00:10:26,880 --> 00:10:32,220
operation which I

255
00:10:29,040 --> 00:10:35,160
uh well a few more bits of a side

256
00:10:32,220 --> 00:10:37,680
uh before we cut to the chase so Linux

257
00:10:35,160 --> 00:10:41,220
is a bit interesting

258
00:10:37,680 --> 00:10:43,800
um so uh above you can correct me on the

259
00:10:41,220 --> 00:10:45,720
Nuance of this later but above other

260
00:10:43,800 --> 00:10:49,500
free and open source operating systems

261
00:10:45,720 --> 00:10:53,160
Linux has a particularly High commitment

262
00:10:49,500 --> 00:10:55,620
to keep a stable Cisco interface that's

263
00:10:53,160 --> 00:10:57,720
because you don't usually as a

264
00:10:55,620 --> 00:11:00,480
programmer invoke sys calls directly you

265
00:10:57,720 --> 00:11:01,920
might use libsy and uh Linux doesn't

266
00:11:00,480 --> 00:11:04,079
doesn't

267
00:11:01,920 --> 00:11:06,240
doesn't in principle prefer a a

268
00:11:04,079 --> 00:11:08,399
canonical lepsy implementation and so

269
00:11:06,240 --> 00:11:11,279
the Cisco interface is deemed to be it's

270
00:11:08,399 --> 00:11:13,200
deemed to be authoritative

271
00:11:11,279 --> 00:11:16,380
um certainly your proprietary operating

272
00:11:13,200 --> 00:11:19,019
systems avoid syscalls at all for user

273
00:11:16,380 --> 00:11:20,760
space so there's lib system on on Mac OS

274
00:11:19,019 --> 00:11:21,899
Cisco interface is not deemed to be

275
00:11:20,760 --> 00:11:25,860
stable

276
00:11:21,899 --> 00:11:28,440
um and then you have nt.dll on Windows

277
00:11:25,860 --> 00:11:30,360
um but yeah Linux has this kind of

278
00:11:28,440 --> 00:11:32,760
extreme attention to detail when it

279
00:11:30,360 --> 00:11:34,860
comes to ABI it's best sum up summed up

280
00:11:32,760 --> 00:11:37,200
by this one particular Linus quote which

281
00:11:34,860 --> 00:11:39,959
is of course in all caps we do not break

282
00:11:37,200 --> 00:11:42,680
user space and uh so that's what I did

283
00:11:39,959 --> 00:11:46,440
all right well

284
00:11:42,680 --> 00:11:48,660
all right um so uh so we have our

285
00:11:46,440 --> 00:11:50,700
interrupt code and um because the kernel

286
00:11:48,660 --> 00:11:53,760
is essentially just this stitch together

287
00:11:50,700 --> 00:11:56,240
big old massive elf binary stitched

288
00:11:53,760 --> 00:11:59,880
together by link scripts

289
00:11:56,240 --> 00:12:02,220
we have these symbols in here for the

290
00:11:59,880 --> 00:12:04,560
actual body of a Cisco our interrupt

291
00:12:02,220 --> 00:12:07,620
code is going to call this body their

292
00:12:04,560 --> 00:12:10,440
inter interrupts are handled by kernel

293
00:12:07,620 --> 00:12:13,440
um and so they used to have uh this sort

294
00:12:10,440 --> 00:12:15,839
of calling convention however to give a

295
00:12:13,440 --> 00:12:20,820
little bit more freedom to how we write

296
00:12:15,839 --> 00:12:23,060
our interrupt code we decided to use

297
00:12:20,820 --> 00:12:25,440
some ugly and horrible macros

298
00:12:23,060 --> 00:12:27,300
to change the convo calling convention

299
00:12:25,440 --> 00:12:30,440
to look like this you don't need to know

300
00:12:27,300 --> 00:12:32,579
what a PT regs is it comes from p-trace

301
00:12:30,440 --> 00:12:33,720
I can't tell you any more of the

302
00:12:32,579 --> 00:12:35,519
entomology

303
00:12:33,720 --> 00:12:36,839
etymology Dragon

304
00:12:35,519 --> 00:12:39,240
um there

305
00:12:36,839 --> 00:12:41,279
um it's it's actual in actual fact just

306
00:12:39,240 --> 00:12:44,700
a structure which represents the

307
00:12:41,279 --> 00:12:46,380
registers of the caller of the Cisco

308
00:12:44,700 --> 00:12:48,779
because it's an interrupt right so we're

309
00:12:46,380 --> 00:12:51,540
we're there's a entirely new context

310
00:12:48,779 --> 00:12:54,300
between user user space code which

311
00:12:51,540 --> 00:12:56,579
called us and kernel code we keep that

312
00:12:54,300 --> 00:12:58,320
context and we're expected to save and

313
00:12:56,579 --> 00:13:00,560
restore it accordingly hence a context

314
00:12:58,320 --> 00:13:00,560
switch

315
00:13:00,899 --> 00:13:04,380
um and so we want to change that calling

316
00:13:02,279 --> 00:13:07,380
convention uh

317
00:13:04,380 --> 00:13:10,380
stay cleared for a bit of motivation on

318
00:13:07,380 --> 00:13:12,839
why we do that exactly but the the kind

319
00:13:10,380 --> 00:13:15,180
of crown jewels is uh I've been inducted

320
00:13:12,839 --> 00:13:18,120
onto the hardening team and uh so we

321
00:13:15,180 --> 00:13:19,860
decided we'd do a hardening and thought

322
00:13:18,120 --> 00:13:24,300
of that included

323
00:13:19,860 --> 00:13:26,880
um uh we wanted to essentially uh

324
00:13:24,300 --> 00:13:27,540
sanitize whatever

325
00:13:26,880 --> 00:13:29,899
um

326
00:13:27,540 --> 00:13:33,000
uh whatever continuity there is between

327
00:13:29,899 --> 00:13:36,480
user space code and kernel code and so

328
00:13:33,000 --> 00:13:39,060
whenever user space can directly or

329
00:13:36,480 --> 00:13:42,180
directly enough in invoke an interrupt

330
00:13:39,060 --> 00:13:44,339
of any sort we don't want the

331
00:13:42,180 --> 00:13:46,500
architected register state of that user

332
00:13:44,339 --> 00:13:48,300
code in any way influencing the

333
00:13:46,500 --> 00:13:50,459
speculative runtime behavior of the

334
00:13:48,300 --> 00:13:52,380
kernel and so what do we do we have to

335
00:13:50,459 --> 00:13:53,480
save even more registers for this

336
00:13:52,380 --> 00:13:56,639
mitigation

337
00:13:53,480 --> 00:13:59,639
we sanitize those registers we set them

338
00:13:56,639 --> 00:14:01,560
to zero and then we restore those

339
00:13:59,639 --> 00:14:04,260
registers before we return

340
00:14:01,560 --> 00:14:05,639
and if it's not obvious to you that

341
00:14:04,260 --> 00:14:09,540
there's an exploit kind of in the

342
00:14:05,639 --> 00:14:12,180
waiting here then um that's fine it

343
00:14:09,540 --> 00:14:14,579
wasn't obvious to me either

344
00:14:12,180 --> 00:14:17,160
um but let's just sort of dive into the

345
