This has been logged as CR 6979725 but not yet fixed by Oracle.
http://bugs.opensolaris.org/bugdatabase/view_bug.do?bug_id=6979725

I experienced the issue during a filebench run on Nexenta in which the file "re-use" operation (msync) on a 20GB file caused all I/O to hang for more than 30 seconds, even other zpools. Stack trace looked like this:

[ ffffff0010429990 _resume_from_idle+0xf1() ]
ffffff0010429b10 page_lookup_create+0xd6(ffffff02d2813c40, 37588b000, 1, 0, 0, 0)
ffffff0010429b40 page_lookup+0x26(ffffff02d2813c40, 375889000, 1)
ffffff0010429c00 zfs_putpage+0x1af(ffffff02d2813c40, 3758a5000, 1000, 10000, ffffff02cf328f20, 0)
ffffff0010429c70 fop_putpage+0x74(ffffff02d2813c40, 3758b0000, 1000, 10000, ffffff02cf328f20, 0)
ffffff0010429d50 segvn_sync+0x459(ffffff02ecb4e7f0, fffffd7f63c00000, 40000000, 0, 6)
ffffff0010429e40 as_ctl+0x200(ffffff02d15efda0, fffffd7f63c00000, 40000000, 1, 0, 6, 0, 0)
ffffff0010429f00 memcntl+0x764(fffffd7f63c00000, 40000000, 1, 2, 0, 0)
ffffff0010429f10 sys_syscall+0x17a()

I created the attached patch and re-ran the filebench tests, and the long I/O stalls were reduced to only 3 or 4 seconds. I'm creating this issue report since the problem most likely exists in Illumos as well. Looking for comments/feedback on whether this fix seems correct and appropriate. Thanks!