Re: [kvm-unit-tests PATCH v2 1/6] s390x: Use interrupts in SCLP and add locking

[David Hildenbrand <david@redhat.com>]

On 05.09.19 12:39, Janosch Frank wrote:
> We need to properly implement interrupt handling for SCLP, because on
> z/VM and LPAR SCLP calls are not synchronous!
> 
> Also with smp CPUs have to compete for sclp. Let's add some locking,
> so they execute sclp calls in an orderly fashion and don't compete for
> the data buffer.
> 
> Signed-off-by: Janosch Frank <frankja@linux.ibm.com>
> ---
>  lib/s390x/asm/interrupt.h |  2 ++
>  lib/s390x/interrupt.c     | 12 +++++++--
>  lib/s390x/sclp-console.c  |  2 ++
>  lib/s390x/sclp.c          | 55 +++++++++++++++++++++++++++++++++++++--
>  lib/s390x/sclp.h          |  3 +++
>  5 files changed, 70 insertions(+), 4 deletions(-)
> 
> diff --git a/lib/s390x/asm/interrupt.h b/lib/s390x/asm/interrupt.h
> index 013709f..f485e96 100644
> --- a/lib/s390x/asm/interrupt.h
> +++ b/lib/s390x/asm/interrupt.h
> @@ -11,6 +11,8 @@
>  #define _ASMS390X_IRQ_H_
>  #include <asm/arch_def.h>
>  
> +#define EXT_IRQ_SERVICE_SIG	0x2401
> +
>  void handle_pgm_int(void);
>  void handle_ext_int(void);
>  void handle_mcck_int(void);
> diff --git a/lib/s390x/interrupt.c b/lib/s390x/interrupt.c
> index cf0a794..7832711 100644
> --- a/lib/s390x/interrupt.c
> +++ b/lib/s390x/interrupt.c
> @@ -12,6 +12,7 @@
>  #include <libcflat.h>
>  #include <asm/interrupt.h>
>  #include <asm/barrier.h>
> +#include <sclp.h>
>  
>  static bool pgm_int_expected;
>  static struct lowcore *lc;
> @@ -107,8 +108,15 @@ void handle_pgm_int(void)
>  
>  void handle_ext_int(void)
>  {
> -	report_abort("Unexpected external call interrupt: at %#lx",
> -		     lc->ext_old_psw.addr);
> +	if (lc->ext_int_code != EXT_IRQ_SERVICE_SIG) {
> +		report_abort("Unexpected external call interrupt: at %#lx",
> +			     lc->ext_old_psw.addr);
> +	} else {
> +		lc->ext_old_psw.mask &= ~PSW_MASK_EXT;
> +		lc->sw_int_cr0 &= ~(1UL << 9);
> +		sclp_handle_ext();
> +		lc->ext_int_code = 0;
> +	}
>  }
>  
>  void handle_mcck_int(void)
> diff --git a/lib/s390x/sclp-console.c b/lib/s390x/sclp-console.c
> index bc01f41..a5ef45f 100644
> --- a/lib/s390x/sclp-console.c
> +++ b/lib/s390x/sclp-console.c
> @@ -17,6 +17,7 @@ static void sclp_set_write_mask(void)
>  {
>  	WriteEventMask *sccb = (void *)_sccb;
>  
> +	sclp_mark_busy();
>  	sccb->h.length = sizeof(WriteEventMask);
>  	sccb->mask_length = sizeof(unsigned int);
>  	sccb->receive_mask = SCLP_EVENT_MASK_MSG_ASCII;
> @@ -37,6 +38,7 @@ void sclp_print(const char *str)
>  	int len = strlen(str);
>  	WriteEventData *sccb = (void *)_sccb;
>  
> +	sclp_mark_busy();
>  	sccb->h.length = sizeof(WriteEventData) + len;
>  	sccb->h.function_code = SCLP_FC_NORMAL_WRITE;
>  	sccb->ebh.length = sizeof(EventBufferHeader) + len;
> diff --git a/lib/s390x/sclp.c b/lib/s390x/sclp.c
> index b60f7a4..56fca0c 100644
> --- a/lib/s390x/sclp.c
> +++ b/lib/s390x/sclp.c
> @@ -14,6 +14,8 @@
>  #include <asm/page.h>
>  #include <asm/arch_def.h>
>  #include <asm/interrupt.h>
> +#include <asm/barrier.h>
> +#include <asm/spinlock.h>
>  #include "sclp.h"
