Date: Wed, 28 Feb 2024 14:33:37 +0530
From: Dibakar Singh <quic_dibasing@...cinc.com>
To: Elliot Berman <quic_eberman@...cinc.com>
CC: <andersson@...nel.org>, <konrad.dybcio@...aro.org>,
        <krzysztof.kozlowski@...aro.org>, <luzmaximilian@...il.com>,
        <bartosz.golaszewski@...aro.org>, <quic_gurus@...cinc.com>,
        <linux-arm-msm@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
        <quic_guptap@...cinc.com>, <quic_pkondeti@...cinc.com>,
        <quic_pheragu@...cinc.com>
Subject: Re: [PATCH] firmware: qcom_scm: Introduce batching of hyp assign
 calls



On 21-Feb-24 12:13 AM, Elliot Berman wrote:
> On Tue, Feb 20, 2024 at 11:24:33AM +0530, Dibakar Singh wrote:
>>
>>
>> On 09-Feb-24 11:36 PM, Elliot Berman wrote:
>>> On Fri, Feb 09, 2024 at 04:55:36PM +0530, Dibakar Singh wrote:
>>>> Expose an API qcom_scm_assign_table to allow client drivers to batch
>>>> multiple memory regions in a single hyp assign call.
>>>>
>>>> In the existing situation, if our goal is to process an sg_table and
>>>> transfer its ownership from the current VM to a different VM, we have a
>>>> couple of strategies. The first strategy involves processing the entire
>>>> sg_table at once and then transferring the ownership. However, this
>>>> method may have an adverse impact on the system because during an SMC
>>>> call, the NS interrupts are disabled, and this delay could be
>>>> significant when dealing with large sg_tables. To address this issue, we
>>>> can adopt a second strategy, which involves processing each sg_list in
>>>> the sg_table individually and reassigning memory ownership. Although
>>>> this method is slower and potentially impacts performance, it will not
>>>> keep the NS interrupts disabled for an extended period.
>>>>
>>>> A more efficient strategy is to process the sg_table in batches. This
>>>> approach addresses both scenarios by involving memory processing in
>>>> batches, thus avoiding prolonged NS interrupt disablement for longer
>>>> duration when dealing with large sg_tables. Moreover, since we process
>>>> in batches, this method is faster compared to processing each item
>>>> individually. The observations on testing both the approaches for
>>>> performance is as follows:
>>>>
>>>> Allocation Size/            256MB            512MB            1024MB
>>>> Algorithm Used           ===========      ===========      ============
>>>>
>>>> Processing each sg_list   73708(us)        149289(us)       266964(us)
>>>> in sg_table one by one
>>>>
>>>> Processing sg_table in    46925(us)         92691(us)       176893(us)
>>>> batches
>>>>
>>>> This implementation serves as a wrapper around the helper function
>>>> __qcom_scm_assign_mem, which takes an sg_list and processes it in
>>>> batches. We’ve set the limit to a minimum of 32 sg_list in a batch or a
>>>> total batch size of 512 pages. The selection of these numbers is
>>>> heuristic, based on the test runs conducted. Opting for a smaller number
>>>> would compromise performance, while a larger number would result in
>>>> non-secure interrupts being disabled for an extended duration.
>>>>
>>>> Co-developed-by: Prakruthi Deepak Heragu <quic_pheragu@...cinc.com>
>>>> Signed-off-by: Prakruthi Deepak Heragu <quic_pheragu@...cinc.com>
>>>> Signed-off-by: Dibakar Singh <quic_dibasing@...cinc.com>