00:14:14,579 --> 00:14:20,279
activity of Kernel hardening right so

346
00:14:17,160 --> 00:14:23,399
there's been a sort of seismic shift in

347
00:14:20,279 --> 00:14:25,500
the world of security research with

348
00:14:23,399 --> 00:14:26,459
Spectra and meltdown which is to say

349
00:14:25,500 --> 00:14:29,720
that

350
00:14:26,459 --> 00:14:33,000
um user space may not be able to elect

351
00:14:29,720 --> 00:14:36,120
leak secrets from the kernel directly

352
00:14:33,000 --> 00:14:39,300
but it might be able to do so in an

353
00:14:36,120 --> 00:14:41,220
implied fashion through timing attacks

354
00:14:39,300 --> 00:14:43,199
and these timing attacks become

355
00:14:41,220 --> 00:14:46,160
increasingly available with an awareness

356
00:14:43,199 --> 00:14:48,120
of of the reality which is that

357
00:14:46,160 --> 00:14:51,420
microarchitectures are necessarily

358
00:14:48,120 --> 00:14:53,160
speculative on on super scalar processes

359
00:14:51,420 --> 00:14:55,380
and um

360
00:14:53,160 --> 00:14:56,820
and so what we're doing is uh is

361
00:14:55,380 --> 00:14:59,160
hardening measures right we're just

362
00:14:56,820 --> 00:15:01,440
trying to produce isolation between user

363
00:14:59,160 --> 00:15:03,540
code and kernel code and so I can't

364
00:15:01,440 --> 00:15:06,300
actually guarantee you that I I've

365
00:15:03,540 --> 00:15:09,600
foreseen exploits per se

366
00:15:06,300 --> 00:15:11,220
but um it's a bit tedious to just chase

367
00:15:09,600 --> 00:15:13,680
after exploits when they become

368
00:15:11,220 --> 00:15:15,660
discovered right you'd rather Harden the

369
00:15:13,680 --> 00:15:18,000
kernel against against the prospect if

370
00:15:15,660 --> 00:15:21,240
people gluing together info leaks and

371
00:15:18,000 --> 00:15:24,420
gadgets and one day creating an exploit

372
00:15:21,240 --> 00:15:26,459
and so a bit of a bit of a of a red flag

373
00:15:24,420 --> 00:15:28,139
which alerts you to to the fact that

374
00:15:26,459 --> 00:15:31,920
this might need hardening is that there

375
00:15:28,139 --> 00:15:37,260
is shared architect at State uh that is

376
00:15:31,920 --> 00:15:40,620
the registers of the calling code are

377
00:15:37,260 --> 00:15:43,620
um are in fact retained although cleared

378
00:15:40,620 --> 00:15:45,959
pretty quickly by by kernel code

379
00:15:43,620 --> 00:15:47,519
if um by the way if there's a if there's

380
00:15:45,959 --> 00:15:50,160
an architectural bug we've got bigger

381
00:15:47,519 --> 00:15:52,500
problems so uh we're on the micro

382
00:15:50,160 --> 00:15:55,440
architectural level here

383
00:15:52,500 --> 00:15:57,060
um okay so a quick aside as to how Cisco

384
00:15:55,440 --> 00:15:59,820
wrappers are implemented in the kernel I

385
00:15:57,060 --> 00:16:01,980
promise to not show you too much code

386
00:15:59,820 --> 00:16:07,139
um so there's this Cisco Define for

387
00:16:01,980 --> 00:16:09,899
macro and so this kind of creates the uh

388
00:16:07,139 --> 00:16:12,480
signature the the Prototype what's it

389
00:16:09,899 --> 00:16:16,620
called and see the the prototype for

390
00:16:12,480 --> 00:16:18,839
um uh for the phallicate F allocates uh

391
00:16:16,620 --> 00:16:20,820
Cisco Handler

392
00:16:18,839 --> 00:16:22,440
um and it's going to prepare this CIS

393
00:16:20,820 --> 00:16:23,760
underscore prefix to the beginning

394
00:16:22,440 --> 00:16:25,800
because no one should be directly

395
00:16:23,760 --> 00:16:27,839
calling it by name

396
00:16:25,800 --> 00:16:30,000
um and so if you're going to create your

397
00:16:27,839 --> 00:16:33,060
own Cisco if you're mucking about the

398
00:16:30,000 --> 00:16:35,940
kernel you'll Define this is called like

399
00:16:33,060 --> 00:16:39,060
such and then you'll add a number in a

400
00:16:35,940 --> 00:16:41,820
in a table to say when r0 is this number

401
00:16:39,060 --> 00:16:42,839
in kernel please call this so it's cool

402
00:16:41,820 --> 00:16:45,540
for me

403
00:16:42,839 --> 00:16:48,060
what we're going to do is uh in order to

404
00:16:45,540 --> 00:16:51,120
get a new calling convention see Cisco

405
00:16:48,060 --> 00:16:53,579
handlers uh like any good C program or

406
00:16:51,120 --> 00:16:55,560
would we're going to write an ugly macro

407
00:16:53,579 --> 00:16:56,940
um and uh that ugly macro better

408
00:16:55,560 --> 00:17:01,440
implement the calling convention

409
00:16:56,940 --> 00:17:03,660
perfectly with no bugs ever for

410
00:17:01,440 --> 00:17:06,900
our Cisco Handler and so clearly this is

411
00:17:03,660 --> 00:17:09,540
this is where my bug came from

412
00:17:06,900 --> 00:17:12,079
um as a bit of an aside so I work on on

413
00:17:09,540 --> 00:17:15,780
Power Systems Linux on power

414
00:17:12,079 --> 00:17:18,419
we have our own subsystem arch power PC

415
00:17:15,780 --> 00:17:21,000
and uh there are about 12 million lines

416
00:17:18,419 --> 00:17:23,040
of code in the kernel uh not including

417
00:17:21,000 --> 00:17:26,459
modules not including user space

418
00:17:23,040 --> 00:17:29,460
libraries tests Etc about four million

419
00:17:26,459 --> 00:17:31,679
of those lines is architecture specific

420
00:17:29,460 --> 00:17:35,760
and about half a million of those lines

421
00:17:31,679 --> 00:17:37,200
sit in Arch power PC so it's a big

422
00:17:35,760 --> 00:17:39,660
subsystem

423
00:17:37,200 --> 00:17:41,840
why is it so big well

424
00:17:39,660 --> 00:17:45,419
you could hint to a few things I suppose

425
00:17:41,840 --> 00:17:48,900
for comparison arm has spun off its

426
00:17:45,419 --> 00:17:51,419
32-bit and 64-bit subsystems

427
00:17:48,900 --> 00:17:53,760
um and so we're like x86 in that respect

428
00:17:51,419 --> 00:17:57,120
we they're both under a common

429
00:17:53,760 --> 00:17:58,799
subdirectory but unlike x86 where by

430
00:17:57,120 --> 00:18:00,419
Indian

431
00:17:58,799 --> 00:18:02,760
so you end up with this kind of horrible

432
00:18:00,419 --> 00:18:04,320
Matrix of kind of platforms running on

433
00:18:02,760 --> 00:18:07,140
Arch powerpc

434
00:18:04,320 --> 00:18:09,140
