Re: [PATCH v12 18/25] irqchip: Add RISC-V incoming MSI controller early driver

[Anup Patel]

On Sat, Feb 17, 2024 at 12:10 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>
> On Sat, Jan 27 2024 at 21:47, Anup Patel wrote:
> > +
> > +#ifdef CONFIG_SMP
> > +static irqreturn_t imsic_local_sync_handler(int irq, void *data)
> > +{
> > +     imsic_local_sync();
> > +     return IRQ_HANDLED;
> > +}
> > +
> > +static void imsic_ipi_send(unsigned int cpu)
> > +{
> > +     struct imsic_local_config *local =
> > +                             per_cpu_ptr(imsic->global.local, cpu);
>
> Let it stick out. We switched to line length 100 quite some time
> ago. Applies to the rest of the series too.

Okay, I will update.

>
> > +     writel_relaxed(IMSIC_IPI_ID, local->msi_va);
> > +}
> > +
> > +static void imsic_ipi_starting_cpu(void)
> > +{
> > +     /* Enable IPIs for current CPU. */
> > +     __imsic_id_set_enable(IMSIC_IPI_ID);
> > +
> > +     /* Enable virtual IPI used for IMSIC ID synchronization */
> > +     enable_percpu_irq(imsic->ipi_virq, 0);
> > +}
> > +
> > +static void imsic_ipi_dying_cpu(void)
> > +{
> > +     /*
> > +      * Disable virtual IPI used for IMSIC ID synchronization so
> > +      * that we don't receive ID synchronization requests.
> > +      */
> > +     disable_percpu_irq(imsic->ipi_virq);
>
> Shouldn't this disable the hardware too, i.e.
>
>           __imsic_id_clear_enable()
>
> ?

Yes, it should but somehow I missed and never saw any issue.

I will update.

>
> > +}
> > +
> > +static int __init imsic_ipi_domain_init(void)
> > +{
> > +     int virq;
> > +
> > +     /* Create IMSIC IPI multiplexing */
> > +     virq = ipi_mux_create(IMSIC_NR_IPI, imsic_ipi_send);
> > +     if (virq <= 0)
> > +             return (virq < 0) ? virq : -ENOMEM;
> > +     imsic->ipi_virq = virq;
> > +
> > +     /* First vIRQ is used for IMSIC ID synchronization */
> > +     virq = request_percpu_irq(imsic->ipi_virq, imsic_local_sync_handler,
> > +                               "riscv-imsic-lsync", imsic->global.local);
> > +     if (virq)
> > +             return virq;
>
> Please use a separate 'ret' variable. I had to read this 3 times to make
> sense of it.

Okay, I will update.

>
> > +     irq_set_status_flags(imsic->ipi_virq, IRQ_HIDDEN);
> > +     imsic->ipi_lsync_desc = irq_to_desc(imsic->ipi_virq);
>
> What's so special about this particular IPI that it can't be handled
> like all the other IPIs?

We are using this special under-the-hood IPI for synchronization
of IRQ enable/disable and IRQ movement across CPUs.

x86 has a more lazy approach of using a per-CPU timer so in
the next revision I will move to a similar approach. This means
both "ipi_virq" and "ipi_lsync_desc" will go away.

>
> > +static int __init imsic_early_probe(struct fwnode_handle *fwnode)
> > +{
> > +     int rc;
> > +     struct irq_domain *domain;
>
> https://www.kernel.org/doc/html/latest/process/maintainer-tip.html#variable-declarations

Okay, I will update.

