Difference between revisions of "Zfs.exports"

Latest revision as of 02:48, 15 April 2021

zfs.exports[edit]

We used to list functions from XNU's private.exports, and using a codeless kext let's us call them. With 2.0 re-port, we no longer use zfs.exports, but it would be nice to have some calls back.

_cpu_number ARM64
 
Comment: Used to spread out lock contention, we use the lock at active-core-index
Alternative: None, only using [0] on arm64. 
Result: performance penalty

_fp_lookup
_fp_drop
_fp_drop_written
_fo_read
_fo_write

Comment: "zfs send" and "zfs recv" will either do IO on a file (vnode) or pipe. There is no
way to do IO on a pipe in a kext without private.exports. 
Alternative: userland will mknod /tmp/.pipe, spawn a relay helper, and open(/tmp/.pipe) which passes
a vnode to kernel. This works, but is complicated, both setting up pipes, and cleaning up. Especially with
SIGINT involved. It would be nice to do pipe IO in kext.
Result: Complicated code, but already working. Pipes love to deadlock, so corner cases could exist.

_vnode_iocount

Comment: Used to know if VFS will re-enter before vnode_put(). 
Alternative: ((int32_t *)vp)[25] - disgusting, but no choice. Really don't want to look in struct vnode.
Result: None, but if struct vnode changes we explode.

_hostname

Comment: uses hostname for diagnostic messages, history logs and multihost mounting (failover would
display which host stopped and which host took over.)
Alternative: None. Some work arounds could be done, like passing hostname from userland. Hashing hostid as string.
Result: user confusion to which host was in use. (Until work around)

_virtual_space_start
_virtual_space_end
_vm_page_free_wanted
_vm_page_free_count
_vm_page_free_min
_vm_page_speculative_count
_vm_pool_low

Comment: used to detect memory pressure well before jettison. 
Alternative: implement with optional hardcap limit, mach_vm_pressure_level_monitor and
mach_vm_pressure_monitor.
Result: Hopefully can be made to run the same, without stutter.

_VFS_ROOT
_rootvnode

Comment: Used with boot to know when root has been mounted. vfs_rootvnode()
does not check for NULL and will just panic.
Alternative: None, It would be better if vfs_rootvnode() would not panic.
Result: No boot support.

_cache_purgevfs

Comment: Used to drop all cache entires for a mount, for example "zfs rollback".
Alternative: Implemented with vnode_iterate(cache_purge_negatives() cache_purge()). Working.
Result: None. Happy to live without.

_vfs_context_kernel
Comment: Used at times when it is correct to use kernel context. Like hfs.kext does as well.
Alternative: None
Result: No idea, seems to run any way

_vnode_lookupat

Comment: Look up a "file" from a directory is useful.
Alternative: Implemented with vn_getpath() strcat("/file" and vnode_lookup().
Result: Some string work in kernel, not "that" expensive. Still, vnode_lookupat would be nice.

_kvtophys
__mh_execute_header
_gLoadedKextSummaries

Comment: Used only in debug/panic. To lookup symbols when dumping stack.
Alternatives: None, and even harder with arm64's PAC. Can live without.
Result: Stack dumps with addresses only, use lldb to resolve symbols.

_kauth_cred_getgroups

Comment: Getting a list of other groups, to check for permission.
Alternative: Get list of permissions, and compare with _kauth_cred_ismember_gid
Result: No difference, just #ifdef in upstream

_cpuid_info

Comment: Get CPU features, sse, avx, aes etc.
Alternative: inline asm { cpuid }, and create own bit checks.
Result: None, just a shame to not be able to use XNU existing API.

 thread_set_cpulimit

Comment: thread_set_cpulimit() is used to emulate Illumos's duty cycle, and asks xnu to limit a thread's scheduling 
to less than 100% of a cpu : if over a period of time the thread uses more than that, it is put to sleep 
for a period before being allowed to continue. The sysdc thread is put into this mode in spa.c. This thread is bursty,
and trips the cpu limit from time to time.
Alternative: none
Result: We have lived (and so can keep living) without it in 2.x, but would be nice to have for smoothness

throttle_set_thread_io_policy

Comment: throttle_set_io_policy() lets us passivate our threads so that if they generate IOKit calls they do not get 
counted by the IO throttle. By default, kernel threads have high priority, so this thread serves to improve the performance
of other I/O versus hdiutil-attached images in datasets (which otherwise are promoted from passive to high priority), 
and zvols with HFS or APFS inside.
Alternative: none
Result: We have lived (and so can keep living) without it in 2.x, but would be nice to have for smoothness