and sometimes the same code Services all

435
00:18:07,140 --> 00:18:12,740
four of these platforms

436
00:18:09,140 --> 00:18:15,660
Ibis are fiddly enough and so with four

437
00:18:12,740 --> 00:18:17,220
sub-architectures of sorts uh yeah

438
00:18:15,660 --> 00:18:20,160
necessarily there there are going to be

439
00:18:17,220 --> 00:18:22,080
at least four abis I need to implement

440
00:18:20,160 --> 00:18:23,460
an API is Right are meant to be

441
00:18:22,080 --> 00:18:28,160
implemented by Tool chains because

442
00:18:23,460 --> 00:18:28,160
they're fiddly and not buy crappy macros

443
00:18:28,200 --> 00:18:33,240
so how did I do in my patch series my

444
00:18:30,900 --> 00:18:34,740
first patch series on the Kernel

445
00:18:33,240 --> 00:18:37,320
um well there's a little babby kernel

446
00:18:34,740 --> 00:18:39,419
developer I had to touch a lot of

447
00:18:37,320 --> 00:18:42,360
um untouched Legacy code

448
00:18:39,419 --> 00:18:45,059
and uh I started off with six patches my

449
00:18:42,360 --> 00:18:47,340
first revision and uh if you don't know

450
00:18:45,059 --> 00:18:51,059
in the kernel we um

451
00:18:47,340 --> 00:18:53,900
we do our communication over email which

452
00:18:51,059 --> 00:18:57,660
is insane to me

453
00:18:53,900 --> 00:19:00,179
and uh once you send off saying hi I'm a

454
00:18:57,660 --> 00:19:01,860
young and not yet cynical kernel

455
00:19:00,179 --> 00:19:03,720
developer and I would like to modernize

456
00:19:01,860 --> 00:19:06,480
your stuff you get a laundry list right

457
00:19:03,720 --> 00:19:08,760
of things which you should fix and so uh

458
00:19:06,480 --> 00:19:12,419
by the time I I submitted part of my

459
00:19:08,760 --> 00:19:14,640
patches it was revision six and uh I had

460
00:19:12,419 --> 00:19:17,640
25 patches on the list which was

461
00:19:14,640 --> 00:19:20,039
ridiculous most of them got into a

462
00:19:17,640 --> 00:19:21,960
kernel six one but the actual hardening

463
00:19:20,039 --> 00:19:23,520
component got into six two because again

464
00:19:21,960 --> 00:19:25,860
if you went to Russell's talk everyone

465
00:19:23,520 --> 00:19:29,460
will complain if you make your CPS like

466
00:19:25,860 --> 00:19:30,059
two percent slower so

467
00:19:29,460 --> 00:19:32,760
um

468
00:19:30,059 --> 00:19:34,980
yeah like where I really

469
00:19:32,760 --> 00:19:37,860
um hit a wall is with this 32-bit

470
00:19:34,980 --> 00:19:39,900
calling conventions I I really

471
00:19:37,860 --> 00:19:42,600
um did work hard to to make these work

472
00:19:39,900 --> 00:19:44,220
did I miss the slide no all right I

473
00:19:42,600 --> 00:19:47,160
really did work hard to make these work

474
00:19:44,220 --> 00:19:50,580
but there are just some cursed

475
00:19:47,160 --> 00:19:53,820
um uh special cases so here's a here's a

476
00:19:50,580 --> 00:19:57,120
relatively normal ciscal CIS F advise

477
00:19:53,820 --> 00:20:00,240
64. where does the 64 come from well

478
00:19:57,120 --> 00:20:02,700
this lofty you kind of you know if you

479
00:20:00,240 --> 00:20:05,760
know that's always 64-bit irrespective

480
00:20:02,700 --> 00:20:08,160
of of what tool chain you're using

481
00:20:05,760 --> 00:20:09,480
um I think that's a posix thing

482
00:20:08,160 --> 00:20:11,940
um and uh

483
00:20:09,480 --> 00:20:15,620
yeah well some architectures actually

484
00:20:11,940 --> 00:20:15,620
want you to implement this one too

485
00:20:16,260 --> 00:20:19,520
um this is this is where it gets bit

486
00:20:17,700 --> 00:20:21,960
cursed you notice that the advice

487
00:20:19,520 --> 00:20:24,480
parameter actually changes its position

488
00:20:21,960 --> 00:20:27,419
this is probably the most attention

489
00:20:24,480 --> 00:20:28,440
you've ever paid to this fact but um the

490
00:20:27,419 --> 00:20:30,960
reason why it does that is because

491
00:20:28,440 --> 00:20:33,419
there's a sort of soft rule of no more

492
00:20:30,960 --> 00:20:35,460
than six parameters to ciscals because

493
00:20:33,419 --> 00:20:38,039
calling conventions on all these

494
00:20:35,460 --> 00:20:39,600
architectures do register bound so it's

495
00:20:38,039 --> 00:20:41,100
called most of them architectures do

496
00:20:39,600 --> 00:20:42,720
registered Bound syscalls for the first

497
00:20:41,100 --> 00:20:45,480
six at least

498
00:20:42,720 --> 00:20:46,919
um and so that's it's kind of common

499
00:20:45,480 --> 00:20:50,160
um and uh

500
00:20:46,919 --> 00:20:53,580
see 32-bit machines still need to be

501
00:20:50,160 --> 00:20:55,679
able to send these uh these two 64-bit

502
00:20:53,580 --> 00:20:57,299
params that's the the lofties I'm

503
00:20:55,679 --> 00:20:59,820
pointing to

504
00:20:57,299 --> 00:21:02,760
how do you sand a 64-bit parameter with

505
00:20:59,820 --> 00:21:05,280
32-bit arguments uh registers well you

506
00:21:02,760 --> 00:21:07,620
obviously split it in two and so that

507
00:21:05,280 --> 00:21:09,480
comes to six and why did we have to move

508
00:21:07,620 --> 00:21:12,360
advice to the back

509
00:21:09,480 --> 00:21:14,179
well it's because these architectures in

510
00:21:12,360 --> 00:21:16,679
their function calling conventions

511
00:21:14,179 --> 00:21:19,679
specify that these registered pairs

512
00:21:16,679 --> 00:21:21,900
which unite to become a single 64-bit

513
00:21:19,679 --> 00:21:24,299
values is register pairs better be odd

514
00:21:21,900 --> 00:21:26,600
even in their in their indexing all

515
00:21:24,299 --> 00:21:26,600
right so

516
00:21:27,360 --> 00:21:32,940
um yeah so some 32-bit args kind of

517
00:21:30,000 --> 00:21:34,620
expect the the Sysco handlers to look a

518
00:21:32,940 --> 00:21:38,039
bit like that

519
00:21:34,620 --> 00:21:39,480
and then the order of uh Lo and hi in

520
00:21:38,039 --> 00:21:41,039
each case is going to be dependent on

521
00:21:39,480 --> 00:21:41,760
endiness

522
00:21:41,039 --> 00:21:44,340
um

523
00:21:41,760 --> 00:21:46,440
yeah well all of these are implemented

524
00:21:44,340 --> 00:21:48,539
not in generic code they're implemented

525
00:21:46,440 --> 00:21:51,059