>
> > +
> > +     /* Find parent domain and register chained handler */
> > +     domain = irq_find_matching_fwnode(riscv_get_intc_hwnode(),
> > +                                       DOMAIN_BUS_ANY);
> > +     if (!domain) {
> > +             pr_err("%pfwP: Failed to find INTC domain\n", fwnode);
> > +             return -ENOENT;
> > +     }
> > +     imsic_parent_irq = irq_create_mapping(domain, RV_IRQ_EXT);
> > +     if (!imsic_parent_irq) {
> > +             pr_err("%pfwP: Failed to create INTC mapping\n", fwnode);
> > +             return -ENOENT;
> > +     }
> > +     irq_set_chained_handler(imsic_parent_irq, imsic_handle_irq);
> > +
> > +     /* Initialize IPI domain */
> > +     rc = imsic_ipi_domain_init();
> > +     if (rc) {
> > +             pr_err("%pfwP: Failed to initialize IPI domain\n", fwnode);
> > +             return rc;
>
> Leaves the chained handler around and enabled.

Okay, I will set the chained hander after imsic_ipi_domain_init().

>
> > diff --git a/drivers/irqchip/irq-riscv-imsic-state.c b/drivers/irqchip/irq-riscv-imsic-state.c
> > +
> > +#define imsic_csr_write(__c, __v)            \
> > +do {                                         \
> > +     csr_write(CSR_ISELECT, __c);            \
> > +     csr_write(CSR_IREG, __v);               \
> > +} while (0)
>
> Any reason why these macros can't be inlines?

No particular reason. I am fine with both maros and inline functions.

I will update in the next revision.

>
> > +const struct imsic_global_config *imsic_get_global_config(void)
> > +{
> > +     return imsic ? &imsic->global : NULL;
> > +}
> > +EXPORT_SYMBOL_GPL(imsic_get_global_config);
>
> Why is this exported?

This is for the KVM RISC-V module. We have follow up
KVM RISC-V patchs which need to know the IMSIC global
configuration so that it can assign IMSIC guest files to a
Guest/VM.

>
> > +#define __imsic_id_read_clear_enabled(__id)          \
> > +     __imsic_eix_read_clear((__id), false)
> > +#define __imsic_id_read_clear_pending(__id)          \
> > +     __imsic_eix_read_clear((__id), true)
>
> Please use inlines.

Okay, I will update.

>
> > +void __imsic_eix_update(unsigned long base_id,
> > +                     unsigned long num_id, bool pend, bool val)
> > +{
> > +     unsigned long i, isel, ireg;
> > +     unsigned long id = base_id, last_id = base_id + num_id;
> > +
> > +     while (id < last_id) {
> > +             isel = id / BITS_PER_LONG;
> > +             isel *= BITS_PER_LONG / IMSIC_EIPx_BITS;
> > +             isel += (pend) ? IMSIC_EIP0 : IMSIC_EIE0;
> > +
> > +             ireg = 0;
> > +             for (i = id & (__riscv_xlen - 1);
> > +                  (id < last_id) && (i < __riscv_xlen); i++) {
> > +                     ireg |= BIT(i);
> > +                     id++;
> > +             }
>
> This lacks a comment what this is doing.

Okay, I will add a comment block.

>
> > +
> > +             /*
> > +              * The IMSIC EIEx and EIPx registers are indirectly
> > +              * accessed via using ISELECT and IREG CSRs so we
> > +              * need to access these CSRs without getting preempted.
> > +              *
> > +              * All existing users of this function call this
> > +              * function with local IRQs disabled so we don't
> > +              * need to do anything special here.
> > +              */
> > +             if (val)
> > +                     imsic_csr_set(isel, ireg);
> > +             else
> > +                     imsic_csr_clear(isel, ireg);
> > +     }
> > +}
> > +
> > +void imsic_local_sync(void)
> > +{
> > +     struct imsic_local_priv *lpriv = this_cpu_ptr(imsic->lpriv);
> > +     struct imsic_local_config *mlocal;
> > +     struct imsic_vector *mvec;
> > +     unsigned long flags;
> > +     int i;
> > +
> > +     raw_spin_lock_irqsave(&lpriv->ids_lock, flags);
> > +     for (i = 1; i <= imsic->global.nr_ids; i++) {
> > +             if (i == IMSIC_IPI_ID)
> > +                     continue;
> > +
> > +             if (test_bit(i, lpriv->ids_enabled_bitmap))
> > +                     __imsic_id_set_enable(i);
> > +             else
> > +                     __imsic_id_clear_enable(i);
> > +
> > +             mvec = lpriv->ids_move[i];
> > +             lpriv->ids_move[i] = NULL;
> > +             if (mvec) {
> > +                     if (__imsic_id_read_clear_pending(i)) {
> > +                             mlocal = per_cpu_ptr(imsic->global.local,
> > +                                                  mvec->cpu);
> > +                             writel_relaxed(mvec->local_id, mlocal->msi_va);
> > +                     }
> > +
> > +                     imsic_vector_free(&lpriv->vectors[i]);
> > +             }
>
> Again an uncommented piece of magic which you will have forgotten what
> it does 3 month down the road :)

