| Message ID | 1571802386-2680-1-git-send-email-yuzenghui@huawei.com (mailing list archive) |
|---|---|
| State | Mainlined |
| Commit | 8424312516e5d9baeeb0a95d0e4523579b7aa395 |
| Headers | show |
| Series | [v2] irqchip/gic-v3-its: Use the exact ITSList for VMOVP | expand |
On 2019-10-23 04:46, Zenghui Yu wrote: > On a system without Single VMOVP support (say GITS_TYPER.VMOVP == 0), > we will map vPEs only on ITSs that will actually control interrupts > for the given VM. And when moving a vPE, the VMOVP command will be > issued only for those ITSs. > > But when issuing VMOVPs we seemed fail to present the exact ITSList > to ITSs who are actually included in the synchronization operation. > The its_list_map we're currently using includes all ITSs in the > system, > even though some of them don't have the corresponding vPE mapping at > all. > > Introduce get_its_list() to get the per-VM its_list_map, to indicate > which ITSs have vPE mappings for the given VM, and use this map as > the expected ITSList when building VMOVP. This is hopefully a > performance > gain not to do some synchronization with those unsuspecting ITSs. > And initialize the whole command descriptor to zero at beginning, > since > the seq_num and its_list should be RES0 when GITS_TYPER.VMOVP == 1. > > Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> I've applied this as a fix for 5.4. In the future, please cc LKML on all IRQ-related patches (as documented in MAINTAINERS). M.
On 2019/10/25 16:24, Marc Zyngier wrote: > On 2019-10-23 04:46, Zenghui Yu wrote: >> On a system without Single VMOVP support (say GITS_TYPER.VMOVP == 0), >> we will map vPEs only on ITSs that will actually control interrupts >> for the given VM. And when moving a vPE, the VMOVP command will be >> issued only for those ITSs. >> >> But when issuing VMOVPs we seemed fail to present the exact ITSList >> to ITSs who are actually included in the synchronization operation. >> The its_list_map we're currently using includes all ITSs in the system, >> even though some of them don't have the corresponding vPE mapping at all. >> >> Introduce get_its_list() to get the per-VM its_list_map, to indicate >> which ITSs have vPE mappings for the given VM, and use this map as >> the expected ITSList when building VMOVP. This is hopefully a performance >> gain not to do some synchronization with those unsuspecting ITSs. >> And initialize the whole command descriptor to zero at beginning, since >> the seq_num and its_list should be RES0 when GITS_TYPER.VMOVP == 1. >> >> Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> > > I've applied this as a fix for 5.4. In the future, please cc LKML on all > IRQ-related patches (as documented in MAINTAINERS). I got it, thanks. Zenghui
diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c index c81da27044bf..0b35fcb71203 100644 --- a/drivers/irqchip/irq-gic-v3-its.c +++ b/drivers/irqchip/irq-gic-v3-its.c @@ -175,6 +175,22 @@ static DEFINE_IDA(its_vpeid_ida); #define gic_data_rdist_rd_base() (gic_data_rdist()->rd_base) #define gic_data_rdist_vlpi_base() (gic_data_rdist_rd_base() + SZ_128K) +static u16 get_its_list(struct its_vm *vm) +{ + struct its_node *its; + unsigned long its_list = 0; + + list_for_each_entry(its, &its_nodes, entry) { + if (!its->is_v4) + continue; + + if (vm->vlpi_count[its->list_nr]) + __set_bit(its->list_nr, &its_list); + } + + return (u16)its_list; +} + static struct its_collection *dev_event_to_col(struct its_device *its_dev, u32 event) { @@ -976,17 +992,15 @@ static void its_send_vmapp(struct its_node *its, static void its_send_vmovp(struct its_vpe *vpe) { - struct its_cmd_desc desc; + struct its_cmd_desc desc = {}; struct its_node *its; unsigned long flags; int col_id = vpe->col_idx; desc.its_vmovp_cmd.vpe = vpe; - desc.its_vmovp_cmd.its_list = (u16)its_list_map; if (!its_list_map) { its = list_first_entry(&its_nodes, struct its_node, entry); - desc.its_vmovp_cmd.seq_num = 0; desc.its_vmovp_cmd.col = &its->collections[col_id]; its_send_single_vcommand(its, its_build_vmovp_cmd, &desc); return; @@ -1003,6 +1017,7 @@ static void its_send_vmovp(struct its_vpe *vpe) raw_spin_lock_irqsave(&vmovp_lock, flags); desc.its_vmovp_cmd.seq_num = vmovp_seq_num++; + desc.its_vmovp_cmd.its_list = get_its_list(vpe->its_vm); /* Emit VMOVPs */ list_for_each_entry(its, &its_nodes, entry) {
On a system without Single VMOVP support (say GITS_TYPER.VMOVP == 0), we will map vPEs only on ITSs that will actually control interrupts for the given VM. And when moving a vPE, the VMOVP command will be issued only for those ITSs. But when issuing VMOVPs we seemed fail to present the exact ITSList to ITSs who are actually included in the synchronization operation. The its_list_map we're currently using includes all ITSs in the system, even though some of them don't have the corresponding vPE mapping at all. Introduce get_its_list() to get the per-VM its_list_map, to indicate which ITSs have vPE mappings for the given VM, and use this map as the expected ITSList when building VMOVP. This is hopefully a performance gain not to do some synchronization with those unsuspecting ITSs. And initialize the whole command descriptor to zero at beginning, since the seq_num and its_list should be RES0 when GITS_TYPER.VMOVP == 1. Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> --- Changes since v1: - Initialize the whole VMOVP command descriptor to zero at beginning, since the seq_num and its_list fields should be RES0 when GITS_TYPER.VMOVP == 1. - Initialize "its_list" to zero in get_its_list(). - Skip non v4 ITSs when computing the per-VM its_list_map. - Other cleanups. drivers/irqchip/irq-gic-v3-its.c | 21 ++++++++++++++++++--- 1 file changed, 18 insertions(+), 3 deletions(-)