libvhost-user: fix postcopy shared memory faults
cover libvhost-user: fix postcopy shared memory faults
These two independent bugs both sit on the libvhost-user postcopy path and together make postcopy migration of a guest with a libvhost-user backend fail. They are only reachable when the backend negotiates VHOST_USER_PROTOCOL_F_CONFIGURE_MEM_SLOTS, i.e. it adds memory regions one by one with VHOST_USER_ADD_MEM_REG instead of VHOST_USER_SET_MEM_TABLE.
Patch 1 stops the backend from panicking on the "all postcopy client bases received" ack, which QEMU sends as an fd-less VHOST_USER_ADD_MEM_REG.
Patch 2 makes the backend return its mapping address for a region added that way, without which QEMU cannot translate the backend's fault addresses back to a RAMBlock.
Reproduced and fixed with contrib/vhost-user-bridge as the net backend of a postcopy migrated guest: before the series the destination backend either panics or QEMU reports "Failed to find region for fault"; after it the migration completes.
Bin Guo (2):
libvhost-user: accept the postcopy client base ack in vu_add_mem_reg()
libvhost-user: return the backend mapping address for added regions
subprojects/libvhost-user/libvhost-user.c | 34 ++++++++++++++--------- 1 file changed, 21 insertions(+), 13 deletions(-)
Signature · 1 line
2.50.1 (Apple Git-155)
PATCH 1/2 libvhost-user: accept the postcopy client base ack in vu_add_mem_reg()
In postcopy mode QEMU signals that it has collected all the postcopy client bases by sending a VHOST_USER_ADD_MEM_REG message with a u64 payload of 0 and no file descriptor (see vhost_user_add_remove_regions()). vu_add_mem_reg() has a case for that message, but only reaches it after validating the fd count of a regular region, so the ack is rejected first:
VHOST_USER_ADD_MEM_REG received 0 fds - only 1 fd should be sent for this message type
This kills the backend during memory table setup, making postcopy unusable for any libvhost-user backend that negotiates VHOST_USER_PROTOCOL_F_CONFIGURE_MEM_SLOTS.
Recognise the ack before validating the fd count.
--- subprojects/libvhost-user/libvhost-user.c | 26 ++++++++++++----------- 1 file changed, 14 insertions(+), 12 deletions(-)
subprojects/libvhost-user/libvhost-user.c +14 -12
@@ -948,6 +948,20 @@ static bool vu_add_mem_reg(VuDev *dev, VhostUserMsg *vmsg) {VhostUserMemoryRegion m = vmsg->payload.memreg.region, *msg_region = &m; +/* +* If we are in postcopy mode and we receive a u64 payload with a 0 value +* we know all the postcopy client bases have been received, and we +* should start generating faults. This message carries no file +* descriptor, so it has to be recognised before the fd count of a real +* region is validated below. +*/ +if (dev->postcopy_listening && + vmsg->size ==sizeof (vmsg->payload.u64) && + vmsg->payload.u64 ==0 ) { + (void )generate_faults (dev); +return false; + } +if (vmsg->fd_num !=1 ) {vmsg_close_fds (vmsg); vu_panic(dev, "VHOST_USER_ADD_MEM_REG received %d fds - only 1 fd "
@@ -971,18 +985,6 @@ vu_add_mem_reg(VuDev *dev, VhostUserMsg *vmsg) {return false; } -/* -* If we are in postcopy mode and we receive a u64 payload with a 0 value -* we know all the postcopy client bases have been received, and we -* should start generating faults. -*/ -if (dev->postcopy_listening && - vmsg->size ==sizeof (vmsg->payload.u64) && - vmsg->payload.u64 ==0 ) { - (void )generate_faults (dev); -return false; - } -_vu_add_mem_reg (dev, msg_region, vmsg->fds[0 ]);close (vmsg->fds[0 ]);
Signature · 1 line
2.50.1 (Apple Git-155)
PATCH 2/2 libvhost-user: return the backend mapping address for added regions
In reply to Bin GuoIn postcopy mode the backend must tell QEMU where it mapped a region, so that QEMU can translate the backend's fault addresses back into a RAMBlock and offset. vu_add_mem_reg() works on a copy of the region and lets _vu_add_mem_reg() fill the mapping address into that copy, but never writes it back into the reply payload. QEMU then stores its own address as the postcopy client base, so fault resolution fails:
vhost_user_postcopy_fault_handler: Failed to find region for fault ...
The VHOST_USER_SET_MEM_TABLE path is unaffected, it fills the payload directly. Put the updated region back into the payload before replying; a pointer into the payload cannot be used instead, as VhostUserMsg is packed.
--- subprojects/libvhost-user/libvhost-user.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-)
subprojects/libvhost-user/libvhost-user.c +7 -1
@@ -989,7 +989,13 @@ vu_add_mem_reg(VuDev *dev, VhostUserMsg *vmsg) {close (vmsg->fds[0 ]);if (dev->postcopy_listening) { -/* Send the message back to qemu with the addresses filled in. */ +/* +* Send the message back to qemu with the addresses filled in. +* _vu_add_mem_reg() worked on our copy of the region, so it has to be +* put back into the message payload. A pointer into the payload +* cannot be handed out instead, VhostUserMsg is packed. +*/ + vmsg->payload.memreg.region = m; vmsg->fd_num =0 ;DPRINT ("Successfully added new region in postcopy\n" );return true;
Signature · 1 line
2.50.1 (Apple Git-155)