xref: /netbsd-src/doc/TODO.smpnet (revision a2e2328d6d63d0479580e9c30629dd48eaadff26)
1$NetBSD: TODO.smpnet,v 1.50 2024/08/12 10:46:40 nia Exp $
2
3MP-safe components
4==================
5
6They work without the big kernel lock (KERNEL_LOCK), i.e., with NET_MPSAFE
7kernel option.  Some components scale up and some don't.
8
9 - Device drivers
10   - aq(4)
11   - awge(4)
12   - bcmgenet(4)
13   - bge(4)
14   - ena(4)
15   - iavf(4)
16   - ixg(4)
17   - ixl(4)
18   - ixv(4)
19   - mcx(4)
20   - rge(4)
21   - se(4)
22   - sunxi_emac(4)
23   - vioif(4)
24   - vmx(4)
25   - wm(4)
26   - xennet(4)
27   - usbnet(4) based adapters:
28     - axe(4)
29     - axen(4)
30     - cdce(4)
31     - cue(4)
32     - kue(4)
33     - mos(4)
34     - mue(4)
35     - smsc(4)
36     - udav(4)
37     - upl(4)
38     - ure(4)
39     - url(4)
40     - urndis(4)
41 - Layer 2
42   - Ethernet (if_ethersubr.c)
43   - bridge(4)
44     - STP
45   - Fast forward (ipflow)
46 - Layer 3
47   - All except for items in the below section
48 - Interfaces
49   - canloop(4)
50   - gif(4)
51   - ipsecif(4)
52   - l2tp(4)
53   - lagg(4)
54   - pppoe(4)
55     - if_spppsubr.c
56   - tap(4)
57   - tun(4)
58   - vether(4)
59   - vlan(4)
60 - Packet filters
61   - npf(7)
62   - ipf(4)
63 - Others
64   - bpf(4)
65   - ipsec(4)
66   - opencrypto(9)
67   - pfil(9)
68
69Non MP-safe components and kernel options
70=========================================
71
72The components and options aren't MP-safe, i.e., requires the big kernel lock,
73yet.  Some of them can be used safely even if NET_MPSAFE is enabled because
74they're still protected by the big kernel lock.  The others aren't protected and
75so unsafe, e.g, they may crash the kernel.
76
77Protected ones
78--------------
79
80 - Device drivers
81   - Most drivers other than ones listed in the above section
82 - Layer 4
83   - DCCP
84   - SCTP
85   - TCP
86   - UDP
87
88Unprotected ones
89----------------
90
91 - Layer 2
92   - ARCNET (if_arcsubr.c)
93   - IEEE 1394 (if_ieee1394subr.c)
94   - IEEE 802.11 (ieee80211(4))
95 - Layer 3
96   - IPSELSRC
97   - MROUTING
98   - PIM
99   - MPLS (mpls(4))
100   - IPv6 address selection policy
101 - Interfaces
102   - agr(4)
103   - carp(4)
104   - faith(4)
105   - gre(4)
106   - ppp(4)
107   - sl(4)
108   - stf(4)
109   - if_srt
110 - Packet filters
111   - pf(4)
112 - Others
113   - AppleTalk (sys/netatalk/)
114   - Bluetooth (sys/netbt/)
115   - altq(4)
116   - kttcp(4)
117   - NFS
118
119Know issues
120===========
121
122NOMPSAFE
123--------
124
125We use "NOMPSAFE" as a mark that indicates that the code around it isn't MP-safe
126yet.  We use it in comments and also use as part of function names, for example
127m_get_rcvif_NOMPSAFE.  Let's use "NOMPSAFE" to make it easy to find non-MP-safe
128codes by grep.
129
130bpf
131---
132
133MP-ification of bpf requires all of bpf_mtap* are called in normal LWP context
134or softint context, i.e., not in hardware interrupt context.  For Tx, all
135bpf_mtap satisfy the requirement.  For Rx, most of bpf_mtap are called in softint.
136Unfortunately some bpf_mtap on Rx are still called in hardware interrupt context.
137
138This is the list of the functions that have such bpf_mtap:
139
140 - sca_frame_process() @ sys/dev/ic/hd64570.c
141
142Ideally we should make the functions run in softint somehow, but we don't have
143actual devices, no time (or interest/love) to work on the task, so instead we
144provide a deferred bpf_mtap mechanism that forcibly runs bpf_mtap in softint
145context.  It's a workaround and once the functions run in softint, we should use
146the original bpf_mtap again.
147
148if_mcast_op() - SIOCADDMULTI/SIOCDELMULTI
149-----------------------------------------
150Helper function is called to add or remove multicast addresses for
151interface.  When called via ioctl it takes IFNET_LOCK(), when called
152via sosetopt() it doesn't.
153
154Various network drivers can't assert IFNET_LOCKED() in their if_ioctl
155because of this. Generally drivers still take care to splnet() even
156with NET_MPSAFE before calling ether_ioctl(), but they do not take
157KERNEL_LOCK(), so this is actually unsafe.
158
159Lingering obsolete variables
160-----------------------------
161
162Some obsolete global variables and member variables of structures remain to
163avoid breaking old userland programs which directly access such variables via
164kvm(3).
165
166The following programs still use kvm(3) to get some information related to
167the network stack.
