[Qemu-devel] [PATCH 3/9] ehci: don't flush cache on doorbell rings.

[Gerd Hoffmann]

Commit 4be23939ab0d7019c7e59a37485b416fbbf0f073 makes ehci instantly
zap any unlinked queue heads when the guest rings the doorbell.

While hacking up uas support this turned out to be a problem.  The linux
kernel can unlink and instantly relink the very same queue head, thereby
killing any async packets in flight.  That alone isn't an issue yet, the
packet will canceled and resubmitted and everything is fine.  We'll run
into trouble though in case the async packet is completed already, so we
can't cancel it any more.  The transaction is simply lost then.

usb_ehci_qh_ptrs q (nil) - QH @ 39c4f000: next 39c4f122 qtds 
00000000,00000001,39c50000
usb_ehci_qh_fields QH @ 39c4f000 - rl 0, mplen 0, eps 0, ep 0, dev 0
usb_ehci_qh_ptrs q 0x7f95feba90a0 - QH @ 39c4f000: next 39c4f122 qtds 
00000000,00000001,39c50000
usb_ehci_qh_fields QH @ 39c4f000 - rl 0, mplen 0, eps 0, ep 0, dev 0
usb_ehci_qh_ptrs q 0x7f95fe515210 - QH @ 39c4f120: next 39c4f0c2 qtds 
29dbce40,29dbc4e0,00000009
usb_ehci_qh_fields QH @ 39c4f120 - rl 4, mplen 512, eps 2, ep 1, dev 2
usb_ehci_packet_action q 0x7f95fe515210 p 0x7f95fdec32a0: alloc
usb_packet_state_change bus 0, port 2, ep 1, packet 0x7f95fdec32e0, state undef 
-> setup
usb_ehci_packet_action q 0x7f95fe515210 p 0x7f95fdec32a0: process
usb_uas_command dev 2, tag 0x2, lun 0, lun64 00000000-00000000
scsi_req_parsed target 0 lun 0 tag 2 command 42 dir 2 length 16384
scsi_req_parsed_lba target 0 lun 0 tag 2 command 42 lba 5933312
scsi_req_alloc target 0 lun 0 tag 2
scsi_req_continue target 0 lun 0 tag 2
scsi_req_data target 0 lun 0 tag 2 len 16384
usb_uas_scsi_data dev 2, tag 0x2, bytes 16384
usb_uas_write_ready dev 2, tag 0x2
usb_packet_state_change bus 0, port 2, ep 1, packet 0x7f95fdec32e0, state setup 
-> complete
usb_ehci_packet_action q 0x7f95fe515210 p 0x7f95fdec32a0: free
usb_ehci_qh_ptrs q 0x7f95fdec3210 - QH @ 39c4f0c0: next 39c4f002 qtds 
29dbce40,00000001,00000009
usb_ehci_qh_fields QH @ 39c4f0c0 - rl 4, mplen 512, eps 2, ep 2, dev 2
usb_ehci_queue_action q 0x7f95fe5152a0: free
usb_packet_state_change bus 0, port 2, ep 2, packet 0x7f95feba9170, state async 
-> complete
^^^ async packets completes.
usb_ehci_packet_action q 0x7f95fdec3210 p 0x7f95feba9130: wakeup

usb_ehci_qh_ptrs q (nil) - QH @ 39c4f000: next 39c4f122 qtds 
00000000,00000001,39c50000
usb_ehci_qh_fields QH @ 39c4f000 - rl 0, mplen 0, eps 0, ep 0, dev 0
usb_ehci_qh_ptrs q 0x7f95feba90a0 - QH @ 39c4f000: next 39c4f122 qtds 
00000000,00000001,39c50000
usb_ehci_qh_fields QH @ 39c4f000 - rl 0, mplen 0, eps 0, ep 0, dev 0
usb_ehci_qh_ptrs q 0x7f95fe515210 - QH @ 39c4f120: next 39c4f002 qtds 
29dbc4e0,29dbc8a0,00000009
usb_ehci_qh_fields QH @ 39c4f120 - rl 4, mplen 512, eps 2, ep 1, dev 2
usb_ehci_queue_action q 0x7f95fdec3210: free
usb_ehci_packet_action q 0x7f95fdec3210 p 0x7f95feba9130: free
^^^ endpoint #2 queue head removed from schedule, doorbell makes ehci zap the 
queue,
    the (completed) usb packet is freed too and gets lost.

usb_ehci_qh_ptrs q (nil) - QH @ 39c4f000: next 39c4f0c2 qtds 
00000000,00000001,39c50000
usb_ehci_qh_fields QH @ 39c4f000 - rl 0, mplen 0, eps 0, ep 0, dev 0
usb_ehci_qh_ptrs q 0x7f95feba90a0 - QH @ 39c4f000: next 39c4f0c2 qtds 
00000000,00000001,39c50000
usb_ehci_qh_fields QH @ 39c4f000 - rl 0, mplen 0, eps 0, ep 0, dev 0
usb_ehci_queue_action q 0x7f9600dff570: alloc
usb_ehci_qh_ptrs q 0x7f9600dff570 - QH @ 39c4f0c0: next 39c4f122 qtds 
29dbce40,00000001,00000009
usb_ehci_qh_fields QH @ 39c4f0c0 - rl 4, mplen 512, eps 2, ep 2, dev 2
usb_ehci_packet_action q 0x7f9600dff570 p 0x7f95feba9130: alloc
usb_packet_state_change bus 0, port 2, ep 2, packet 0x7f95feba9170, state undef 
-> setup
usb_ehci_packet_action q 0x7f9600dff570 p 0x7f95feba9130: process
usb_packet_state_change bus 0, port 2, ep 2, packet 0x7f95feba9170, state setup 
-> async
usb_ehci_packet_action q 0x7f9600dff570 p 0x7f95feba9130: async
^^^ linux kernel relinked the queue head, ehci creates a new usb packet,
    but we should have delivered the completed one instead.
usb_ehci_qh_ptrs q 0x7f95fe515210 - QH @ 39c4f120: next 39c4f002 qtds 
29dbc4e0,29dbc8a0,00000009
usb_ehci_qh_fields QH @ 39c4f120 - rl 4, mplen 512, eps 2, ep 1, dev 2

So instead of instantly zapping the queue we'll set a flag that the
queue needs revalidation in case we'll see it again in the schedule.
ehci then checks that the queue head fields addressing / describing the
endpoint and the qtd pointer match the cached content before reusing it.

