diff options
| author | Dean Sanner <dsanner@us.ibm.com> | 2018-02-15 07:04:43 -0600 |
|---|---|---|
| committer | William G. Hoffa <wghoffa@us.ibm.com> | 2018-02-16 10:02:29 -0500 |
| commit | 1eac1bea6c0960e47d3929f519cbb78071e40b7f (patch) | |
| tree | 72ad4a39fa44534a8e896fcf1c4583b65b1a3102 /src/kernel | |
| parent | fe7571698e857cdcafceac0f120e66293892d00e (diff) | |
| download | talos-hostboot-1eac1bea6c0960e47d3929f519cbb78071e40b7f.tar.gz talos-hostboot-1eac1bea6c0960e47d3929f519cbb78071e40b7f.zip | |
Close race condition in multi instance IPC
P9 Doorbells operate differently than P8 IPI mechanism, creating
a race condition. Basically because the HB kernel turns a
doorbell into a message, and then is ready for another one -- it
can lose the next one since repeated messages with same key are
dropped (IPC handling clears memory comm area, other HB instances
notice, fill and drop a doorbell before getting back into kernel)
Change-Id: I8eb6f010600afb9de365c8942fcd24227903436f
Reviewed-on: http://ralgit01.raleigh.ibm.com/gerrit1/54118
Reviewed-by: Prachi Gupta <pragupta@us.ibm.com>
Tested-by: Jenkins Server <pfd-jenkins+hostboot@us.ibm.com>
Tested-by: Jenkins OP Build CI <op-jenkins+hostboot@us.ibm.com>
Tested-by: Jenkins OP HW <op-hw-jenkins+hostboot@us.ibm.com>
Tested-by: FSP CI Jenkins <fsp-CI-jenkins+hostboot@us.ibm.com>
Reviewed-by: William G. Hoffa <wghoffa@us.ibm.com>
Diffstat (limited to 'src/kernel')
| -rw-r--r-- | src/kernel/intmsghandler.C | 17 |
1 files changed, 14 insertions, 3 deletions
diff --git a/src/kernel/intmsghandler.C b/src/kernel/intmsghandler.C index b4b5254a7..805b0d761 100644 --- a/src/kernel/intmsghandler.C +++ b/src/kernel/intmsghandler.C @@ -37,6 +37,7 @@ const char* VFS_ROOT_MSG_INTR = "/msg/interrupt"; InterruptMsgHdlr * InterruptMsgHdlr::cv_instance = NULL; uint64_t InterruptMsgHdlr::cv_ipc_base_address = 0; +uint64_t InterruptMsgHdlr::cv_ipc_salt = 0; void InterruptMsgHdlr::create(MessageQueue * i_msgQ, uint64_t i_ipc_addr) { @@ -193,9 +194,19 @@ void InterruptMsgHdlr::sendIpcMsg(uint64_t i_pir) { if(cv_instance) { - // To avoid conflict with interrupts on thread i_pir, change the key - // for the message to be an invalid PIR. - uint64_t pir_key = i_pir | MSG_KEY_IPC_MSG; + //Note that due to how IPC works between independent HB + //Instances, their is a race between when the data area + //"lock" is released and when the doorbell handled response + //is sent back to the kernel. Basically the other instances + //pounce on the data area as soon as it is unlocked, and + //a duplicate doorbell happens before kernel clears first + //message. + //Since the kernel will drop any message with the same PIR + //key on the floor, need to make it unique with a incrementing + //counter + + cv_ipc_salt += MSG_IPC_SALT; + uint64_t pir_key = i_pir | MSG_KEY_IPC_MSG | cv_ipc_salt; cv_instance->iv_lock.lock(); cv_instance->sendMessage(MSG_INTR_IPC, |
