We have 3 mutexes: mheap_.lock, stackpoolmu and persistent.lock.
mheap_.lock is locked under stackpoolmu when we allocate new stack segments.
persistent.lock is locked under mheap_.lock when we allocate new mspans.
stackpoolmu is locked under persistent.lock when persistentalloc grows stack.
The exact stacks are provided below.
This is a deadly cycle.

That last edge (stackpoolmu under persistent.lock) is not possible on linux as far as I
see. But it seem to be possible today on plan9, where sysAlloc has deeper stack than
lock. Also it can fire on any OS if/when we change code of persistentalloc or sysAlloc,
or do more aggressive stack shrinking.

Mheap has to run on system stack and mheap uses persistentalloc. So the solution is to
run persistentalloc on system stack.

=======

runtime: stackpoolmu must not be locked while holding persistent.lock
fatal error: runtime deadlock

runtime stack:
runtime.gothrow(0x488970, 0x10)
    src/runtime/panic.go:507 +0x8e
runtime.ddLock(0x666c48)
    src/runtime/dd.go:361 +0x28e
runtime.lock(0x666c48)
    src/runtime/lock_sema.go:42 +0x87
runtime.newstack()
    src/runtime/stack1.go:661 +0x4a7
runtime.morestack()
    src/runtime/asm_amd64.s:311 +0x7e

goroutine 43 [running]:
runtime.unlock(0x6527a0)
    src/runtime/lock_sema.go:97 +0x14 fp=0xc208095120 sp=0xc208095118
runtime.persistentalloc(0x28, 0x8, 0x66a680, 0x1977ee)
    src/runtime/malloc.go:809 +0x201 fp=0xc208095148 sp=0xc208095120
runtime.getitab(0x3a6460, 0x3ea260, 0x0, 0x3f7620)
    src/runtime/iface.go:79 +0x68c fp=0xc2080951c8 sp=0xc208095148
runtime.convI2I(0x3a6460, 0x7253c0, 0x652160, 0x0, 0x0)
    src/runtime/iface.go:310 +0xdc fp=0xc208095200 sp=0xc2080951c8

==============

runtime: mheap_.lock must not be locked while holding stackpoolmu
fatal error: runtime deadlock

runtime stack:
runtime.gothrow(0x488970, 0x10)
    src/runtime/panic.go:507 +0x8e
runtime.ddLock(0x65a960)
    src/runtime/dd.go:361 +0x28e
runtime.lock(0x65a960)
    src/runtime/lock_sema.go:42 +0x87
runtime.mHeap_AllocStack(0x65a960, 0x4, 0x1)
    src/runtime/mheap.go:248 +0x66
runtime.stackpoolalloc(0x666c00, 0x0)
    src/runtime/stack1.go:66 +0x61
runtime.stackcacherefill(0x710000, 0x0)
    src/runtime/stack1.go:130 +0x5b
runtime.stackalloc(0x800, 0x0, 0x0)
    src/runtime/stack1.go:230 +0x1bd
runtime.func·036()
    src/runtime/proc1.go:1875 +0x31
runtime.systemstack(0x7fff5fbff8a0)
    src/runtime/asm_amd64.s:260 +0xa3
runtime.malg(0xc200000800, 0x0)
    src/runtime/proc1.go:1876 +0x72
runtime.newproc1(0x5149b0, 0x7fff5fbff9b0, 0x0, 0x47f35, 0x0)
    src/runtime/proc1.go:1928 +0xe9

==============

runtime: persistent.lock must not be locked while holding mheap_.lock
fatal error: runtime deadlock

runtime stack:
runtime.gothrow(0x488970, 0x10)
    src/runtime/panic.go:507 +0x8e
runtime.ddLock(0x6527a0)
    src/runtime/dd.go:361 +0x28e
runtime.lock(0x6527a0)
    src/runtime/lock_sema.go:42 +0x87
runtime.persistentalloc(0x4000, 0x8, 0x66a658, 0x1d7cb)
    src/runtime/malloc.go:797 +0xbe
runtime.fixAlloc_Alloc(0x665df8, 0x1)
    src/runtime/mfixalloc.go:39 +0xeb
runtime.mHeap_AllocSpanLocked(0x65a960, 0x1, 0xc208013201)
    src/runtime/mheap.go:306 +0x187
runtime.mHeap_Alloc_m(0x65a960, 0x1, 0x13, 0x652f00)
    src/runtime/mheap.go:194 +0x12c
runtime.func·019()
    src/runtime/mheap.go:231 +0x56
runtime.systemstack(0x7fff5fbff868)
    src/runtime/asm_amd64.s:260 +0xa3
runtime.mHeap_Alloc(0x65a960, 0x1, 0x10000000013, 0x107a2)
    src/runtime/mheap.go:232 +0x68
runtime.mCentral_Grow(0x6627f8, 0x1a5ec)
    src/runtime/mcentral.go:181 +0x96
runtime.mCentral_CacheSpan(0x6627f8, 0x428)
    src/runtime/mcentral.go:77 +0x358
runtime.mCache_Refill(0x710000, 0x13, 0x428)
    src/runtime/mcache.go:65 +0xd6
runtime.func·007()
    src/runtime/malloc.go:175 +0x35
runtime.systemstack(0x652f30)
    src/runtime/asm_amd64.s:244 +0x71
runtime.mstart()
    src/runtime/proc1.go:693