Sure, I will add a comment block.

>
> > +
> > +     }
> > +     raw_spin_unlock_irqrestore(&lpriv->ids_lock, flags);
> > +}
> > +
> > +void imsic_local_delivery(bool enable)
> > +{
> > +     if (enable) {
> > +             imsic_csr_write(IMSIC_EITHRESHOLD, IMSIC_ENABLE_EITHRESHOLD);
> > +             imsic_csr_write(IMSIC_EIDELIVERY, IMSIC_ENABLE_EIDELIVERY);
> > +             return;
> > +     }
> > +
> > +     imsic_csr_write(IMSIC_EIDELIVERY, IMSIC_DISABLE_EIDELIVERY);
> > +     imsic_csr_write(IMSIC_EITHRESHOLD, IMSIC_DISABLE_EITHRESHOLD);
> > +}
> > +
> > +#ifdef CONFIG_SMP
> > +static void imsic_remote_sync(unsigned int cpu)
> > +{
> > +     /*
> > +      * We simply inject ID synchronization IPI to a target CPU
> > +      * if it is not same as the current CPU. The ipi_send_mask()
> > +      * implementation of IPI mux will inject ID synchronization
> > +      * IPI only for CPUs that have enabled it so offline CPUs
> > +      * won't receive IPI. An offline CPU will unconditionally
> > +      * synchronize IDs through imsic_starting_cpu() when the
> > +      * CPU is brought up.
> > +      */
> > +     if (cpu_online(cpu)) {
> > +             if (cpu != smp_processor_id())
> > +                     __ipi_send_mask(imsic->ipi_lsync_desc, cpumask_of(cpu));
>
> Still wondering why this can't use the regular API. There might be a
> reason, but then it wants to be documented.

As mentioned above, the "ipi_virq" and "irq_lsync_desc" will
be replaced by a per-CPU timer in the next revision.

>
> > +             else
> > +                     imsic_local_sync();
> > +     }
> > +}
> > +#else
> > +static inline void imsic_remote_sync(unsigned int cpu)
> > +{
> > +     imsic_local_sync();
> > +}
> > +#endif
> > +
> > +void imsic_vector_mask(struct imsic_vector *vec)
> > +{
> > +     struct imsic_local_priv *lpriv;
> > +     unsigned long flags;
> > +
> > +     lpriv = per_cpu_ptr(imsic->lpriv, vec->cpu);
> > +     if (WARN_ON(&lpriv->vectors[vec->local_id] != vec))
> > +             return;
> > +
> > +     raw_spin_lock_irqsave(&lpriv->ids_lock, flags);
>
> AFAICT, this is used from an irqchip callback:
>
> static void imsic_irq_mask(struct irq_data *d)
> {
>         imsic_vector_mask(irq_data_get_irq_chip_data(d));
> }
>
> So no need to use irqsave() here. Those callbacks run always with
> interrupts disabled when called from the core.

Okay, I will update.

