bhyve: ahci: wait for the commands in progress when pausing

[wanpengqian_gmail.com]

Summary

A command whose data takes more than BLOCKIF_IOV_MAX segments is done
in several block requests, the completion of one issuing the next. The
completion of a command also issues the commands that the guest queued
in the meantime. pci_ahci_pause() only called blockif_pause(), which
marks the block device as paused and waits for the requests in flight,
so such a completion issued a request on a paused block device:

Assertion failed: (!bc->bc_paused), function blockif_request, file
block_if.c, line 781.

A guest with load on an AHCI disk could not be checkpointed.

When the pause begins, stop issuing commands and wait until the ones
in progress are complete; only then pause the block devices.
pci_ahci_resume() issues the commands that were held back. They are in
PxCI, which is part of the snapshot, so they are issued after a restore
as well.

Signed-off-by: Wanpeng Qian <[e-mail]>
Sponsored by: keelos.dev

Test Plan

main (f958aa7e7) with WITH_BHYVE_SNAPSHOT, FreeBSD 16.0-CURRENT guest with a second disk -s 4,ahci-hd,/dev/md1.nop.

To get commands of more than one block request, a small program in the guest reads and writes 1 MB at random offsets of /dev/ada1 from a buffer whose 256 pages are scattered in physical memory (single-page shm objects mapped in a random order), so that each command has 256 PRDT entries; it writes a pattern, reads it back and compares. To have a request in flight when the pause begins, the disk is slow: gnop(8) with a delay of 100 ms on every read and write.

Before: the first bhyvectl --checkpoint under that load ends bhyve:

Assertion failed: (!bc->bc_paused), function blockif_request, file /usr/src/usr.sbin/bhyve/block_if.c, line 781.

After: 10 checkpoints under the same load, then bhyvectl --suspend and a restore with bhyve -r: the program keeps running in the restored guest and finds no mismatch.

The same failure was first seen on FreeBSD 14.5 with a guest that was copying files on an AHCI disk (a zvol) when it was suspended.

[markj]

usr.sbin/bhyve/pci_ahci.c

1922

We should also verify that p->pr_sc != NULL.

[wanpengqian_gmail.com]

Return early from ahci_handle_port() when the port has no softc (markj)

[wanpengqian_gmail.com]

Thanks, done: ahci_handle_port() now returns when p->pr_sc is NULL. The ports between the configured ones have no softc, and pci_ahci_port_write() takes writes to any port below sc->ports, so the new check (which runs before the one for PxCMD.ST) could dereference NULL on a write to such a port. Built with and without WITH_BHYVE_SNAPSHOT.