in power PC and they were last edited in

526
00:21:48,539 --> 00:21:52,799
2005. so so I had to clean up a whole

527
00:21:51,059 --> 00:21:55,620
bunch of them they weren't even using

528
00:21:52,799 --> 00:21:57,360
the macro I showed earlier so yeah a lot

529
00:21:55,620 --> 00:21:59,280
of fix up I won't believe with the

530
00:21:57,360 --> 00:22:01,980
details and then there was this little

531
00:21:59,280 --> 00:22:04,440
gem all right well so it turns out

532
00:22:01,980 --> 00:22:07,020
because this is a Cisco right and we

533
00:22:04,440 --> 00:22:08,580
don't break user space this was ABI for

534
00:22:07,020 --> 00:22:10,679
a while this was this was

535
00:22:08,580 --> 00:22:14,220
I couldn't change this without much

536
00:22:10,679 --> 00:22:17,400
arguing so in short there was this crazy

537
00:22:14,220 --> 00:22:19,919
um like power PC 32-bit only uh Cisco

538
00:22:17,400 --> 00:22:21,720
where um it would check whether the

539
00:22:19,919 --> 00:22:23,640
first parameter is a pointer or Not by

540
00:22:21,720 --> 00:22:27,059
saying is it bigger than a prohibitively

541
00:22:23,640 --> 00:22:29,220
large number of file descriptors and um

542
00:22:27,059 --> 00:22:32,220
if so it would arbitrarily decide use

543
00:22:29,220 --> 00:22:36,179
these old semantics if not use these new

544
00:22:32,220 --> 00:22:37,799
ones and that's nuts so I after quite a

545
00:22:36,179 --> 00:22:39,120
bit of arguing and then resuscitating

546
00:22:37,799 --> 00:22:40,980
all the conversations on the mailing

547
00:22:39,120 --> 00:22:42,480
list and saying no look we really were

548
00:22:40,980 --> 00:22:44,159
going in this direction just people

549
00:22:42,480 --> 00:22:45,840
didn't have time

550
00:22:44,159 --> 00:22:48,659
um we should just replace this with

551
00:22:45,840 --> 00:22:51,000
something sensible and so I

552
00:22:48,659 --> 00:22:52,880
finally cut out this thing and replaced

553
00:22:51,000 --> 00:22:55,980
it with say select

554
00:22:52,880 --> 00:22:58,260
besides the way I'd set things up you

555
00:22:55,980 --> 00:23:00,000
couldn't reliably recurse this call

556
00:22:58,260 --> 00:23:03,360
handlers so this kind of had to go

557
00:23:00,000 --> 00:23:04,919
anyway details

558
00:23:03,360 --> 00:23:07,980
all right so how did we do with

559
00:23:04,919 --> 00:23:10,679
performance right so we're having to

560
00:23:07,980 --> 00:23:13,260
save more registers in the way and that

561
00:23:10,679 --> 00:23:14,539
we hope to sanitize all of the users

562
00:23:13,260 --> 00:23:17,820
registers

563
00:23:14,539 --> 00:23:19,919
and then eventually restore them and so

564
00:23:17,820 --> 00:23:23,220
saving to the stack restoring from the

565
00:23:19,919 --> 00:23:26,400
stack it's going to be hella slow right

566
00:23:23,220 --> 00:23:28,620
oh sorry sanitizing more registers

567
00:23:26,400 --> 00:23:30,080
clearly that's more work on every

568
00:23:28,620 --> 00:23:34,260
contact switch

569
00:23:30,080 --> 00:23:35,640
and it turns out it made things faster

570
00:23:34,260 --> 00:23:36,900
um I made context which is faster

571
00:23:35,640 --> 00:23:40,440
accidentally

572
00:23:36,900 --> 00:23:42,120
um and so I've hence been summarily um

573
00:23:40,440 --> 00:23:44,940
removed from the hardening team because

574
00:23:42,120 --> 00:23:46,860
that goes against our our secondary goal

575
00:23:44,940 --> 00:23:50,039
primary goal

576
00:23:46,860 --> 00:23:52,500
just making context switches slower

577
00:23:50,039 --> 00:23:54,780
um it turns out that I made them 14

578
00:23:52,500 --> 00:23:56,460
faster you can ask me in the questions

579
00:23:54,780 --> 00:23:59,100
how

580
00:23:56,460 --> 00:24:01,559
uh and uh even with the mitigation

581
00:23:59,100 --> 00:24:04,799
kicked in contact switches are still 5.6

582
00:24:01,559 --> 00:24:07,380
faster so that's nice

583
00:24:04,799 --> 00:24:09,840
all right so you didn't come here to

584
00:24:07,380 --> 00:24:12,179
hear me waffle about cescals you came to

585
00:24:09,840 --> 00:24:13,200
see me break things so

586
00:24:12,179 --> 00:24:16,140
um

587
00:24:13,200 --> 00:24:19,400
right well this is the feedback I got

588
00:24:16,140 --> 00:24:21,360
and um after a little bit of a frantic

589
00:24:19,400 --> 00:24:23,340
slurry of trying to figure out what was

590
00:24:21,360 --> 00:24:25,620
going wrong this is what I figured out

591
00:24:23,340 --> 00:24:27,720
what my macro was doing if you

592
00:24:25,620 --> 00:24:29,159
um don't like debugging macros then uh

593
00:24:27,720 --> 00:24:32,460
good

594
00:24:29,159 --> 00:24:34,200
um so so I my macro spits out these

595
00:24:32,460 --> 00:24:35,940
three symbols

596
00:24:34,200 --> 00:24:38,760
um just in various stages of trying to

597
00:24:35,940 --> 00:24:40,620
unpack this args parameter

598
00:24:38,760 --> 00:24:42,000
and uh

599
00:24:40,620 --> 00:24:44,039
you don't need to see all of them

600
00:24:42,000 --> 00:24:45,120
because uh right here you can probably

601
00:24:44,039 --> 00:24:47,700
see the problem

602
00:24:45,120 --> 00:24:49,559
if you're facilitating a 32-bit calling

603
00:24:47,700 --> 00:24:52,080
convention by the way you should be able

604
00:24:49,559 --> 00:24:54,600
to run 32-bit programs on a 64-bit

605
00:24:52,080 --> 00:24:55,440
machine there needs to be compatibility

606
00:24:54,600 --> 00:24:58,980
um

607
00:24:55,440 --> 00:25:00,480
if you're running 32-bit my macro

608
00:24:58,980 --> 00:25:02,900
basically

609
00:25:00,480 --> 00:25:07,740
it cuts out here it sees that this is a

610
00:25:02,900 --> 00:25:09,659
uh for arity Sysco and so it only emits

611
00:25:07,740 --> 00:25:11,580
the first four registers so it just it

612
00:25:09,659 --> 00:25:13,740
breaks calling convention and I knew

613
00:25:11,580 --> 00:25:17,640
about this right I just I just didn't

614
00:25:13,740 --> 00:25:20,240
know that there were any ciscals left in

615
00:25:17,640 --> 00:25:23,039
the kernel after I had done all my work

616
00:25:20,240 --> 00:25:24,840
that were expected to run on both 32 and

617
00:25:23,039 --> 00:25:27,720
64.