>  #include <alloc_phys.h>
>  #include <alloc_page.h>
> @@ -25,6 +27,8 @@ static uint64_t max_ram_size;
>  static uint64_t ram_size;
>  
>  char _sccb[PAGE_SIZE] __attribute__((__aligned__(4096)));
> +static volatile bool sclp_busy;
> +static struct spinlock sclp_lock;
>  
>  static void mem_init(phys_addr_t mem_end)
>  {
> @@ -41,17 +45,62 @@ static void mem_init(phys_addr_t mem_end)
>  	page_alloc_ops_enable();
>  }
>  
> +static void sclp_setup_int(void)
> +{
> +	uint64_t mask;
> +
> +	ctl_set_bit(0, 9);
> +
> +	mask = extract_psw_mask();
> +	mask |= PSW_MASK_EXT;
> +	load_psw_mask(mask);
> +}
> +
> +void sclp_handle_ext(void)
> +{
> +	ctl_clear_bit(0, 9);
> +	spin_lock(&sclp_lock);
> +	sclp_busy = false;
> +	spin_unlock(&sclp_lock);
> +}
> +
> +void sclp_wait_busy(void)
> +{
> +	while (sclp_busy)
> +		mb();
> +}
> +
> +void sclp_mark_busy(void)
> +{
> +	/*
> +	 * With multiple CPUs we might need to wait for another CPU's
> +	 * request before grabbing the busy indication.
> +	 */
> +	while (true) {
> +		sclp_wait_busy();
> +		spin_lock(&sclp_lock);
> +		if (!sclp_busy) {
> +			sclp_busy = true;
> +			spin_unlock(&sclp_lock);
> +			return;
> +		}
> +		spin_unlock(&sclp_lock);
> +	}
> +}
> +
>  static void sclp_read_scp_info(ReadInfo *ri, int length)
>  {
>  	unsigned int commands[] = { SCLP_CMDW_READ_SCP_INFO_FORCED,
>  				    SCLP_CMDW_READ_SCP_INFO };
> -	int i;
> +	int i, cc;
>  
>  	for (i = 0; i < ARRAY_SIZE(commands); i++) {
> +		sclp_mark_busy();
>  		memset(&ri->h, 0, sizeof(ri->h));
>  		ri->h.length = length;
>  
> -		if (sclp_service_call(commands[i], ri))
> +		cc = sclp_service_call(commands[i], ri);
> +		if (cc)
>  			break;
>  		if (ri->h.response_code == SCLP_RC_NORMAL_READ_COMPLETION)
>  			return;
> @@ -66,12 +115,14 @@ int sclp_service_call(unsigned int command, void *sccb)
>  {
>  	int cc;
>  
> +	sclp_setup_int();
>  	asm volatile(
>  		"       .insn   rre,0xb2200000,%1,%2\n"  /* servc %1,%2 */
>  		"       ipm     %0\n"
>  		"       srl     %0,28"
>  		: "=&d" (cc) : "d" (command), "a" (__pa(sccb))
>  		: "cc", "memory");
> +	sclp_wait_busy();
>  	if (cc == 3)
>  		return -1;
>  	if (cc == 2)
> diff --git a/lib/s390x/sclp.h b/lib/s390x/sclp.h
> index 583c4e5..63cf609 100644
> --- a/lib/s390x/sclp.h
> +++ b/lib/s390x/sclp.h
> @@ -213,6 +213,9 @@ typedef struct ReadEventData {
>  } __attribute__((packed)) ReadEventData;
>  
>  extern char _sccb[];
> +void sclp_handle_ext(void);
> +void sclp_wait_busy(void);
> +void sclp_mark_busy(void);
>  void sclp_console_setup(void);
>  void sclp_print(const char *str);
>  int sclp_service_call(unsigned int command, void *sccb);
> 

I was wondering whether it would make sense to enable sclp interrupts as
default for all CPUs (once in a reasonable state after brought up), and
simply let any CPU process the request. Initially, we could only let the
boot CPU handle them.

You already decoupled sclp_mark_busy() and sclp_setup_int() already. The
part would have to be moved to the CPU init stage and sclp_handle_ext()
would simply not clear the interrupt-enable flag.

Opinions?

-- 

Thanks,

David / dhildenb