[PATCH v3 04/10] powerpc/pseries/vas: Reopen windows with DLPAR core add

Nicholas Piggin npiggin at gmail.com
Mon Feb 14 14:08:53 AEDT 2022


Excerpts from Haren Myneni's message of January 22, 2022 5:56 am:
> 
> VAS windows can be closed in the hypervisor due to lost credits
> when the core is removed. If these credits are available later
> for core add, reopen these windows and set them active. When the
> kernel sees page fault on the paste address, it creates new mapping
> on the new paste address. Then the user space can continue to use
> these windows and send HW compression requests to NX successfully.

Any reason to put this before the close windows patch? It would be
more logical to put it afterwards AFAIKS.

> 
> Signed-off-by: Haren Myneni <haren at linux.ibm.com>
> ---
>  arch/powerpc/include/asm/vas.h          |  16 +++
>  arch/powerpc/platforms/book3s/vas-api.c |   1 +
>  arch/powerpc/platforms/pseries/vas.c    | 144 ++++++++++++++++++++++++
>  arch/powerpc/platforms/pseries/vas.h    |   8 +-
>  4 files changed, 163 insertions(+), 6 deletions(-)
> 
> diff --git a/arch/powerpc/include/asm/vas.h b/arch/powerpc/include/asm/vas.h
> index 57573d9c1e09..f1efe86563cc 100644
> --- a/arch/powerpc/include/asm/vas.h
> +++ b/arch/powerpc/include/asm/vas.h
> @@ -29,6 +29,19 @@
>  #define VAS_THRESH_FIFO_GT_QTR_FULL	2
>  #define VAS_THRESH_FIFO_GT_EIGHTH_FULL	3
>  
> +/*
> + * VAS window status
> + */
> +#define VAS_WIN_ACTIVE		0x0	/* Used in platform independent */
> +					/* vas mmap() */
> +/* The hypervisor returns these values */
> +#define VAS_WIN_CLOSED		0x00000001
> +#define VAS_WIN_INACTIVE	0x00000002 /* Inactive due to HW failure */
> +#define VAS_WIN_MOD_IN_PROCESS	0x00000003 /* Process of being modified, */

While you're moving these and adding a comment, it would be good to 
list what hcalls they are relevant to. H_QUERY_VAS_WINDOW (which is not
used anywhere yet?) These are also a 1-byte field, so '0x00', '0x01' etc
would be more appropriate.

> +					   /* deallocated, or quiesced */
> +/* Linux status bits */
> +#define VAS_WIN_NO_CRED_CLOSE	0x00000004 /* Window is closed due to */
> +					   /* lost credit */

This is mixing a user defined bit field with hcall API value field. You
also AFAIKS as yet don't fill in the hypervisor status anywhere.

I would make this it's own field entirely. A boolean would be nice, if
possible.