618
00:25:24,840 --> 00:25:31,080
and so the rest of this is all a wash

619
00:25:27,720 --> 00:25:33,480
um but yeah that's some cursed macro

620
00:25:31,080 --> 00:25:34,860
and uh yeah just to kind of spell it out

621
00:25:33,480 --> 00:25:39,240
a little bit further

622
00:25:34,860 --> 00:25:41,460
uh we only pass in four registers uh the

623
00:25:39,240 --> 00:25:43,140
the last two are five and R6 should be

624
00:25:41,460 --> 00:25:44,820
parts of offset

625
00:25:43,140 --> 00:25:48,059
of the offset parameter but instead

626
00:25:44,820 --> 00:25:51,120
they're interpreted as a random bit of

627
00:25:48,059 --> 00:25:53,880
offset and around a bit of Len and so

628
00:25:51,120 --> 00:25:55,919
this this is what the macro should spit

629
00:25:53,880 --> 00:25:59,760
out but it doesn't

630
00:25:55,919 --> 00:26:02,340
which is not good and uh yeah so the the

631
00:25:59,760 --> 00:26:07,020
exact part of offset and Lan which are

632
00:26:02,340 --> 00:26:10,799
represented by each each of R5 and R6 uh

633
00:26:07,020 --> 00:26:12,840
basically that's sort of random but um

634
00:26:10,799 --> 00:26:15,960
but not quite random that's determined

635
00:26:12,840 --> 00:26:19,080
by our endian-ness which brings us to

636
00:26:15,960 --> 00:26:20,940
the next bug I wrote

637
00:26:19,080 --> 00:26:22,679
um and uh there's nothing really deep

638
00:26:20,940 --> 00:26:25,260
here but I thought I would give back to

639
00:26:22,679 --> 00:26:30,480
the community and um I would Define this

640
00:26:25,260 --> 00:26:32,100
compat ARG u64 dual macro there's a

641
00:26:30,480 --> 00:26:34,320
subsystem in the kernel you may be

642
00:26:32,100 --> 00:26:38,760
interested to know called ASM generic

643
00:26:34,320 --> 00:26:41,820
and uh azim Sans stands for assembly and

644
00:26:38,760 --> 00:26:45,539
it's you know it's like generic assembly

645
00:26:41,820 --> 00:26:47,700
um essentially it's bits of macros

646
00:26:45,539 --> 00:26:49,919
um which uh all architectures could

647
00:26:47,700 --> 00:26:52,320
benefit from there being common code and

648
00:26:49,919 --> 00:26:54,240
people override their own macros and so

649
00:26:52,320 --> 00:26:55,679
I overload these macros

650
00:26:54,240 --> 00:26:58,260
so that everyone can have a

651
00:26:55,679 --> 00:27:00,600
compatibility version of f allocate

652
00:26:58,260 --> 00:27:04,860
everyone who chooses to

653
00:27:00,600 --> 00:27:07,559
uh and that compat are q64 Jewel that's

654
00:27:04,860 --> 00:27:09,960
going to split offset into a low portion

655
00:27:07,559 --> 00:27:12,000
and a high portion and then the order of

656
00:27:09,960 --> 00:27:14,580
those low portion and high portion is

657
00:27:12,000 --> 00:27:16,380
going to be determined by endiness well

658
00:27:14,580 --> 00:27:17,880
I got that wrong I got that flipped

659
00:27:16,380 --> 00:27:19,440
around the other side and my patch

660
00:27:17,880 --> 00:27:21,299
series went through six revisions and

661
00:27:19,440 --> 00:27:22,020
somehow no one caught that

662
00:27:21,299 --> 00:27:23,820
um

663
00:27:22,020 --> 00:27:26,159
so so I really was just hitting high

664
00:27:23,820 --> 00:27:30,360
schoolers here and I had in fact bricked

665
00:27:26,159 --> 00:27:31,740
risk five 32-bit and power PC 32-bit at

666
00:27:30,360 --> 00:27:34,200
this stage

667
00:27:31,740 --> 00:27:35,159
um a prominent member of the colonel

668
00:27:34,200 --> 00:27:37,020
community

669
00:27:35,159 --> 00:27:39,240
made a comment about this see the

670
00:27:37,020 --> 00:27:41,220
blunders which I'd made right they were

671
00:27:39,240 --> 00:27:43,919
they were great

672
00:27:41,220 --> 00:27:46,559
um in nature but not great enough to

673
00:27:43,919 --> 00:27:48,120
topple two-haul architectures because um

674
00:27:46,559 --> 00:27:50,580
no one's really running these Upstream

675
00:27:48,120 --> 00:27:53,580
kernels but a certain prominent member

676
00:27:50,580 --> 00:27:56,159
of the community decided to chime in

677
00:27:53,580 --> 00:27:58,080