>>>> ---
>>>>    drivers/firmware/qcom/qcom_scm.c       | 211 +++++++++++++++++++++++++
>>>>    include/linux/firmware/qcom/qcom_scm.h |   7 +
>>>>    2 files changed, 218 insertions(+)
>>>>
>>>> diff --git a/drivers/firmware/qcom/qcom_scm.c b/drivers/firmware/qcom/qcom_scm.c
>>>> index 520de9b5633a..038b96503d65 100644
>>>> --- a/drivers/firmware/qcom/qcom_scm.c
>>>> +++ b/drivers/firmware/qcom/qcom_scm.c
>>>> @@ -21,6 +21,8 @@
>>>>    #include <linux/platform_device.h>
>>>>    #include <linux/reset-controller.h>
>>>>    #include <linux/types.h>
>>>> +#include <linux/scatterlist.h>
>>>> +#include <linux/slab.h>
>>>>    #include "qcom_scm.h"
>>>> @@ -1048,6 +1050,215 @@ int qcom_scm_assign_mem(phys_addr_t mem_addr, size_t mem_sz,
>>>>    }
>>>>    EXPORT_SYMBOL_GPL(qcom_scm_assign_mem);
>>>> +/**
>>>> + * qcom_scm_assign_mem_batch() - Make a secure call to reassign memory
>>>> + *				   ownership of several memory regions
>>>> + * @mem_regions:    A buffer describing the set of memory regions that need to
>>>> + *		    be reassigned
>>>> + * @nr_mem_regions: The number of memory regions that need to be reassigned
>>>> + * @srcvms:	    A buffer populated with he vmid(s) for the current set of
>>>> + *		    owners
>>>> + * @src_sz:	    The size of the srcvms buffer (in bytes)
>>>> + * @destvms:	    A buffer populated with the new owners and corresponding
>>>> + *		    permission flags.
>>>> + * @dest_sz:	    The size of the destvms buffer (in bytes)
>>>> + *
>>>> + * Return negative errno on failure, 0 on success.
>>>> + */
>>>> +static int qcom_scm_assign_mem_batch(struct qcom_scm_mem_map_info *mem_regions,
>>>> +				     size_t nr_mem_regions, u32 *srcvms,
>>>> +				     size_t src_sz,
>>>> +				     struct qcom_scm_current_perm_info *destvms,
>>>> +				     size_t dest_sz)
>>>> +{
>>>> +	dma_addr_t mem_dma_addr;
>>>> +	size_t mem_regions_sz;
>>>> +	int ret = 0, i;
>>>> +
>>>> +	for (i = 0; i < nr_mem_regions; i++) {
>>>> +		mem_regions[i].mem_addr = cpu_to_le64(mem_regions[i].mem_addr);
>>>> +		mem_regions[i].mem_size = cpu_to_le64(mem_regions[i].mem_size);
>>>> +	}
>>>> +
>>>> +	mem_regions_sz = nr_mem_regions * sizeof(*mem_regions);
>>>> +	mem_dma_addr = dma_map_single(__scm->dev, mem_regions, mem_regions_sz,
>>>> +				      DMA_TO_DEVICE);
>>>> +	if (dma_mapping_error(__scm->dev, mem_dma_addr)) {
>>>> +		dev_err(__scm->dev, "mem_dma_addr mapping failed\n");
>>>> +		return -ENOMEM;
>>>> +	}
>>>> +
>>>> +	ret = __qcom_scm_assign_mem(__scm->dev, virt_to_phys(mem_regions),
>>>> +				    mem_regions_sz, virt_to_phys(srcvms), src_sz,
>>>> +				    virt_to_phys(destvms), dest_sz);
>>>> +
>>>> +	dma_unmap_single(__scm->dev, mem_dma_addr, mem_regions_sz, DMA_TO_DEVICE);
>>>> +	return ret;
>>>> +}
>>>> +
>>>> +/**
>>>> + * qcom_scm_prepare_mem_batch() - Prepare batches of memory regions
>>>> + * @sg_table:       A scatter list whose memory needs to be reassigned
>>>> + * @srcvms:	    A buffer populated with he vmid(s) for the current set of
>>>> + *		    owners
>>>> + * @nr_src:	    The number of the src_vms buffer
>>>> + * @destvms:	    A buffer populated with he vmid(s) for the new owners
>>>> + * @destvms_perms:  A buffer populated with the permission flags of new owners
>>>> + * @nr_dest:	    The number of the destvms