>  /*
>   * Get/Set bit fields
>   */
> @@ -59,6 +72,8 @@ struct vas_user_win_ref {
>  	struct pid *pid;	/* PID of owner */
>  	struct pid *tgid;	/* Thread group ID of owner */
>  	struct mm_struct *mm;	/* Linux process mm_struct */
> +	struct mutex mmap_mutex;	/* protects paste address mmap() */
> +					/* with DLPAR close/open windows */
>  };
>  
>  /*
> @@ -67,6 +82,7 @@ struct vas_user_win_ref {
>  struct vas_window {
>  	u32 winid;
>  	u32 wcreds_max;	/* Window credits */
> +	u32 status;
>  	enum vas_cop_type cop;
>  	struct vas_user_win_ref task_ref;
>  	char *dbgname;
> diff --git a/arch/powerpc/platforms/book3s/vas-api.c b/arch/powerpc/platforms/book3s/vas-api.c
> index 4d82c92ddd52..2b0ced611f32 100644
> --- a/arch/powerpc/platforms/book3s/vas-api.c
> +++ b/arch/powerpc/platforms/book3s/vas-api.c
> @@ -316,6 +316,7 @@ static int coproc_ioc_tx_win_open(struct file *fp, unsigned long arg)
>  		return PTR_ERR(txwin);
>  	}
>  
> +	mutex_init(&txwin->task_ref.mmap_mutex);
>  	cp_inst->txwin = txwin;
>  
>  	return 0;
> diff --git a/arch/powerpc/platforms/pseries/vas.c b/arch/powerpc/platforms/pseries/vas.c
> index 2ef56157634f..d9ff73d7704d 100644
> --- a/arch/powerpc/platforms/pseries/vas.c
> +++ b/arch/powerpc/platforms/pseries/vas.c
> @@ -501,6 +501,7 @@ static int __init get_vas_capabilities(u8 feat, enum vas_cop_feat_type type,
>  	memset(vcaps, 0, sizeof(*vcaps));
>  	INIT_LIST_HEAD(&vcaps->list);
>  
> +	vcaps->feat = feat;
>  	caps = &vcaps->caps;
>  
>  	rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES, feat,
> @@ -539,6 +540,145 @@ static int __init get_vas_capabilities(u8 feat, enum vas_cop_feat_type type,
>  	return 0;
>  }
>  
> +/*
> + * VAS windows can be closed due to lost credits when the core is
> + * removed. So reopen them if credits are available due to DLPAR
> + * core add and set the window active status. When NX sees the page
> + * fault on the unmapped paste address, the kernel handles the fault
> + * by setting the remapping to new paste address if the window is
> + * active.
> + */
> +static int reconfig_open_windows(struct vas_caps *vcaps, int creds)
> +{
> +	long domain[PLPAR_HCALL9_BUFSIZE] = {VAS_DEFAULT_DOMAIN_ID};
> +	struct vas_cop_feat_caps *caps = &vcaps->caps;
> +	struct pseries_vas_window *win = NULL, *tmp;
> +	int rc, mv_ents = 0;
> +
> +	/*
> +	 * Nothing to do if there are no closed windows.
> +	 */
> +	if (!vcaps->close_wins)
> +		return 0;
> +
> +	/*
> +	 * For the core removal, the hypervisor reduces the credits
> +	 * assigned to the LPAR and the kernel closes VAS windows
> +	 * in the hypervisor depends on reduced credits. The kernel
> +	 * uses LIFO (the last windows that are opened will be closed
> +	 * first) and expects to open in the same order when credits
> +	 * are available.
> +	 * For example, 40 windows are closed when the LPAR lost 2 cores
> +	 * (dedicated). If 1 core is added, this LPAR can have 20 more
> +	 * credits. It means the kernel can reopen 20 windows. So move
> +	 * 20 entries in the VAS windows lost and reopen next 20 windows.
> +	 */
> +	if (vcaps->close_wins > creds)
> +		mv_ents = vcaps->close_wins - creds;
> +
> +	list_for_each_entry_safe(win, tmp, &vcaps->list, win_list) {
> +		if (!mv_ents)
> +			break;
> +
> +		mv_ents--;
> +	}
> +
> +	list_for_each_entry_safe_from(win, tmp, &vcaps->list, win_list) {
> +		/*
> +		 * Nothing to do on this window if it is not closed
> +		 * with VAS_WIN_NO_CRED_CLOSE
> +		 */
> +		if (!(win->vas_win.status & VAS_WIN_NO_CRED_CLOSE))
> +			continue;
> +
> +		rc = allocate_setup_window(win, (u64 *)&domain[0],
> +					   caps->win_type);
> +		if (rc)
> +			return rc;
> +
> +		rc = h_modify_vas_window(win);
> +		if (rc)
> +			goto out;
> +
> +		mutex_lock(&win->vas_win.task_ref.mmap_mutex);
> +		/*
> +		 * Set window status to active
> +		 */
> +		win->vas_win.status &= ~VAS_WIN_NO_CRED_CLOSE;
> +		mutex_unlock(&win->vas_win.task_ref.mmap_mutex);

What's the mutex protecting? If status can change, what happens if it 
changed between checking it above and this point?