168
169 - netstat(1)
170 - vmstat(1)
171 - fstat(1)
172
173netstat(1) accesses ifnet_list, the head of a list of interface objects
174(struct ifnet), and traverses each object through ifnet#if_list member variable.
175ifnet_list and ifnet#if_list is obsoleted by ifnet_pslist and
176ifnet#if_pslist_entry respectively. netstat also accesses the IP address list
177of an interface through ifnet#if_addrlist. struct ifaddr, struct in_ifaddr
178and struct in6_ifaddr are accessed and the following obsolete member variables
179are stuck: ifaddr#ifa_list, in_ifaddr#ia_hash, in_ifaddr#ia_list,
180in6_ifaddr#ia_next and in6_ifaddr#_ia6_multiaddrs. Note that netstat already
181implements alternative methods to fetch the above information via sysctl(3).
182
183vmstat(1) shows statistics of hash tables created by hashinit(9) in the kernel.
184The statistic information is retrieved via kvm(3). The global variables
185in_ifaddrhash and in_ifaddrhashtbl, which are for a hash table of IPv4
186addresses and obsoleted by in_ifaddrhash_pslist and in_ifaddrhashtbl_pslist,
187are kept for this purpose. We should provide a means to fetch statistics of
188hash tables via sysctl(3).
189
190fstat(1) shows information of bpf instances. Each bpf instance (struct bpf) is
191obtained via kvm(3). bpf_d#_bd_next, bpf_d#_bd_filter and bpf_d#_bd_list
192member variables are obsolete but remain. ifnet#if_xname is also accessed
193via struct bpf_if and obsolete ifnet#if_list is required to remain to not change
194the offset of ifnet#if_xname. The statistic counters (bpf#bd_rcount,
195bpf#bd_dcount and bpf#bd_ccount) are also victims of this restriction; for
196scalability the statistic counters should be per-CPU and we should stop using
197atomic operations for them however we have to remain the counters and atomic
198operations.
199
200Scalability
201-----------
202
203 - Per-CPU rtcaches (used in say IP forwarding) aren't scalable on multiple
204   flows per CPU
205 - ipsec(4) isn't scalable on the number of SA/SP; the cost of a look-up
206   is O(n)
207 - opencrypto(9)'s crypto_newsession()/crypto_freesession() aren't scalable
208   as they are serialized by one mutex
209
210ALTQ
211----
212
213If ALTQ is enabled in the kernel, it enforces to use just one Tx queue (if_snd)
214for packet transmissions, resulting in serializing all Tx packet processing on
215the queue.  We should probably design and implement an alternative queuing
216mechanism that deals with multi-core systems at the first place, not making the
217existing ALTQ MP-safe because it's just annoying.
218
219Using kernel modules
220--------------------
221
222Please note that if you enable NET_MPSAFE in your kernel, and you use and
223loadable kernel modules (including compat_xx modules or individual network
224interface if_xxx device driver modules), you will need to build custom
225modules.  For each module you will need to add the following line to its
226Makefile:
227
228	CPPFLAGS+=	NET_MPSAFE
229
230Failure to do this may result in unpredictable behavior.
231
232IPv4 address initialization atomicity
233-------------------------------------
234
235An IPv4 address is referenced by several data structures: an associated
236interface, its local route, a connected route (if necessary), the global list,
237the global hash table, etc.  These data structures are not updated atomically,
238i.e., there can be inconsistent states on an IPv4 address in the kernel during
239the initialization of an IPv4 address.
240
241One known failure of the issue is that incoming packets destinating to an
242initializing address can loop in the network stack in a short period of time.
243The address initialization creates an local route first and then registers an
244initializing address to the global hash table that is used to decide if an
245incoming packet destinates to the host by checking the destination of the packet
246is registered to the hash table.  So, if the host allows forwarding, an incoming
247packet can match on a local route of an initializing address at ip_output while
248it fails the to-self check described above at ip_input.  Because a matched local
249route points a loopback interface as its destination interface, an incoming
250packet sends to the network stack (ip_input) again, which results in looping.
251The loop stops once an initializing address is registered to the hash table.
252
253One solution of the issue is to reorder the address initialization instructions,
254first register an address to the hash table then create its routes.  Another
255solution is to use the routing table for the to-self check instead of using the
256global hash table, like IPv6.
257
258if_flags
259--------
260
261To avoid data race on if_flags it should be protected by a lock (currently it's
262IFNET_LOCK).  Thus, if_flags should not be accessed on packet processing to
263avoid performance degradation by lock contentions.  Traditionally IFF_RUNNING,
264IFF_UP and IFF_OACTIVE flags of if_flags are checked on packet processing.  If
265you make a driver MP-safe you must remove such checks.
266
267Drivers should not touch IFF_ALLMULTI.  They are tempted to do so when updating
268hardware multicast filters on SIOCADDMULTI/SIOCDELMULTI.  Instead, they should
269use the ETHER_F_ALLMULTI bit in struct ethercom::ec_flags, under ETHER_LOCK.
270ether_ioctl takes care of presenting IFF_ALLMULTI according to the current state
271of ETHER_F_ALLMULTI when queried with SIOCGIFFLAGS.
272
273Also IFF_PROMISC is checked in ether_input and we should get rid of it somehow.
274