>>>> + * @last_sgl:	    Denotes to the last scatter list element. Used in case of rollback
>>>> + * @roll_back:	    Identifies whether we are executing rollback in case of failure
>>>> + *
>>>> + * Return negative errno on failure, 0 on success.
>>>> + */
>>>> +static int qcom_scm_prepare_mem_batch(struct sg_table *table,
>>>> +				      u32 *srcvms, int nr_src,
>>>> +				      int *destvms, int *destvms_perms,
>>>> +				      int nr_dest,
>>>> +				      struct scatterlist *last_sgl, bool roll_back)
>>>> +{
>>>> +	struct qcom_scm_current_perm_info *destvms_cp;
>>>> +	struct qcom_scm_mem_map_info *mem_regions_buf;
>>>> +	struct scatterlist *curr_sgl = table->sgl;
>>>> +	dma_addr_t source_dma_addr, dest_dma_addr;
>>>> +	size_t batch_iterator;
>>>> +	size_t batch_start = 0;
>>>> +	size_t destvms_cp_sz;
>>>> +	size_t srcvms_cp_sz;
>>>> +	size_t batch_size;
>>>> +	u32 *srcvms_cp;
>>>> +	int ret = 0;
>>>> +	int i;
>>>> +
>>>> +	if (!table || !table->sgl || !srcvms || !nr_src ||
>>>> +	    !destvms || !destvms_perms || !nr_dest || !table->nents)
>>>> +		return -EINVAL;
>>>> +
>>>> +	srcvms_cp_sz = sizeof(*srcvms_cp) * nr_src;
>>>> +	srcvms_cp = kmemdup(srcvms, srcvms_cp_sz, GFP_KERNEL);
>>>> +	if (!srcvms_cp)
>>>> +		return -ENOMEM;
>>>> +
>>>> +	for (i = 0; i < nr_src; i++)
>>>> +		srcvms_cp[i] = cpu_to_le32(srcvms_cp[i]);
>>>> +
>>>> +	source_dma_addr = dma_map_single(__scm->dev, srcvms_cp,
>>>> +					 srcvms_cp_sz, DMA_TO_DEVICE);
>>>
>>> Please use the new tzmem allocator:
>>>
>>> https://lore.kernel.org/all/20240205182810.58382-1-brgl@bgdev.pl/
>>
>> Noted, I will use this allocator in V2 patch.
>>
>>>
>>>> +
>>>> +	if (dma_mapping_error(__scm->dev, source_dma_addr)) {
>>>> +		ret = -ENOMEM;
>>>> +		goto out_free_source;
>>>> +	}
>>>> +
>>>> +	destvms_cp_sz = sizeof(*destvms_cp) * nr_dest;
>>>> +	destvms_cp = kzalloc(destvms_cp_sz, GFP_KERNEL);
>>>> +
>>>> +	if (!destvms_cp) {
>>>> +		ret = -ENOMEM;
>>>> +		goto out_unmap_source;
>>>> +	}
>>>> +
>>>> +	for (i = 0; i < nr_dest; i++) {
>>>> +		destvms_cp[i].vmid = cpu_to_le32(destvms[i]);
>>>> +		destvms_cp[i].perm = cpu_to_le32(destvms_perms[i]);
>>>> +		destvms_cp[i].ctx = 0;
>>>> +		destvms_cp[i].ctx_size = 0;
>>>> +	}
>>>> +
>>>> +	dest_dma_addr = dma_map_single(__scm->dev, destvms_cp,
>>>> +				       destvms_cp_sz, DMA_TO_DEVICE);
>>>> +	if (dma_mapping_error(__scm->dev, dest_dma_addr)) {
>>>> +		ret = -ENOMEM;
>>>> +		goto out_free_dest;
>>>> +	}
>>>> +
>>>> +	mem_regions_buf = kcalloc(QCOM_SCM_MAX_BATCH_SECTION, sizeof(*mem_regions_buf),
>>>> +				  GFP_KERNEL);
>>>> +	if (!mem_regions_buf)
>>>> +		return -ENOMEM;
>>>> +
>>>> +	while (batch_start < table->nents) {
>>>> +		batch_size = 0;
>>>> +		batch_iterator = 0;
>>>> +
>>>> +		do {
>>>> +			mem_regions_buf[batch_iterator].mem_addr = page_to_phys(sg_page(curr_sgl));
>>>> +			mem_regions_buf[batch_iterator].mem_size = curr_sgl->length;
>>>> +			batch_size += curr_sgl->length;
>>>> +			batch_iterator++;
>>>> +			if (roll_back && curr_sgl == last_sgl)
>>>> +				break;
> 
> Is roll_back bool necessary? last_sgl is NULL in case of not
> rolling back and I think is good sentinel value when not rolling back.
> 
> If last_sgl is NULL, then all sgl entries are assigned. If last_sgl is
> not NULL, then all the sgl entries are assigned until last_sgl is
> encountered (inclusive).