saying you know is it just me or did

678
00:27:56,159 --> 00:28:00,179
this break every architecture did did

679
00:27:58,080 --> 00:28:01,440
this grad just manage to destroy

680
00:28:00,179 --> 00:28:03,240
everything

681
00:28:01,440 --> 00:28:06,059
um and people just didn't scream because

682
00:28:03,240 --> 00:28:08,159
32-bit code has become so rare

683
00:28:06,059 --> 00:28:11,220
um thankfully it was it was just risk

684
00:28:08,159 --> 00:28:14,340
five and powerpc so no one noticed but

685
00:28:11,220 --> 00:28:14,340
um but

686
00:28:15,000 --> 00:28:17,120
um

687
00:28:17,159 --> 00:28:20,880
but you know every cloud has a silver

688
00:28:19,380 --> 00:28:23,539
lining and a guy I got my first shout

689
00:28:20,880 --> 00:28:23,539
out from Linus

690
00:28:25,679 --> 00:28:30,900
so uh by means of conclusion essentially

691
00:28:28,440 --> 00:28:34,440
yeah I'm a first year grad

692
00:28:30,900 --> 00:28:36,840
um I go to work on the Linux kernel

693
00:28:34,440 --> 00:28:38,940
which supports the vast majority of of

694
00:28:36,840 --> 00:28:40,679
web servers which is kind of amazing and

695
00:28:38,940 --> 00:28:42,900
it's just kind of amazing how how easily

696
00:28:40,679 --> 00:28:45,360
someone can throw their hat into the

697
00:28:42,900 --> 00:28:48,600
ring and work on

698
00:28:45,360 --> 00:28:51,000
um essentially just infrastructure which

699
00:28:48,600 --> 00:28:53,580
which everyone depends upon

700
00:28:51,000 --> 00:28:56,100
and um essentially my ability to

701
00:28:53,580 --> 00:28:59,940
contribute as much as I have is really

702
00:28:56,100 --> 00:29:01,860
owed greatly to the rigorous uh

703
00:28:59,940 --> 00:29:04,980
reviewing this is not a joke by the way

704
00:29:01,860 --> 00:29:08,580
these these these tests uh the people

705
00:29:04,980 --> 00:29:09,779
really did boot kernels on 32-bit and uh

706
00:29:08,580 --> 00:29:12,480
it turns out they they didn't actually

707
00:29:09,779 --> 00:29:14,700
corrupt enough to not boot

708
00:29:12,480 --> 00:29:16,500
um which is delightful so it was one of

709
00:29:14,700 --> 00:29:18,419
those kind of wicked problems

710
00:29:16,500 --> 00:29:21,059
um but they certainly were corrupting

711
00:29:18,419 --> 00:29:24,120
um and really anyway my ability to get

712
00:29:21,059 --> 00:29:26,159
this far is very much I'm thankful for

713
00:29:24,120 --> 00:29:28,080
for the testers the reviewers and the

714
00:29:26,159 --> 00:29:30,840
maintainers who makes this possible

715
00:29:28,080 --> 00:29:32,580
essentially what I want to say is that

716
00:29:30,840 --> 00:29:34,980
persistence seems to be a sort of key

717
00:29:32,580 --> 00:29:36,539
skill for kernel developers

718
00:29:34,980 --> 00:29:39,000
um and that if you if you're someone who

719
00:29:36,539 --> 00:29:41,820
enjoys low level programming and you're

720
00:29:39,000 --> 00:29:44,039
willing to go through a mailing list to

721
00:29:41,820 --> 00:29:45,899
go through multiple lines of reviews you

722
00:29:44,039 --> 00:29:48,360
you really can actually contribute to

723
00:29:45,899 --> 00:29:50,340
some pretty significant software

724
00:29:48,360 --> 00:29:52,740
projects

725
00:29:50,340 --> 00:29:55,399
and uh there's always work it seems

726
00:29:52,740 --> 00:29:57,480
manually manipulating registers

727
00:29:55,399 --> 00:30:00,360
and but if you have this sort of

728
00:29:57,480 --> 00:30:02,340
meticulous attention to detail yeah you

729
00:30:00,360 --> 00:30:03,179
can you can contribute on most things so

730
00:30:02,340 --> 00:30:05,700
I hope

731
00:30:03,179 --> 00:30:07,799
in hearing this talk uh this encourages

732
00:30:05,700 --> 00:30:09,539
you to contribute to the projects which

733
00:30:07,799 --> 00:30:10,440
interest you

734
00:30:09,539 --> 00:30:12,000
um

735
00:30:10,440 --> 00:30:15,260
and I hope that your first patch series

736
00:30:12,000 --> 00:30:15,260
is not quite this Coast

737
00:30:15,419 --> 00:30:18,790
um

738
00:30:16,200 --> 00:30:26,820
any questions yep

739
00:30:18,790 --> 00:30:28,799
[Applause]

