1: [[!template id=project
3: title="SMP Networking (aka remove the big network lock)"
16: Traditionally, the NetBSD kernel code had been protected by a single,
17: global lock. This lock ensured that, on a multiprocessor system, two
18: different threads of execution did not access the kernel concurrently and
19: thus simplified the internal design of the kernel. However, such design
20: does not scale to multiprocessor machines because, effectively, the kernel
21: is restricted to run on a single processor at any given time.
23: The NetBSD kernel has been modified to use fine grained locks in many of
24: its different subsystems, achieving good performance on today's
25: multiprocessor machines. Unfotunately, these changes have not yet been
26: applied to the networking code, which remains protected by the single lock.
27: In other words: NetBSD networking has evolved to work in a uniprocessor
28: envionment; switching it to use fine-grained locked is a hard and complex
31: # This project is currently claimed
33: # Funding
35: At this time, The NetBSD Foundation is accepting project specifications to
36: remove the single networking lock. If you want to apply for this project,
37: please send your proposal to the contact addresses listed above.
39: Due to the size of this project, your proposal does not need to cover
40: everything to qualify for funding. We have attempted to split the work
41: into smaller units, and **you can submit funding applications for these
42: smaller subtasks independently** as long as the work you deliver fits in
43: the grand order of this project. For example, you could send an
44: application to make the network interfaces alone MP-friendly (see the *work
45: plan* below).
47: What follows is a particular design proposal, extracted from an
48: [original text](http://www.NetBSD.org/~matt/smpnet.html) written by
49: [Matt Thomas](mailto:matt@NetBSD.org). You may choose to work on this
50: particular proposal or come up with your own.
52: # Tentative specification
54: The future of NetBSD network infrastructure has to efficiently embrace two
55: major design criteria: Symmetric Multi-Processing (SMP) and modularity.
56: Other design considerations include not only supporting but taking
57: advantage of the capability of newer network devices to do packet
58: classification, payload splitting, and even full connection offload.
60: You can divide the network infrastructure into 5 major components:
62: * Interfaces (both real devices and pseudo-devices)
63: * Socket code
64: * Protocols
65: * Routing code
66: * mbuf code.
68: Part of the complexity is that, due to the monolithic nature of the kernel,
69: each layer currently feels free to call any other layer. This makes
70: designing a lock hierarchy difficult and likely to fail.
72: Part of the problem are asynchonous upcalls, among which include:
74: * `ifa->ifa_rtrequest` for route changes.
75: * `pr_ctlinput` for interface events.
77: Another source of complexity is the large number of global variables
78: scattered throughout the source files. This makes putting locks around
79: them difficult.
81: ## Subtasks
83: The proposed solution presented here include the following tasks (in no
84: particular order) to achieve the desired goals of SMP support and
87: [[!map show="title" pages="projects/project/* and tagged(project) and tagged(smp_networking) and tagged(status:active)"]]
89: ## Work plan
91: Aside from the list of tasks above, the work to be done for this project
92: can be achieved by following these steps:
94: 1. Move ARP out of the routing table. See the [[nexthop_cache]] project.
96: 1. Make the network interfaces MP, which are one of the few users of the
97: big kernel lock left. This needs to support multiple receive and
98: transmit queues to help reduce locking contention. This also includes
99: changing more of the common interfaces to do what the `tsec` driver does
100: (basically do everything with softints). This also needs to change the
101: `*_input` routines to use a table to do dispatch instead of the current
102: switch code so domain can be dynamically loaded.
104: 1. Collect global variables in the IP/UDP/TCP protocols into structures.
105: This helps the following items.
107: 1. Make IPV4/ICMP/IGMP/REASS MP-friendly.
109: 1. Make IPV6/ICMP/IGMP/ND MP-friendly.
111: 1. Make TCP MP-friendly.
113: 1. Make UDP MP-friendly.
115: # Radical thoughts
117: You should also consider the following ideas:
119: ## LWPs in user space do not need a kernel stack
121: Those pages are only being used in case the an exception happens.
122: Interrupts are probably going to their own dedicated stack. One could just
123: keep a set of kernel stacks around. Each CPU has one, when a user
124: exception happens, that stack is assigned to the current LWP and removed as
125: the active CPU one. When that CPU next returns to user space, the kernel
126: stack it was using is saved to be used for the next user exception. The
127: idle lwp would just use the current kernel stack.
129: ## LWPs waiting for kernel condition shouldn't need a kernel stack
131: If an LWP is waiting on a kernel condition variable, it is expecting to be
132: inactive for some time, possibly a long time. During this inactivity, it
133: does not really need a kernel stack.
135: When the exception handler get an usermode exeception, it sets LWP
136: restartable flag that indicates that the exception is restartable, and then
137: services the exception as normal. As routines are called, they can clear
138: the LWP restartable flag as needed. When an LWP needs to block for a long
139: time, instead of calling `cv_wait`, it could call `cv_restart`. If
140: `cv_restart` returned false, the LWPs restartable flag was clear so
141: `cv_restart` acted just like `cv_wait`. Otherwise, the LWP and CV would
142: have been tied together (big hand wave), the lock had been released and the
143: routine should have returned `ERESTART`. `cv_restart` could also wait for
144: a small amount of time like .5 second, and only if the timeout expires.
146: As the stack unwinds, eventually, it would return to the last the exception
147: handler. The exception would see the LWP has a bound CV, save the LWP's
148: user state into the PCB, set the LWP to sleeping, mark the lwp's stack as
149: idle, and call the scheduler to find more work. When called,
150: `cpu_switchto` would notice the stack is marked idle, and detach it from
151: the LWP.
153: When the condition times out or is signalled, the first LWP attached to the
154: condition variable is marked runnable and detached from the CV. When the
155: `cpu_switchto` routine is called, the it would notice the lack of a stack
156: so it would grab one, restore the trapframe, and reinvoke the exception
CVSweb for NetBSD wikisrc <wikimaster@NetBSD.org> software: FreeBSD-CVSweb