I no longer know why a validation check ensured the size of an entry
passed to gsi_trans_pool_init() was restricted to be a multiple of 8.
For 32-bit builds, this condition doesn't always hold, and for DMA
pools, the size is rounded up to a power of 2 anyway.
Remove this restriction.
Signed-off-by: Alex Elder <elder@linaro.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
A recent patch avoided doing 64-bit modulo operations by checking
the alignment of some DMA allocations using only the lower 32 bits
of the address.
David Laight pointed out (after the fix was committed) that DMA
allocations might already satisfy the alignment requirements. And
he was right.
Remove the alignment checks that occur after DMA allocation requests,
and update comments to explain why the constraint is satisfied. The
only place IPA_TABLE_ALIGN was used was to check the alignment; it is
therefore no longer needed, so get rid of it.
Add comments where GSI_RING_ELEMENT_SIZE and the tre_count and
event_count channel data fields are defined to make explicit they
are required to be powers of 2.
Revise a comment in gsi_trans_pool_init_dma(), taking into account
that dma_alloc_coherent() guarantees its result is aligned to a page
size (or order thereof).
Don't bother printing an error if a DMA allocation fails.
Suggested-by: David Laight <David.Laight@ACULAB.COM>
Signed-off-by: Alex Elder <elder@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
When the coherent memory is freed in gsi_trans_pool_exit_dma(), we
are mistakenly passing the size of a single element in the pool
rather than the actual allocated size. Fix this bug.
Fixes: 9dd441e4ed ("soc: qcom: ipa: GSI transactions")
Reported-by: Stephen Boyd <swboyd@chromium.org>
Tested-by: Sujit Kautkar <sujitka@chromium.org>
Signed-off-by: Alex Elder <elder@linaro.org>
Reviewed-by: Bjorn Andersson <bjorn.andersson@linaro.org>
Link: https://lore.kernel.org/r/20201203215106.17450-1-elder@linaro.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Transactions sit on one of several lists, depending on their state
(allocated, pending, complete, or polled). A spinlock protects
against concurrent access when transactions are moved between these
lists.
Transactions are also reference counted. A newly-allocated
transaction has an initial count of 1; a transaction is released in
gsi_trans_free() only if its decremented reference count reaches 0.
Releasing a transaction includes removing it from the polled (or if
unused, allocated) list, so the spinlock is acquired when we release
a transaction.
The reference count is used to allow a caller to synchronously wait
for a committed transaction to complete. In this case, the waiter
takes an extra reference to the transaction *before* committing it
(so it won't be freed), and releases its reference (calls
gsi_trans_free()) when it is done with it.
Similarly, gsi_channel_update() takes an extra reference to ensure a
transaction isn't released before the function is done operating on
it. Until the transaction is moved to the completed list (by this
function) it won't be freed, so this reference is taken "safely."
But in the quiesce path, we want to wait for the "last" transaction,
which we find in the completed or polled list. Transactions on
these lists can be freed at any time, so we (try to) prevent that
by taking the reference while holding the spinlock.
Currently gsi_trans_free() decrements a transaction's reference
count unconditionally, acquiring the lock to remove the transaction
from its list *only* when the count reaches 0. This does not
protect the quiesce path, which depends on the lock to ensure its
extra reference prevents release of the transaction.
Fix this by only dropping the last reference to a transaction
in gsi_trans_free() while holding the spinlock.
Fixes: 9dd441e4ed ("soc: qcom: ipa: GSI transactions")
Reported-by: Stephen Boyd <swboyd@chromium.org>
Signed-off-by: Alex Elder <elder@linaro.org>
Link: https://lore.kernel.org/r/20201114182017.28270-1-elder@linaro.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
IPA transactions describe actions to be performed by the IPA
hardware. Three cases use IPA transactions: transmitting a socket
buffer; providing a page to receive packet data; and issuing an IPA
immediate command. An IPA transaction contains a scatter/gather
list (SGL) to hold the set of actions to be performed.
We map buffers in the SGL for DMA at the time they are added to the
transaction. For skb TX transactions, we fill the SGL with a call
to skb_to_sgvec(). Page RX transactions involve a single page
pointer, and that is recorded in the SGL with sg_set_page(). In
both of these cases we then map the SGL for DMA with a call to
dma_map_sg().
Immediate commands are different. The payload for an immediate
command comes from a region of coherent DMA memory, which must
*not* be mapped for DMA. For that reason, gsi_trans_cmd_add()
sort of hand-crafts each SGL entry added to a command transaction.
This patch fixes a problem with the code that crafts the SGL entry
for an immediate command. Previously a portion of the SGL entry was
updated using sg_set_buf(). However this is not valid because it
includes a call to virt_to_page() on the buffer, but the command
buffer pointer is not a linear address.
Since we never actually map the SGL for command transactions, there
are very few fields in the SGL we need to fill. Specifically, we
only need to record the DMA address and the length, so they can be
used by __gsi_trans_commit() to fill a TRE. We additionally need to
preserve the SGL flags so for_each_sg() still works. For that we
can simply assign a null page pointer for command SGL entries.
Fixes: 9dd441e4ed ("soc: qcom: ipa: GSI transactions")
Reported-by: Stephen Boyd <swboyd@chromium.org>
Tested-by: Stephen Boyd <swboyd@chromium.org>
Signed-off-by: Alex Elder <elder@linaro.org>
Link: https://lore.kernel.org/r/20201022010029.11877-1-elder@linaro.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In "gsi_trans.c", the field mask TRE_FLAGS_IEOB_FMASK is defined but
never used. Although there's no harm in defining this, remove it
for now and redefine it at some future date if it becomes needed.
This is warned about if "W=2" is added to the build command.
Signed-off-by: Alex Elder <elder@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a command gets added to a transaction for the AP->command
channel we set the DMA address of its scatterlist entry, but not
its DMA length. Fix this bug.
Signed-off-by: Alex Elder <elder@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch implements GSI transactions. A GSI transaction is a
structure that represents a single request (consisting of one or
more TREs) sent to the GSI hardware. The last TRE in a transaction
includes a flag requesting that the GSI interrupt the AP to notify
that it has completed.
TREs are executed and completed strictly in order. For this reason,
the completion of a single TRE implies that all previous TREs (in
particular all of those "earlier" in a transaction) have completed.
Whenever there is a need to send a request (a set of TREs) to the
IPA, a GSI transaction is allocated, specifying the number of TREs
that will be required. Details of the request (e.g. transfer offsets
and length) are represented by in a Linux scatterlist array that is
incorporated in the transaction structure.
Once all commands (TREs) are added to a transaction it is committed.
When the hardware signals that the request has completed, a callback
function allows for cleanup or followup activity to be performed
before the transaction is freed.
Signed-off-by: Alex Elder <elder@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>