diff options
Diffstat (limited to 'kernel/gsp_rm_emulation.c')
| -rw-r--r-- | kernel/gsp_rm_emulation.c | 42 |
1 files changed, 27 insertions, 15 deletions
diff --git a/kernel/gsp_rm_emulation.c b/kernel/gsp_rm_emulation.c index 764ef05..cd9cbf2 100644 --- a/kernel/gsp_rm_emulation.c +++ b/kernel/gsp_rm_emulation.c @@ -571,27 +571,39 @@ static int gsp_rm_send_flags(struct gsp_emu_context *ctx, u32 function, u32 used, space, rxoff = 0; u8 shdr[32]; - if (gsp_rm_read_buf(ctx, rm->stat_off, shdr, - sizeof(shdr))) + /* + * r332 SWAP_RX cursor fix: the guest's consumption + * cursor for the STATUS queue lives in the COMMAND + * queue's RX header, not in the status queue's own + * rxHdrOff. Under MSGQ_FLAGS_SWAP_RX (both sides set + * it: flags=1 in both headers) the driver's + * msgqRxMarkConsumed writes its status readPtr to + * pOurRxHdr->readPtr, and the driver's pOurRxHdr is + * the COMMAND-queue buffer's RX header (msgq.c + * msgqRxLink swap branch: pReadOutgoing = + * &pOurRxHdr->readPtr with pOurRxHdr from + * msgqTxCreate's pBackingStore = pCommandQueue). The + * status queue's own +rxHdrOff is the mirror we + * publish: OUR command-queue read position + * (written below after every consumed command + * record). Reading that mirror here made the + * free-space math track our own writes: the ring was + * declared "full" with the guest mid-drain, replies + * were written into it anyway, unread records were + * overwritten, and the guest RM read a newer record + * ("Bad sequence number. Expected N got N+63") - the + * first-client cudaMalloc then fails with "device + * busy or unavailable" (the boot lottery) and every + * later RPC fails until a fresh boot. + */ + if (gsp_rm_read_buf(ctx, rm->cmd_off, shdr, sizeof(shdr))) break; memcpy(&rxoff, shdr + 24, 4); if (!rxoff || rxoff >= 0x1000) break; - if (gsp_rm_read_buf(ctx, rm->stat_off + rxoff, - &peer_read, 4) || + if (gsp_rm_read_buf(ctx, rm->cmd_off + rxoff, &peer_read, 4) || peer_read >= rm->stat_msgcount) break; /* no usable cursor: blind write */ - /* - * fail4: only a cursor the guest is actually BEHIND - * (peer <= our writePtr) is a consumption cursor. - * The observed value at stat+rxHdrOff under this - * queue's flags tracks writePtr+1 (a mirror), and - * gating on it silenced every wake-up event and - * wedged the clients. A mirror (peer > write) falls - * back to the historical blind write. - */ - if (peer_read > rm->stat_write) - break; have_peer = true; used = (rm->stat_write + rm->stat_msgcount - peer_read) % rm->stat_msgcount; |