740
00:30:26,820 --> 00:30:33,299
so how did you speed up context

741
00:30:28,799 --> 00:30:34,919
yeah yeah good one so uh should I show

742
00:30:33,299 --> 00:30:37,559
some kernel code for this one that might

743
00:30:34,919 --> 00:30:42,419
take a little a little while go ahead

744
00:30:37,559 --> 00:30:46,140
okay uh all right I closed my editor

745
00:30:42,419 --> 00:30:47,460
um okay well in in short

746
00:30:46,140 --> 00:30:49,860
uh

747
00:30:47,460 --> 00:30:51,840
there's been a number of reforms to

748
00:30:49,860 --> 00:30:55,380
replace what is a whole bunch of

749
00:30:51,840 --> 00:30:57,320
assembly so we we we use uh gnu

750
00:30:55,380 --> 00:30:59,820
assembler to

751
00:30:57,320 --> 00:31:01,860
write our interrupt handlers at the

752
00:30:59,820 --> 00:31:04,980
exact region in virtual memory where

753
00:31:01,860 --> 00:31:06,600
intra handlers are expected and then we

754
00:31:04,980 --> 00:31:08,460
just have this we have a number of

755
00:31:06,600 --> 00:31:10,260
trampolines which eventually call into C

756
00:31:08,460 --> 00:31:13,620
code and then the rest of everything is

757
00:31:10,260 --> 00:31:15,659
performed in in C code in short when I

758
00:31:13,620 --> 00:31:16,799
changed calling conventions I know what

759
00:31:15,659 --> 00:31:20,720
I'll do

760
00:31:16,799 --> 00:31:20,720
I'll just show this slide

761
00:31:21,419 --> 00:31:25,260
yeah so when I change calling

762
00:31:23,220 --> 00:31:27,779
conventions

763
00:31:25,260 --> 00:31:29,580
for the assist call handlers I was also

764
00:31:27,779 --> 00:31:32,760
able to call it change the calling

765
00:31:29,580 --> 00:31:34,500
Convention of some of these uh C

766
00:31:32,760 --> 00:31:35,760
functions which facilitate these

767
00:31:34,500 --> 00:31:38,520
interrupts

768
00:31:35,760 --> 00:31:42,600
and that might not sound like a lot but

769
00:31:38,520 --> 00:31:45,360
um in principle these Cisco handlers can

770
00:31:42,600 --> 00:31:48,000
take up to six arguments so that's six

771
00:31:45,360 --> 00:31:49,500
of the scratch registers uh consumed and

772
00:31:48,000 --> 00:31:53,279
so what I did is I just freed up

773
00:31:49,500 --> 00:31:56,820
register allocation in short so uh it it

774
00:31:53,279 --> 00:31:59,039
uh decreased the the total stack

775
00:31:56,820 --> 00:32:02,700
utilization and it turns out that yeah I

776
00:31:59,039 --> 00:32:04,860
got more than 10 speed up which is kind

777
00:32:02,700 --> 00:32:06,840
of phenomenal there actually not at all

778
00:32:04,860 --> 00:32:08,880
what I expected

779
00:32:06,840 --> 00:32:12,740
um yeah there's another part to that but

780
00:32:08,880 --> 00:32:12,740
I've gone any other questions

781
00:32:15,899 --> 00:32:18,440
hey

782
00:32:25,140 --> 00:32:29,880
um sort of a two-parter like uh how much

783
00:32:27,480 --> 00:32:32,039
background did you have like in kernel

784
00:32:29,880 --> 00:32:34,080
development or like

785
00:32:32,039 --> 00:32:36,899
um looking at this kind of low-level

786
00:32:34,080 --> 00:32:40,820
development before you got stuck into

787
00:32:36,899 --> 00:32:43,500
this work and secondly why was risk five

788
00:32:40,820 --> 00:32:45,480
impacted along with powerpc good one

789
00:32:43,500 --> 00:32:48,539
okay uh

790
00:32:45,480 --> 00:32:50,820
uh yes so what background did I have I

791
00:32:48,539 --> 00:32:54,000
primarily got into low-level programming

792
00:32:50,820 --> 00:32:55,740
actually by just writing simulations

793
00:32:54,000 --> 00:32:57,840
um I've touched a bit of Fortran but

794
00:32:55,740 --> 00:33:00,419
thankfully mostly in C

795
00:32:57,840 --> 00:33:02,100
um and uh so yeah just working with low

796
00:33:00,419 --> 00:33:06,240
level libraries like

797
00:33:02,100 --> 00:33:08,659
um MPI uh and openmp and whatnot yeah

798
00:33:06,240 --> 00:33:11,460
and I just discovered that I I just like

799
00:33:08,659 --> 00:33:14,159
tuning optimizing things

800
00:33:11,460 --> 00:33:15,720
um which is cool uh

801
00:33:14,159 --> 00:33:17,460
yeah

802
00:33:15,720 --> 00:33:20,460
um and then later on I got into Colonel

803
00:33:17,460 --> 00:33:22,679
Dev I took a class at Uni

804
00:33:20,460 --> 00:33:24,899
um that kind of interested me to read a

805
00:33:22,679 --> 00:33:26,519
whole bunch of Kernel code but didn't

806
00:33:24,899 --> 00:33:28,320
have any projects immediately which I

807
00:33:26,519 --> 00:33:29,159
necessarily wanted to write in the

808
00:33:28,320 --> 00:33:30,720
kernel

809
00:33:29,159 --> 00:33:33,179
um and so yeah I really did come in

810
00:33:30,720 --> 00:33:35,700
blind at some rate in terms of the the

811
00:33:33,179 --> 00:33:37,559
context of what I was looking at and um

812
00:33:35,700 --> 00:33:41,519
actually this was a this is a really

813
00:33:37,559 --> 00:33:42,799
good first task working on uh Cisco

814
00:33:41,519 --> 00:33:46,860
handlers

815
00:33:42,799 --> 00:33:49,019
in that uh essentially so the The Cisco

816
00:33:46,860 --> 00:33:52,880
interfaces I imagine probably the

817
00:33:49,019 --> 00:33:57,000
largest user to to core kernel

818
00:33:52,880 --> 00:33:59,940
interface and so uh the user user code

819
00:33:57,000 --> 00:34:01,620
needs to request resources of of all

820
00:33:59,940 --> 00:34:04,080
sorts and so it should eventually touch

821
00:34:01,620 --> 00:34:05,580
all subsystems in the core kernel and so

822
00:34:04,080 --> 00:34:07,980
yeah I got to read quite a bit of code

823
00:34:05,580 --> 00:34:09,300
that way so it's a good induction

824
00:34:07,980 --> 00:34:12,480
um and then your second question how did

825
00:34:09,300 --> 00:34:14,159
I break risk five as well

826
00:34:12,480 --> 00:34:16,440
um let's go back to that slide because I

827
00:34:14,159 --> 00:34:18,500
don't think I explained it super well

