Annotation of wikisrc/projects/project/smp_networking.mdwn, revision 1.6

1.1       jmmv        1: [[!template id=project
                      2: 
                      3: title="SMP Networking (aka remove the big network lock)"
                      4: 
                      5: contact="""
                      6: [tech-kern](mailto:tech-kern@NetBSD.org),
1.2       jmmv        7: [tech-net](mailto:tech-net@NetBSD.org),
1.1       jmmv        8: [board](mailto:board@NetBSD.org),
                      9: [core](mailto:core@NetBSD.org)
                     10: """
                     11: 
1.2       jmmv       12: category="networking"
1.1       jmmv       13: difficulty="hard"
                     14: 
                     15: description="""
1.2       jmmv       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.
1.1       jmmv       22: 
                     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.
1.2       jmmv       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
                     29: problem.
1.1       jmmv       30: 
1.6     ! wiki       31: # This project is currently claimed
        !            32: 
1.2       jmmv       33: # Funding
1.1       jmmv       34: 
                     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,
1.2       jmmv       37: please send your proposal to the contact addresses listed above.
                     38: 
1.3       jmmv       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).
                     46: 
1.2       jmmv       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.
                     51: 
                     52: # Tentative specification
                     53: 
                     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.
                     59: 
                     60: You can divide the network infrastructure into 5 major components:
                     61: 
                     62: * Interfaces (both real devices and pseudo-devices)
                     63: * Socket code
                     64: * Protocols
                     65: * Routing code
                     66: * mbuf code.
                     67: 
                     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.
                     71: 
                     72: Part of the problem are asynchonous upcalls, among which include:
                     73: 
                     74: * `ifa->ifa_rtrequest` for route changes.
                     75: * `pr_ctlinput` for interface events.
                     76: 
                     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.
                     80: 
1.3       jmmv       81: ## Subtasks
                     82: 
1.2       jmmv       83: The proposed solution presented here include the following tasks (in no
                     84: particular order) to achieve the desired goals of SMP support and
                     85: modularity:
                     86: 
1.5       jmmv       87: [[!map show="title" pages="projects/project/* and tagged(project) and tagged(smp_networking) and tagged(status:active)"]]
1.2       jmmv       88: 
1.3       jmmv       89: ## Work plan
                     90: 
                     91: Aside from the list of tasks above, the work to be done for this project
                     92: can be achieved by following these steps:
                     93: 
                     94: 1. Move ARP out of the routing table.  See the [[nexthop_cache]] project.
                     95: 
                     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.
                    103: 
                    104: 1. Collect global variables in the IP/UDP/TCP protocols into structures.
                    105:    This helps the following items.
                    106: 
                    107: 1. Make IPV4/ICMP/IGMP/REASS MP-friendly.
                    108: 
                    109: 1. Make IPV6/ICMP/IGMP/ND MP-friendly.
                    110: 
                    111: 1. Make TCP MP-friendly.
                    112: 
                    113: 1. Make UDP MP-friendly.
                    114: 
1.2       jmmv      115: # Radical thoughts
                    116: 
                    117: You should also consider the following ideas:
                    118: 
                    119: ## LWPs in user space do not need a kernel stack
                    120: 
                    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.
                    128: 
                    129: ## LWPs waiting for kernel condition shouldn't need a kernel stack
                    130: 
                    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.
                    134: 
                    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.
                    145: 
                    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.
                    152: 
                    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
                    157: handler.
1.1       jmmv      158: """
                    159: ]]

CVSweb for NetBSD wikisrc <wikimaster@NetBSD.org> software: FreeBSD-CVSweb