>
> > +void imsic_vector_move(struct imsic_vector *old_vec,
> > +                     struct imsic_vector *new_vec)
> > +{
> > +     struct imsic_local_priv *old_lpriv, *new_lpriv;
> > +     unsigned long flags, flags1;
> > +
> > +     if (WARN_ON(old_vec->cpu == new_vec->cpu))
> > +             return;
> > +
> > +     old_lpriv = per_cpu_ptr(imsic->lpriv, old_vec->cpu);
> > +     if (WARN_ON(&old_lpriv->vectors[old_vec->local_id] != old_vec))
> > +             return;
> > +
> > +     new_lpriv = per_cpu_ptr(imsic->lpriv, new_vec->cpu);
> > +     if (WARN_ON(&new_lpriv->vectors[new_vec->local_id] != new_vec))
> > +             return;
> > +
> > +     raw_spin_lock_irqsave(&old_lpriv->ids_lock, flags);
> > +     raw_spin_lock_irqsave(&new_lpriv->ids_lock, flags1);
>
> Lockdep should yell at you for this, rightfully so. And not only because
> of the missing nested() annotation.
>
> Assume there are two CPUs setting affinity for two different interrupts.
>
> CPU0 moves an interrupt to CPU1 and CPU1 moves another interrupt to
> CPU0. The resulting lock order is:
>
> CPU0                     CPU1
> lock(lpriv[CPU0]);       lock(lpriv[CPU1]);
> lock(lpriv[CPU1]);       lock(lpriv[CPU0]);
>
> a classic ABBA deadlock.
>
> You need to take those locks always in the same order. Look at
> double_raw_lock() in kernel/sched/sched.h.

I have simplified the locking to avoid this nested locks so this
will be much simpler without any lock nesting.

>
> > +     /* Unmask the new vector entry */
> > +     if (test_bit(old_vec->local_id, old_lpriv->ids_enabled_bitmap))
> > +             bitmap_set(new_lpriv->ids_enabled_bitmap,
> > +                        new_vec->local_id, 1);
>
> Either make that one line or please add brackets. See:
>
>   https://www.kernel.org/doc/html/latest/process/maintainer-tip.html#bracket-rules

Okay, I will update.

>
> > +static int __init imsic_local_init(void)
> > +{
> > +     struct imsic_global_config *global = &imsic->global;
> > +     struct imsic_local_priv *lpriv;
> > +     struct imsic_vector *vec;
> > +     int cpu, i;
> > +
> > +     /* Allocate per-CPU private state */
> > +     imsic->lpriv = alloc_percpu(typeof(*(imsic->lpriv)));
> > +     if (!imsic->lpriv)
> > +             return -ENOMEM;
> > +
> > +     /* Setup per-CPU private state */
> > +     for_each_possible_cpu(cpu) {
> > +             lpriv = per_cpu_ptr(imsic->lpriv, cpu);
> > +
> > +             raw_spin_lock_init(&lpriv->ids_lock);
> > +
> > +             /* Allocate enabled bitmap */
> > +             lpriv->ids_enabled_bitmap = bitmap_zalloc(global->nr_ids + 1,
> > +                                                       GFP_KERNEL);
> > +             if (!lpriv->ids_enabled_bitmap) {
> > +                     imsic_local_cleanup();
> > +                     return -ENOMEM;
> > +             }
> > +
> > +             /* Allocate move array */
> > +             lpriv->ids_move = kcalloc(global->nr_ids + 1,
> > +                                     sizeof(*lpriv->ids_move), GFP_KERNEL);
> > +             if (!lpriv->ids_move) {
> > +                     imsic_local_cleanup();
> > +                     return -ENOMEM;
> > +             }
> > +
> > +             /* Allocate vector array */
> > +             lpriv->vectors = kcalloc(global->nr_ids + 1,
> > +                                      sizeof(*lpriv->vectors), GFP_KERNEL);
> > +             if (!lpriv->vectors) {
> > +                     imsic_local_cleanup();
> > +                     return -ENOMEM;
>
> Third instance of the same pattern. goto cleanup; perhaps?