Cc: Hans de Goede <address@hidden>
Signed-off-by: Gerd Hoffmann <address@hidden>
---
 hw/usb/hcd-ehci.c |   35 +++++++++++++++++++++++++++++------
 1 files changed, 29 insertions(+), 6 deletions(-)

diff --git a/hw/usb/hcd-ehci.c b/hw/usb/hcd-ehci.c
index 1406b84..0bdfbe8 100644
--- a/hw/usb/hcd-ehci.c
+++ b/hw/usb/hcd-ehci.c
@@ -365,6 +365,7 @@ struct EHCIQueue {
     uint32_t seen;
     uint64_t ts;
     int async;
+    int revalidate;
 
     /* cached data from guest - needs to be flushed
      * when guest removes an entry (doorbell, handshake sequence)
@@ -775,7 +776,18 @@ static EHCIQueue *ehci_find_queue_by_qh(EHCIState *ehci, 
uint32_t addr,
     return NULL;
 }
 
-static void ehci_queues_rip_unused(EHCIState *ehci, int async, int flush)
+static void ehci_queues_tag_unused_async(EHCIState *ehci)
+{
+    EHCIQueue *q;
+
+    QTAILQ_FOREACH(q, &ehci->aqueues, next) {
+        if (!q->seen) {
+            q->revalidate = 1;
+        }
+    }
+}
+
+static void ehci_queues_rip_unused(EHCIState *ehci, int async)
 {
     EHCIQueueHead *head = async ? &ehci->aqueues : &ehci->pqueues;
     uint64_t maxage = FRAME_TIMER_NS * ehci->maxframes * 4;
@@ -787,7 +799,7 @@ static void ehci_queues_rip_unused(EHCIState *ehci, int 
async, int flush)
             q->ts = ehci->last_run_ns;
             continue;
         }
-        if (!flush && ehci->last_run_ns < q->ts + maxage) {
+        if (ehci->last_run_ns < q->ts + maxage) {
             continue;
         }
         ehci_free_queue(q);
@@ -1631,7 +1643,7 @@ static int ehci_state_waitlisthead(EHCIState *ehci,  int 
async)
         ehci_set_usbsts(ehci, USBSTS_REC);
     }
 
-    ehci_queues_rip_unused(ehci, async, 0);
+    ehci_queues_rip_unused(ehci, async);
 
     /*  Find the head of the list (4.9.1.1) */
     for(i = 0; i < MAX_QH; i++) {
@@ -1716,6 +1728,7 @@ static EHCIQueue *ehci_state_fetchqh(EHCIState *ehci, int 
async)
     EHCIPacket *p;
     uint32_t entry, devaddr;
     EHCIQueue *q;
+    EHCIqh qh;
 
     entry = ehci_get_fetch_addr(ehci, async);
     q = ehci_find_queue_by_qh(ehci, entry, async);
@@ -1733,7 +1746,17 @@ static EHCIQueue *ehci_state_fetchqh(EHCIState *ehci, 
int async)
     }
 
     get_dwords(ehci, NLPTR_GET(q->qhaddr),
-               (uint32_t *) &q->qh, sizeof(EHCIqh) >> 2);
+               (uint32_t *) &qh, sizeof(EHCIqh) >> 2);
+    if (q->revalidate && (q->qh.epchar      != qh.epchar ||
+                          q->qh.epcap       != qh.epcap  ||
+                          q->qh.current_qtd != qh.current_qtd)) {
+        ehci_free_queue(q);
+        q = ehci_alloc_queue(ehci, entry, async);
+        q->seen++;
+        p = NULL;
+    }
+    q->qh = qh;
+    q->revalidate = 0;
     ehci_trace_qh(q, NLPTR_GET(q->qhaddr), &q->qh);
 
     devaddr = get_field(q->qh.epchar, QH_EPCHAR_DEVADDR);
@@ -2228,7 +2251,7 @@ static void ehci_advance_async_state(EHCIState *ehci)
          */
         if (ehci->usbcmd & USBCMD_IAAD) {
             /* Remove all unseen qhs from the async qhs queue */
-            ehci_queues_rip_unused(ehci, async, 1);
+            ehci_queues_tag_unused_async(ehci);
             DPRINTF("ASYNC: doorbell request acknowledged\n");
             ehci->usbcmd &= ~USBCMD_IAAD;
             ehci_set_interrupt(ehci, USBSTS_IAA);
@@ -2281,7 +2304,7 @@ static void ehci_advance_periodic_state(EHCIState *ehci)
         ehci_set_fetch_addr(ehci, async,entry);
         ehci_set_state(ehci, async, EST_FETCHENTRY);
         ehci_advance_state(ehci, async);
-        ehci_queues_rip_unused(ehci, async, 0);
+        ehci_queues_rip_unused(ehci, async);
         break;
 
     default:
-- 
1.7.1