The SCL clock is always generated by the I2C master. The specification requires minimum periods for the low and high phases of the clock signal. Hence, the actual clock rate may be lower than the nominal clock rate e.g. in I2C buses with large rise times due to high capacitances.
I2C devices can slow down communication by stretching SCL: During an SCL low phase, any I2C device on the bus may additionally hold down SCL to prevent it from rising again, enabling it to slow down the SCL clock rate or to stop I2C communication for a while. This is also referred to as clock synchronization.
Note: The I2C specification does not specify any timeout conditions for clock stretching, i.e. any device can hold down SCL as long as it likes.
Several I2C multi-masters can be connected to the same I2C bus and operate concurrently. By constantly monitoring SDA and SCL for start and stop conditions, they can determine whether the bus is currently idle or not. If the bus is busy, masters delay pending I2C transfers until a stop condition indicates that the bus is free again.
However, it may happen that two masters start a transfer at the same time. During the transfer, the masters constantly monitor SDA and SCL. If one of them detects that SDA is low when it should actually be high, it assumes that another master is active and immediately stops its transfer. This process is called arbitration.