Okay, I will add goto here.

>
> > +struct imsic_vector *imsic_vector_alloc(unsigned int hwirq,
> > +                                     const struct cpumask *mask)
> > +{
> > +     struct imsic_vector *vec = NULL;
> > +     struct imsic_local_priv *lpriv;
> > +     unsigned long flags;
> > +     unsigned int cpu;
> > +     int local_id;
> > +
> > +     raw_spin_lock_irqsave(&imsic->matrix_lock, flags);
> > +     local_id = irq_matrix_alloc(imsic->matrix, mask, false, &cpu);
> > +     raw_spin_unlock_irqrestore(&imsic->matrix_lock, flags);
> > +     if (local_id < 0)
> > +             return NULL;
> > +
> > +     lpriv = per_cpu_ptr(imsic->lpriv, cpu);
> > +     vec = &lpriv->vectors[local_id];
> > +     vec->hwirq = hwirq;
> > +
> > +     return vec;
> > +}
>
> ...
>
> > +int imsic_hwirq_alloc(void)
> > +{
> > +     int ret;
> > +     unsigned long flags;
> > +
> > +     raw_spin_lock_irqsave(&imsic->hwirqs_lock, flags);
> > +     ret = bitmap_find_free_region(imsic->hwirqs_used_bitmap,
> > +                                   imsic->nr_hwirqs, 0);
> > +     raw_spin_unlock_irqrestore(&imsic->hwirqs_lock, flags);
> > +
> > +     return ret;
> > +}
>
> This part is just to create a unique hwirq number, right?

Yes, this is only for unique hwirq. We can directly use virq
instead of hwirq so this hwirq allocation/management will
go away in the next revision.

>
> > +
> > +     /* Find number of guest index bits in MSI address */
> > +     rc = of_property_read_u32(to_of_node(fwnode),
> > +                               "riscv,guest-index-bits",
> > +                               &global->guest_index_bits);
> > +     if (rc)
> > +             global->guest_index_bits = 0;
>
> So here you get the index bits, but then 50 lines further down you do
> sanity checking. Wouldn't it make sense to do that right here?
>
> Same for the other bits.

This is intentional because we already have a AIA ACPI series
where this helps to reduce the number of "if (acpi_disabled)"
checks.

>
> > +
> > +/*
> > + * The IMSIC driver uses 1 IPI for ID synchronization and
> > + * arch/riscv/kernel/smp.c require 6 IPIs so we fix the
> > + * total number of IPIs to 8.
> > + */
> > +#define IMSIC_IPI_ID                         1
> > +#define IMSIC_NR_IPI                         8
> > +
> > +struct imsic_vector {
> > +     /* Fixed details of the vector */
> > +     unsigned int cpu;
> > +     unsigned int local_id;
> > +     /* Details saved by driver in the vector */
> > +     unsigned int hwirq;
> > +};
> > +
> > +struct imsic_local_priv {
> > +     /* Local state of interrupt identities */
> > +     raw_spinlock_t ids_lock;
> > +     unsigned long *ids_enabled_bitmap;
> > +     struct imsic_vector **ids_move;
> > +
> > +     /* Local vector table */
> > +     struct imsic_vector *vectors;
>
> Please make those structs tabular:
>
> https://www.kernel.org/doc/html/latest/process/maintainer-tip.html#struct-declarations-and-initializers

Okay, I will update.

>
> > +void __imsic_eix_update(unsigned long base_id,
> > +                     unsigned long num_id, bool pend, bool val);
> > +
> > +#define __imsic_id_set_enable(__id)          \
> > +     __imsic_eix_update((__id), 1, false, true)
> > +#define __imsic_id_clear_enable(__id)        \
> > +     __imsic_eix_update((__id), 1, false, false)
>
> inlines please.

Okay, I will update.

Regards,
Anup