828
00:34:16,440 --> 00:34:18,500
um

829
00:34:19,980 --> 00:34:27,060
in short I um I influented this compat

830
00:34:23,099 --> 00:34:29,580
ARG u64 Jewel macro that was already

831
00:34:27,060 --> 00:34:32,760
implemented and being consumed by risk

832
00:34:29,580 --> 00:34:35,700
five and um uh

833
00:34:32,760 --> 00:34:37,460
I opted in powerpc to using it so we

834
00:34:35,700 --> 00:34:40,379
could remove some of our own custom

835
00:34:37,460 --> 00:34:41,220
Cisco handlers which were basically the

836
00:34:40,379 --> 00:34:46,139
same

837
00:34:41,220 --> 00:34:47,820
uh but critically I I exchanged their um

838
00:34:46,139 --> 00:34:50,240
working version of that Macro for a one

839
00:34:47,820 --> 00:34:50,240
which broke

840
00:34:51,980 --> 00:35:01,619
uh yeah I um I misplaced a um

841
00:34:58,200 --> 00:35:04,980
yeah if deaf big Indian

842
00:35:01,619 --> 00:35:07,440
um turns out not all architectures have

843
00:35:04,980 --> 00:35:09,240
a if def little Indian which is which is

844
00:35:07,440 --> 00:35:13,020
weird and more architectures are likely

845
00:35:09,240 --> 00:35:14,880
to have a a macro called defined config

846
00:35:13,020 --> 00:35:16,020
big endian then config little ending

847
00:35:14,880 --> 00:35:18,000
which is weird because more

848
00:35:16,020 --> 00:35:19,980
architectures I imagine a little Indian

849
00:35:18,000 --> 00:35:21,839
than and whatever

850
00:35:19,980 --> 00:35:24,359
fun

851
00:35:21,839 --> 00:35:26,640
yes indeed well I suppose if Little

852
00:35:24,359 --> 00:35:27,599
Indians the default then

853
00:35:26,640 --> 00:35:31,680
anyway

854
00:35:27,599 --> 00:35:34,680
whatever but um in short I uh

855
00:35:31,680 --> 00:35:36,780
I I realize this that oh if I have if

856
00:35:34,680 --> 00:35:38,460
config little endian I'm gonna break a

857
00:35:36,780 --> 00:35:40,859
whole bunch of architectures and so I

858
00:35:38,460 --> 00:35:41,700
decided let's not use that so I decided

859
00:35:40,859 --> 00:35:44,400
to use

860
00:35:41,700 --> 00:35:46,260
if and death big end in except I missed

861
00:35:44,400 --> 00:35:48,180
the N is down the end and so did my

862
00:35:46,260 --> 00:35:50,700
reviewers apparently so

863
00:35:48,180 --> 00:35:52,280
kind of been that important yeah

864
00:35:50,700 --> 00:35:55,579
um

865
00:35:52,280 --> 00:35:55,579
any more questions

866
00:35:57,960 --> 00:36:01,980
so I have a fully non non-technical

867
00:35:59,940 --> 00:36:04,200
question as a newcomer to the Linux

868
00:36:01,980 --> 00:36:05,820
Community what if anything do you think

869
00:36:04,200 --> 00:36:08,700
should be different about the Linux

870
00:36:05,820 --> 00:36:10,320
development culture to get more people

871
00:36:08,700 --> 00:36:13,380
like yourselves who are newcomers to

872
00:36:10,320 --> 00:36:14,820
involved and working Upstream yeah I

873
00:36:13,380 --> 00:36:17,640
would say that I mean the grads are

874
00:36:14,820 --> 00:36:19,440
today are natives right in in at very

875
00:36:17,640 --> 00:36:21,780
least GitHub

876
00:36:19,440 --> 00:36:24,540
um I mean sorry I think I'm only

877
00:36:21,780 --> 00:36:27,780
touching on concerns which are

878
00:36:24,540 --> 00:36:29,460
I think pretty broadly felt within even

879
00:36:27,780 --> 00:36:31,920
within the kernel Community who've

880
00:36:29,460 --> 00:36:34,859
solidified their workflows now for for

881
00:36:31,920 --> 00:36:38,579
some years but um yeah I think the key

882
00:36:34,859 --> 00:36:41,460
thing is um uh the way people contribute

883
00:36:38,579 --> 00:36:43,800
uh new grads will be more familiar of

884
00:36:41,460 --> 00:36:45,180
course with with merge requests

885
00:36:43,800 --> 00:36:46,619
um or pull requests or if you're on

886
00:36:45,180 --> 00:36:50,640
GitHub

887
00:36:46,619 --> 00:36:52,619
um and uh the way CI Works kernel CI is

888
00:36:50,640 --> 00:36:53,579
a talk um happening in this conference

889
00:36:52,619 --> 00:36:57,480
about it

890
00:36:53,579 --> 00:36:59,460
um Colonel CI uh has quite a quite a

891
00:36:57,480 --> 00:37:02,940
long history there are some really good

892
00:36:59,460 --> 00:37:06,240
examples of it with um uh so for example

893
00:37:02,940 --> 00:37:08,579
the zero day CI Which Intel does is

894
00:37:06,240 --> 00:37:10,380
um has been very helpful

895
00:37:08,579 --> 00:37:11,700
um but ultimately you know we built up

896
00:37:10,380 --> 00:37:14,160
this Matrix right of the sub

897
00:37:11,700 --> 00:37:15,839
architectures present present on on

898
00:37:14,160 --> 00:37:18,420
powerpc that's really the tip of the

899
00:37:15,839 --> 00:37:20,760
iceberg because there are any number of

900
00:37:18,420 --> 00:37:23,579
modules and configs of any sort so

901
00:37:20,760 --> 00:37:25,380
essentially the The Matrix of things

902
00:37:23,579 --> 00:37:26,820
which you might want to test become

903
00:37:25,380 --> 00:37:27,660
enormous

904
00:37:26,820 --> 00:37:30,240
um

905
00:37:27,660 --> 00:37:32,460
yeah yeah there are there are some very

906
00:37:30,240 --> 00:37:33,780
good talks from previous lcas on on the

907
00:37:32,460 --> 00:37:36,420
state of testing but yeah it's

908
00:37:33,780 --> 00:37:38,400
fundamentally it's it's testing and how

909
00:37:36,420 --> 00:37:40,859
do you contribute

910
00:37:38,400 --> 00:37:42,660
um setting up your your mail client to

911
00:37:40,859 --> 00:37:44,400
be able to contribute is it's kind of

912
00:37:42,660 --> 00:37:46,079
yeah

913
00:37:44,400 --> 00:37:48,839
um non-trivial

914
00:37:46,079 --> 00:37:51,240
um and so in just decreasing that that

915
00:37:48,839 --> 00:37:55,560
barrier to entry

916
00:37:51,240 --> 00:37:57,820
um yeah I'd say barrier to entry yep

917
00:37:55,560 --> 00:38:01,030
just thank you

918
00:37:57,820 --> 00:38:01,030
[Applause]