Thanks Elliot for pointing it out. However, in next patchset, I will be 
changing the strategy a little bit. Currently, I am rolling back till 
the current failed batch, but I think this is not correct because we are 
not sure of which state the current batch is in. So, in this case, I am 
planning to implement it in a way that it’ll roll back to the last 
successful batch and leak the current batch by setting it private.

> 
>>>> +			curr_sgl = sg_next(curr_sgl);
>>>> +		} while (curr_sgl && batch_iterator < QCOM_SCM_MAX_BATCH_SECTION &&
>>>> +				curr_sgl->length + batch_size < QCOM_SCM_MAX_BATCH_SIZE);
>>>> +
>>>> +		batch_start += batch_iterator;
>>>> +
>>>> +		ret = qcom_scm_assign_mem_batch(mem_regions_buf, batch_iterator,
>>>> +						srcvms_cp, srcvms_cp_sz, destvms_cp, destvms_cp_sz);
>>>> +
>>>> +		if (ret) {
>>>> +			dev_info(__scm->dev, "Failed to assign memory protection, ret = %d\n", ret);
>>>> +			last_sgl = curr_sgl;
>>>> +			ret = -EADDRNOTAVAIL;
>>>
>>> Probably should not be changing the ret value.
>>
>> We are adjusting the return value here because we have already printed the
>> failed return value earlier. Our goal is to ensure that clients invoking
>> this API consistently receive the same failed values, without needing to
>> worry about the precise details of the failure.
>>
> 
> Clients will need special handling for ENOMEM, ENODEV, or EPROBEDEFER.
> Clients can ignore the details of the error if they don't care, but
> don't mask the error.
> 

Noted. I'll take care of this in next patchset.

>>>
>>> What happens to the memory that has been assigned so far? How do we
>>> reclaim that?
>>
>> We’ve already added rollback functionality in qcom_scm_assign_table that
>> allows batches to be rolled back to the HLOS if there’s a failure.
>>
>>>
>>>> +			break;
>>>> +		}
>>>> +		if (roll_back && curr_sgl == last_sgl)
>>>> +			break;
>>>> +	}
>>>> +	kfree(mem_regions_buf);
>>>> +
>>>> +	dma_unmap_single(__scm->dev, dest_dma_addr,
>>>> +			 destvms_cp_sz, DMA_TO_DEVICE);
>>>> +
>>>> +out_free_dest:
>>>> +	kfree(destvms_cp);
>>>> +out_unmap_source:
>>>> +	dma_unmap_single(__scm->dev, source_dma_addr,
>>>> +			 srcvms_cp_sz, DMA_TO_DEVICE);
>>>> +out_free_source:
>>>> +	kfree(srcvms_cp);
>>>> +	return ret;
>>>> +}
>>>> +
>>>> +/**
>>>> + * qcom_scm_assign_table() - Make a call to prepare batches of memory regions
>>>> + *			     and reassign memory ownership of several memory regions at once
>>>> + * @sg_table:       A scatter list whose memory needs to be reassigned
>>>> + * @srcvms:	    A buffer populated with he vmid(s) for the current set of
>>>> + *		    owners
>>>> + * @nr_src:	    The number of the src_vms buffer
>>>> + * @destvms:	    A buffer populated with he vmid(s) for the new owners
>>>> + * @destvms_perms:  A buffer populated with the permission flags of new owners
>>>> + * @nr_dest:	    The number of the destvms
>>>> + *
>>>> + * Return negative errno on failure, 0 on success.
>>>> + */
>>>> +int qcom_scm_assign_table(struct sg_table *table,
>>>> +			  u32 *srcvms, int nr_src,
>>>> +			  int *destvms, int *destvms_perms,
>>>> +			  int nr_dest)
>>>> +{
>>>> +	struct scatterlist *last_sgl = NULL;
>>>> +	int rb_ret = 0;
>>>> +	u32 new_dests;
>>>> +	int new_perms;
>>>> +	int ret = 0;
>>>> +
>>>> +	ret = qcom_scm_prepare_mem_batch(table, srcvms, nr_src,
>>>> +					 destvms, destvms_perms, nr_dest, last_sgl, false);
>>>> +
>>>> +	if (!ret)
>>>> +		goto out;
>>>> +	new_dests = QCOM_SCM_VMID_HLOS;
>>>
>>> We have the original srcvms. Will it always be HLOS? If so, then do we
>>> need to pass srcvms at all?
>>
>> No, srcvms can be VMs other than HLOS. However, this API does not expect
>> clients to provide srcvms permissions. As a result, when such permissions
>> are absent, we proceed by assigning the previously processed batches to
>> HLOS.
>>
> 
> This is a wrong assumption though, right? You are allowing srcvms to be
> something other than HLOS but assuming HLOS VMID is the only one in the
> error path.
> 