> +		win->win_type = caps->win_type;
> +		if (!--vcaps->close_wins)
> +			break;
> +	}
> +
> +	return 0;
> +out:
> +	/*
> +	 * Window modify HCALL failed. So close the window to the
> +	 * hypervisor and return.
> +	 */
> +	free_irq_setup(win);
> +	h_deallocate_vas_window(win->vas_win.winid);
> +	return rc;
> +}
> +
> +/*
> + * Get new VAS capabilities when the core add/removal configuration
> + * changes. Reconfig window configurations based on the credits
> + * availability from this new capabilities.
> + */
> +static int vas_reconfig_capabilties(u8 type)
> +{
> +	struct hv_vas_cop_feat_caps *hv_caps;
> +	struct vas_cop_feat_caps *caps;
> +	int lpar_creds, new_creds;
> +	struct vas_caps *vcaps;
> +	int rc = 0;
> +
> +	if (type >= VAS_MAX_FEAT_TYPE) {
> +		pr_err("Invalid credit type %d\n", type);
> +		return -EINVAL;
> +	}
> +
> +	vcaps = &vascaps[type];
> +	caps = &vcaps->caps;
> +
> +	hv_caps = kmalloc(sizeof(*hv_caps), GFP_KERNEL);
> +	if (!hv_caps)
> +		return -ENOMEM;
> +
> +	mutex_lock(&vas_pseries_mutex);
> +	rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES, vcaps->feat,
> +				      (u64)virt_to_phys(hv_caps));
> +	if (rc)
> +		goto out;
> +
> +	new_creds = be16_to_cpu(hv_caps->target_lpar_creds);
> +
> +	lpar_creds = atomic_read(&caps->target_creds);

NBD but it would be slightly clearer that new_creds is the absolute 
value not the number of new credits if you call these two
new_nr_creds and old_nr_creds or prev_nr_creds.

> +
> +	atomic_set(&caps->target_creds, new_creds);
> +	/*
> +	 * The total number of available credits may be decreased or
> +	 * inceased with DLPAR operation. Means some windows have to be
> +	 * closed / reopened. Hold the vas_pseries_mutex so that the
> +	 * the user space can not open new windows.
> +	 */
> +	if (lpar_creds <  new_creds) {
> +		/*
> +		 * If the existing target credits is less than the new
> +		 * target, reopen windows if they are closed due to
> +		 * the previous DLPAR (core removal).
> +		 */
> +		rc = reconfig_open_windows(vcaps, new_creds - lpar_creds);
> +	}
> +
> +out:
> +	mutex_unlock(&vas_pseries_mutex);
> +	kfree(hv_caps);
> +	return rc;
> +}
> +
>  /*
>   * Total number of default credits available (target_credits)
>   * in LPAR depends on number of cores configured. It varies based on
> @@ -565,6 +705,10 @@ static int pseries_vas_notifier(struct notifier_block *nb,
>  	if (!intserv)
>  		return NOTIFY_OK;
>  
> +	rc = vas_reconfig_capabilties(VAS_GZIP_DEF_FEAT_TYPE);
> +	if (rc)
> +		pr_err("Failed reconfig VAS capabilities with DLPAR\n");
> +
>  	return rc;
>  }
>  
> diff --git a/arch/powerpc/platforms/pseries/vas.h b/arch/powerpc/platforms/pseries/vas.h
> index 0538760d13be..45b62565955b 100644
> --- a/arch/powerpc/platforms/pseries/vas.h
> +++ b/arch/powerpc/platforms/pseries/vas.h
> @@ -21,12 +21,6 @@
>  #define VAS_MOD_WIN_FLAGS	(VAS_MOD_WIN_JOBS_KILL | VAS_MOD_WIN_DR | \
>  				VAS_MOD_WIN_PR | VAS_MOD_WIN_SF)
>  
> -#define VAS_WIN_ACTIVE		0x0
> -#define VAS_WIN_CLOSED		0x1
> -#define VAS_WIN_INACTIVE	0x2	/* Inactive due to HW failure */
> -/* Process of being modified, deallocated, or quiesced */
> -#define VAS_WIN_MOD_IN_PROCESS	0x3
> -
>  #define VAS_COPY_PASTE_USER_MODE	0x00000001
>  #define VAS_COP_OP_USER_MODE		0x00000010
>  
> @@ -84,6 +78,8 @@ struct vas_cop_feat_caps {
>  struct vas_caps {
>  	struct vas_cop_feat_caps caps;
>  	struct list_head list;	/* List of open windows */
> +	int close_wins;		/* closed windows in the hypervisor for DLPAR */

'nr_closed_wins'

This does not look like a capability. Although I guess neither do some
of the fields (used_creds) in vas_cop_feat_caps. Why not at least put
this there with the rest of those credits fields?

Thanks,
Nick

> +	u8 feat;		/* Feature type */
>  };
>  
>  /*
> -- 
> 2.27.0
> 
> 
> 


More information about the Linuxppc-dev mailing list