Annotation of wikisrc/projects/project/smp_networking.mdwn, revision 1.5

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: funded="The NetBSD Foundation"
                     15: 
                     16: description="""
1.2       jmmv       17: Traditionally, the NetBSD kernel code had been protected by a single,
                     18: global lock.  This lock ensured that, on a multiprocessor system, two
                     19: different threads of execution did not access the kernel concurrently and
                     20: thus simplified the internal design of the kernel.  However, such design
                     21: does not scale to multiprocessor machines because, effectively, the kernel
                     22: is restricted to run on a single processor at any given time.
1.1       jmmv       23: 
                     24: The NetBSD kernel has been modified to use fine grained locks in many of
                     25: its different subsystems, achieving good performance on today's
                     26: multiprocessor machines.  Unfotunately, these changes have not yet been
                     27: applied to the networking code, which remains protected by the single lock.
1.2       jmmv       28: In other words: NetBSD networking has evolved to work in a uniprocessor
                     29: envionment; switching it to use fine-grained locked is a hard and complex
                     30: problem.
1.1       jmmv       31: 
1.2       jmmv       32: # Funding
1.1       jmmv       33: 
                     34: At this time, The NetBSD Foundation is accepting project specifications to
                     35: remove the single networking lock.  If you want to apply for this project,
1.2       jmmv       36: please send your proposal to the contact addresses listed above.
                     37: 
1.3       jmmv       38: Due to the size of this project, your proposal does not need to cover
                     39: everything to qualify for funding.  We have attempted to split the work
                     40: into smaller units, and **you can submit funding applications for these
                     41: smaller subtasks independently** as long as the work you deliver fits in
                     42: the grand order of this project.  For example, you could send an
                     43: application to make the network interfaces alone MP-friendly (see the *work
                     44: plan* below).
                     45: 
1.2       jmmv       46: What follows is a particular design proposal, extracted from an
                     47: [original text](http://www.NetBSD.org/~matt/smpnet.html) written by
                     48: [Matt Thomas](mailto:matt@NetBSD.org).  You may choose to work on this
                     49: particular proposal or come up with your own.
                     50: 
                     51: # Tentative specification
                     52: 
                     53: The future of NetBSD network infrastructure has to efficiently embrace two
                     54: major design criteria: Symmetric Multi-Processing (SMP) and modularity.
                     55: Other design considerations include not only supporting but taking
                     56: advantage of the capability of newer network devices to do packet
                     57: classification, payload splitting, and even full connection offload.
                     58: 
                     59: You can divide the network infrastructure into 5 major components:
                     60: 
                     61: * Interfaces (both real devices and pseudo-devices)
                     62: * Socket code
                     63: * Protocols
                     64: * Routing code
                     65: * mbuf code.
                     66: 
                     67: Part of the complexity is that, due to the monolithic nature of the kernel,
                     68: each layer currently feels free to call any other layer.  This makes
                     69: designing a lock hierarchy difficult and likely to fail.
                     70: 
                     71: Part of the problem are asynchonous upcalls, among which include:
                     72: 
                     73: * `ifa->ifa_rtrequest` for route changes.
                     74: * `pr_ctlinput` for interface events.
                     75: 
                     76: Another source of complexity is the large number of global variables
                     77: scattered throughout the source files.  This makes putting locks around
                     78: them difficult.
                     79: 
1.3       jmmv       80: ## Subtasks
                     81: 
1.2       jmmv       82: The proposed solution presented here include the following tasks (in no
                     83: particular order) to achieve the desired goals of SMP support and
                     84: modularity:
                     85: 
1.5     ! jmmv       86: [[!map show="title" pages="projects/project/* and tagged(project) and tagged(smp_networking) and tagged(status:active)"]]
1.2       jmmv       87: 
1.3       jmmv       88: ## Work plan
                     89: 
                     90: Aside from the list of tasks above, the work to be done for this project
                     91: can be achieved by following these steps:
                     92: 
                     93: 1. Move ARP out of the routing table.  See the [[nexthop_cache]] project.
                     94: 
                     95: 1. Make the network interfaces MP, which are one of the few users of the
                     96:    big kernel lock left.  This needs to support multiple receive and
                     97:    transmit queues to help reduce locking contention.  This also includes
                     98:    changing more of the common interfaces to do what the `tsec` driver does
                     99:    (basically do everything with softints).  This also needs to change the
                    100:    `*_input` routines to use a table to do dispatch instead of the current
                    101:    switch code so domain can be dynamically loaded.
                    102: 
                    103: 1. Collect global variables in the IP/UDP/TCP protocols into structures.
                    104:    This helps the following items.
                    105: 
                    106: 1. Make IPV4/ICMP/IGMP/REASS MP-friendly.
                    107: 
                    108: 1. Make IPV6/ICMP/IGMP/ND MP-friendly.
                    109: 
                    110: 1. Make TCP MP-friendly.
                    111: 
                    112: 1. Make UDP MP-friendly.
                    113: 
1.2       jmmv      114: # Radical thoughts
                    115: 
                    116: You should also consider the following ideas:
                    117: 
                    118: ## LWPs in user space do not need a kernel stack
                    119: 
                    120: Those pages are only being used in case the an exception happens.
                    121: Interrupts are probably going to their own dedicated stack.  One could just
                    122: keep a set of kernel stacks around.  Each CPU has one, when a user
                    123: exception happens, that stack is assigned to the current LWP and removed as
                    124: the active CPU one.  When that CPU next returns to user space, the kernel
                    125: stack it was using is saved to be used for the next user exception.  The
                    126: idle lwp would just use the current kernel stack.
                    127: 
                    128: ## LWPs waiting for kernel condition shouldn't need a kernel stack
                    129: 
                    130: If an LWP is waiting on a kernel condition variable, it is expecting to be
                    131: inactive for some time, possibly a long time.  During this inactivity, it
                    132: does not really need a kernel stack.
                    133: 
                    134: When the exception handler get an usermode exeception, it sets LWP
                    135: restartable flag that indicates that the exception is restartable, and then
                    136: services the exception as normal.  As routines are called, they can clear
                    137: the LWP restartable flag as needed.  When an LWP needs to block for a long
                    138: time, instead of calling `cv_wait`, it could call `cv_restart`.  If
                    139: `cv_restart` returned false, the LWPs restartable flag was clear so
                    140: `cv_restart` acted just like `cv_wait`.  Otherwise, the LWP and CV would
                    141: have been tied together (big hand wave), the lock had been released and the
                    142: routine should have returned `ERESTART`.  `cv_restart` could also wait for
                    143: a small amount of time like .5 second, and only if the timeout expires.
                    144: 
                    145: As the stack unwinds, eventually, it would return to the last the exception
                    146: handler.  The exception would see the LWP has a bound CV, save the LWP's
                    147: user state into the PCB, set the LWP to sleeping, mark the lwp's stack as
                    148: idle, and call the scheduler to find more work.  When called,
                    149: `cpu_switchto` would notice the stack is marked idle, and detach it from
                    150: the LWP.
                    151: 
                    152: When the condition times out or is signalled, the first LWP attached to the
                    153: condition variable is marked runnable and detached from the CV.  When the
                    154: `cpu_switchto` routine is called, the it would notice the lack of a stack
                    155: so it would grab one, restore the trapframe, and reinvoke the exception
                    156: handler.
1.1       jmmv      157: """
                    158: ]]

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