Yes, assuming that this API allows the transfer of page ownership from 
any VM to any other VM, rolling it back to HLOS in case of failure 
doesn’t make much sense. I am thinking of doing it in a way where if the 
source VM is HLOS, then we roll it back to HLOS; otherwise, since we 
don’t have permission for the source VMs, we’ll make the pages private 
and let the clients handle it in such cases.

>>>
>>>> +	new_perms = QCOM_SCM_PERM_EXEC | QCOM_SCM_PERM_WRITE | QCOM_SCM_PERM_READ;
>>>> +	rb_ret = qcom_scm_prepare_mem_batch(table, destvms, nr_dest, &new_dests,
>>>> +					    &new_perms, nr_src, last_sgl, true);
>>>> +	WARN_ON_ONCE(rb_ret);
>>>> +out:
>>>> +	return ret;
>>>> +}
>>>> +EXPORT_SYMBOL_GPL(qcom_scm_assign_table);
>>>
>>> Who uses this function?
>>
>> https://lore.kernel.org/all/cover.1700544802.git.quic_vjitta@quicinc.com/
>>
>> This patch responsible for extending qcom secure heap support which will be
>> the user of this API. In the current implementation, we use
>> qcom_scm_assign_mem and iterate over each sg_list in the sg_table
>> individually to secure the memory. Going forward, we will replace this
>> approach with our API.
>>
>>>
>>>> +
>>>>    /**
>>>>     * qcom_scm_ocmem_lock_available() - is OCMEM lock/unlock interface available
>>>>     */
>>>> diff --git a/include/linux/firmware/qcom/qcom_scm.h b/include/linux/firmware/qcom/qcom_scm.h
>>>> index ccaf28846054..abd675c7ef49 100644
>>>> --- a/include/linux/firmware/qcom/qcom_scm.h
>>>> +++ b/include/linux/firmware/qcom/qcom_scm.h
>>>> @@ -8,6 +8,7 @@
>>>>    #include <linux/err.h>
>>>>    #include <linux/types.h>
>>>>    #include <linux/cpumask.h>
>>>> +#include <linux/scatterlist.h>
>>>>    #include <dt-bindings/firmware/qcom,scm.h>
>>>> @@ -15,6 +16,8 @@
>>>>    #define QCOM_SCM_CPU_PWR_DOWN_L2_ON	0x0
>>>>    #define QCOM_SCM_CPU_PWR_DOWN_L2_OFF	0x1
>>>>    #define QCOM_SCM_HDCP_MAX_REQ_CNT	5
>>>> +#define QCOM_SCM_MAX_BATCH_SECTION	32
>>>> +#define QCOM_SCM_MAX_BATCH_SIZE		SZ_2M
> 
> Move the macros to qcom_scm.c

Noted. I'll take care of this in the next patchset.

> 
>>>>    struct qcom_scm_hdcp_req {
>>>>    	u32 addr;
>>>> @@ -93,6 +96,10 @@ int qcom_scm_mem_protect_video_var(u32 cp_start, u32 cp_size,
>>>>    int qcom_scm_assign_mem(phys_addr_t mem_addr, size_t mem_sz, u64 *src,
>>>>    			const struct qcom_scm_vmperm *newvm,
>>>>    			unsigned int dest_cnt);
>>>> +int qcom_scm_assign_table(struct sg_table *table,
>>>> +			  u32 *srcvms, int nr_src,
>>>> +			  int *destvms, int *destvms_perms,
>>>> +			  int nr_dest);
>>>>    bool qcom_scm_ocmem_lock_available(void);
>>>>    int qcom_scm_ocmem_lock(enum qcom_scm_ocmem_client id, u32 offset, u32 size,
>>>> -- 
>>>> 2.34.1
>>>>
>>

Powered by blists - more mailing lists